How AI vendors are
actually pricing.

The SPP read

Anthropic cancelled a scheduled price increase, then removed the evidence there had been one. The Sonnet 5 page carried a dated footnote: $2/$10 per million input/output tokens through August 31, with $3/$15 standard pricing after it. On August 11 the increase was cancelled and the introductory rate made permanent. Between August 9 and 16 the footnote came off the page entirely, along with the asterisks pointing at it. A buyer arriving today sees $2/$10 as the price with no way to learn it was ever temporary, or that the scheduled reversion would have been a 50% rise. The cancellation is the concession. The deletion is what makes the next dated price easier to set.

The SPP read

Anthropic made Claude Sonnet 5's introductory rate of $2 per million input tokens and $10 per million output permanent on August 10, withdrawing the standard $3/$15 pricing scheduled for September 1. No existing invoice moves; what moved is the effective date: a planned 50% step-up canceled three weeks before it took effect. Holding the introductory price protects the displacement math Sonnet 5 was priced for.

The SPP read

A new flagship-class model at the predecessor's exact price is a price hold doing strategic work: near-Fable capability now sits at half Fable's rate, undercutting the flagship from inside the house. When capability compounds while the price line holds, the effective price of the tier is falling, and the metered flagship above it has to re-justify its premium every release cycle.

The SPP read

From July 20 the terms split by plan. Max, Team Premium and premium Enterprise seats keep the model included at up to half of their weekly limit, with usage above it metered at the model's API rate; Pro and Team Standard move to usage credits, with a one-time $100 credit. Read as a sequence: a launch with inclusion and no cap, a return with a cap and a dated meter, two extensions of that date, and a split that kept the cap-and-meter design for the top plans and moved the other plans onto the meter. The one-time credit is the only element of the final terms that appeared in none of the earlier versions.

The SPP read

The second extension came five days after the first and moved the switch again, from July 12 to July 19, with the temporary +50% Claude Code weekly limit extended to the same date. Two extensions in six days, with the cap and the credit rate unchanged through both. The terms that replaced the window were announced on July 17 and 18.

The SPP read

The extension was announced on the day the window was due to close, and it moved the switch to usage credits from July 7 to July 12. Nothing else in the terms moved: the cap stayed at half the weekly limit and the credit rate stayed where it was. The date was the only variable that changed.

The SPP read

The model returned nineteen days after it was withdrawn, on different terms than it left with. At launch it was included in paid plans with no cap on how much of the weekly limit it could use, for a two-week window. On return it was included at up to half of each plan's weekly limit, for six days, with usage credits at roughly API rates taking over after that. Enterprise Standard seats were metered from the first day back. Two things changed at once: how much of the model a subscription carries, and what usage beyond that amount costs. Both were in place on the day access resumed.

The SPP read

An introductory rate with a published end date is a promise about a future price, and this one landed in the same season the vendor was extending a different dated promise rather than keeping it. The workhorse model gets a discount with a calendar attached while the flagship's own deadline keeps moving. Buyers read both signals together. A promotional price only teaches the market what the real price is if the reversion actually happens, and every extension elsewhere makes the reversion look negotiable. The mechanism here is ordinary. The credibility cost of running it in this particular window is not.

The SPP read

Clearance restored the product, not the offer. Subscribers came back to a capped allocation and a dated meter in place of the unmetered window promised at launch. When an interruption forces a relaunch, the relaunch terms are a repricing decision, and the market reads them as one.

The SPP read

No pricing architecture plans for a regulator switching the product off three days into a launch window. The suspension turned a simple included-access promise into a public obligation, and every later move in this arc is Anthropic repricing that obligation. For your own launches, the design question is how much promotional promise remains outstanding if access is ever interrupted.

The SPP read

The API rate is the meter the credits translate to: ten dollars in, fifty out, the top of Anthropic's list and roughly twice the workhorse model's rate. Pricing the same tokens two ways, bundled and metered, makes the gap between those two prices the real product decision of this arc.

The SPP read

Included access arrived with an expiry date attached: the new flagship shipped at no extra cost to paid subscribers for a fixed window, with usage credits taking over on a named date. Free-for-now with the meter's start date published up front is a different promise than free, and it sets the reference point buyers carry into the first metered month.

The SPP read

Anthropic has extended the window of its Claude Code weekly-limits promotion, which raises usage allowances for Pro, Max, and Team plans as well as legacy Enterprise seats. The promotion's end date has been pushed out and the support article now reflects the updated close date. Five-hour limits remain outside the promotion's scope.

The SPP read

Anthropic raised the ceiling on how much Claude Code usage subscribers can consume within a rolling five-hour window across its Pro, Max, Team, and seat-based Enterprise plans, while simultaneously retiring the peak-hour reduction that had been constraining Pro and Max users. Weekly caps introduced in a prior cycle were left in place, and the same announcement extended higher rate limits to Claude Opus on the API. The move repositions the allowance structure without touching the base subscription price, adding a passthrough option for Max subscribers who exhaust their included usage.

The SPP read

A price hold at a moment of capability compression is still a pricing decision even though no number on the list changes: Sonnet 4.6 arrives twelve days after the flagship at Sonnet's unchanged rate, closing on Opus 4.6's capability without closing on its price. Every time the mid-tier model catches the flagship this fast, the flagship's premium has to re-earn its place on the list, the same pressure Opus 5 later resolved by holding its own price against Fable 5.

The SPP read

The ladder-fill finished at the org level with two seat types rather than one: standard seats keep the existing plan, premium seats add usage headroom and Claude Code access for a single incremental price per user. Individual subscribers got the bundle outright across Max and Pro; teams get it as a paid upgrade path admins assign seat by seat. The premium seat becomes the anchor for every Team and Enterprise pricing conversation that follows, the point where more usage and agent access are sold as one line rather than two.

The SPP read

A weekly limit with paid overage at published API rates bounds the vendor's exposure. It does not bound the buyer's. Anthropic ran that shape a year before the sequence that made it famous, and the asymmetry runs one direction: the ceiling is denominated in tokens, a unit no buyer can forecast from the work they actually intend to do, so the limit lands at a moment they cannot anticipate and the spend above it accrues at a rate they cannot plan. Pricing the overage at list rather than a marked-up surrogate does treat the heavy user squarely. What it cannot fix is a meter counting something the buyer has no way to see coming, which is a metric problem rather than a budgeting one. The same mechanism returns in mid 2026 behind a model at roughly twice the workhorse rate, with inclusion capped at half the weekly limit and everything above it metered.

The SPP read

The bundle moved one plan down the ladder with a model ceiling attached: five weeks after Max absorbed Claude Code, the same at-no-extra-charge access reached the Pro plan, capped to Sonnet rather than Max's full model range. The entry subscription now buys agent access, not just chat, and the ceiling is what keeps Pro's price from buying Max's depth. Two plans in, the differentiator has already shifted from access itself to which model the access runs on.

The SPP read

A metered research preview became a subscription line item: three weeks after Max launched, Claude Code moved off pay-per-token Console billing and into Max's shared rate limits at no separate charge. The chat app and the coding agent had priced access two different ways since Code's February preview; this collapsed them into one bill and one limit pool. It is the opening beat of a ladder-fill: Pro follows within five weeks, then a Team and Enterprise premium seat within four months.

The SPP read

GitHub extended its downloadable AI usage report to expose token consumption at the per-model level, splitting activity into four distinct categories. For enterprise buyers, the observable unit moves from aggregate model usage to the input, output, and cache dimensions that underlie Copilot billing. The change sits on the governance arc: the meter itself did not move, but the buyer's ability to read and act on it materially expanded.

The SPP read

GitHub has added Kimi K3 to Copilot and placed it on a per-token consumption meter covering input, output, and cached-input traffic. This is the first published rate for this model inside the Copilot marketplace, establishing a new entry on GitHub's model-selection price surface. Buyers choosing between Copilot-available models now have a stated token cost for Kimi K3 to weigh against alternatives.

The SPP read

The value metric does the work here: per active committer prices the people whose code the product touches, not the whole seat roster. It self-scopes to usage, which softens procurement pushback, and it rides the metric Copilot already trained GitHub's buyers to accept. The exposure is the agent question: when AI agents commit code, the definition of an active committer becomes a licensing decision, and today's metric language decides whether that expansion arrives priced or free.

The SPP read

Budgets and usage caps attach to the cost center giving more finely grained controls on variability. Rather than change the pricing architecture, GitHub is gearing more tools for customers to manage their own spend. Expect this trend to continue until a breakout vendor’s alternative model is to absorb the risk, ditch the tools (and the requirement for customers to have staff managing these tools) and make it easier for customers to handle variability.

The SPP read

An open-weight model entering the picker puts a lower-priced option on the same menu as the frontier models, and GitHub hosts this one itself while still describing the charge as provider list pricing. That phrase does quiet work: it frames the credit draw as a relay of someone else's price even where GitHub is the provider, which makes the rate read as arithmetic rather than a decision. Once customers pay by consumption, the model picker stops being a quality control and becomes a price list they operate. The vendor that writes the menu sets the spread, whatever the menu says about who set the price.

The SPP read

BYOK routes inference through the customer's own provider account three weeks after credits went live, moving the token bill off GitHub's meter entirely. What stays billable is the harness: the agent loop, the client, the repository context, the governance. That splits what a Copilot subscription had been selling as one product, and it hands the largest buyers a lever they will eventually pull at renewal. A vendor that lets you bring your own tokens has told you which half of the product it believes is defensible.

The SPP read

The effective date is where the architecture sharpened: multipliers rose the same day billing went live, and code review began drawing on two meters at once, Actions minutes and AI Credits. Notice what is absent: a legacy conversion. The pattern that holds across our engagements runs new-logo pricing first, with the installed base following on a modeled schedule. GitHub instead repriced the installed base in place on a date, cushioned by promotional and flex credits that expire on a calendar. A flash-cut of the base is the most damaging path a repricing can take, and doubly so in a competitive market where the smallest customers carry the lowest switching costs. The cushions soften the first invoice, not the decision to leave.

The SPP read

The announcement did the heavy lifting a quarter early: plan prices held still while the unit underneath them changed from requests to AI Credits, and promotional credit cushions bought the installed base a quiet first quarter. Swapping the meter while the visible price stands still is the least-resistance path through a repricing, and the cushion is its anesthetic. The bill arrives when the promo expires, not when the announcement ships.

The SPP read

The annual cohort was not migrated. It was left on the old meter with its multipliers raised, then dropped to the free plan at expiry unless it re-bought. That turns a renewal into a fresh purchase decision at the moment the customer has the least leverage and the vendor has the most: mid-arc, on a repriced surface, with the prior contract already gone. The cohort that had been paying annually was the one asked to buy again at the new rates.

The SPP read

GitHub stopped selling the individual plans, tightened their limits, and pulled the most expensive model out of Pro, then said why in pricing terms: a handful of requests could now cost more than the plan itself. The stated reason is the mechanism: a flat monthly fee with per-request costs the fee did not bound, and the model pulled from Pro was the one carrying the largest of those costs. The repricing followed one week later. The sequence on the record: access tightened first, with the cost reason stated, and the repricing dated one week later.

The SPP read

Auto bills at whatever model it picks, and at launch it picked only from the low end of the menu: models carrying a zero-to-one multiplier. Choosing auto then cost ten percent less than choosing the same model directly. The discount prices a decision rather than a resource. The customer gives up model choice, GitHub keeps the spread between the model the task needs and the model the customer would have reached for, and the router is bounded so that spread can only run one way. Paying customers to stop choosing is how a reseller manufactures a margin the meter does not hand it.

The SPP read

Enforcement is the move, not the announcement: on June 18 the allowance became a wall, and passing through it cost four cents a request. Two design choices sit underneath. The overage tap shipped switched off, so a developer had to opt into a variable bill rather than discover one, and completions stayed unmetered so the daily experience never felt counted. Meter the frontier, leave the habit free, and the bill lands only on work the buyer already believes is expensive.

The SPP read

The plan split the product surface in two: unlimited use of the base model, and a counted allowance of premium requests for everything else. That is a new metric introduced under cover of a new top tier, where a higher price presents the unit as generosity rather than a meter. Every later step in GitHub's arc runs through that unit. The first time a vendor counts something, the number is almost never the point; the habit of counting is.

The SPP read

AWS did not cap what a penetration test can cost; it let the buyer set a hard stop on it. A maximum task-hour limit ends the test when the limit is reached and keeps the findings gathered up to that point, which bounds the spend but says nothing about whether the test finished: coverage is whatever the hours bought. The meter itself did not move, and the variance stays where the metric put it, with the buyer, who can now truncate the work rather than forecast it. Revalidation runs on the same meter, re-testing individual findings by task-hour duration. Read together with the launch of the meter and its migration into Continuum, this is the vendor answering the boundedness question by handing the buyer the switch, not by redesigning the unit.

The SPP read

Amazon listed xAI's Grok 4.6 on Bedrock under three cross-region inference tiers, each carrying its own per-token rate card, with Standard serving as the baseline from which Priority and Flex rates are derived. The move brings a new model onto Bedrock's token meter and introduces a named service-tier multiplier structure on top of it, a mechanism not previously documented for xAI models on the platform. The pricing page was reorganized by provider at the same time, surfacing Grok 4.6 and Grok 4.3 through a model selector.

The SPP read

AWS has stated the commercial model for Continuum for code vulnerabilities before publishing a price: the capability is in gated preview, and its VP says the customer buys Continuum at one price while AWS chooses which frontier model runs each phase and carries those model costs itself. That is a decision about which unit the buyer sees, made ahead of the rate card. The task-hour meter already carried from Security Agent into Continuum pen testing stays the platform's only published price, so the open question this move sets up is whether the flagship capability will land on that same meter or on a unit of its own. Either way the supplier's token variability stops at AWS, which is the harness absorbing the metered input and selling in its own terms.

Related: AI Harness Pricing

The SPP read

AWS relaunched its security agent products under the Continuum platform, preserving the existing task-hour meter for penetration testing and code scanning under new product names rather than resetting commercial terms. Simultaneously, Continuum for code vulnerabilities introduced a model-agnostic agent-team architecture that routes each phase to whichever frontier model performs best, entering gated preview without a published price. The move places an established outcome meter inside a broader platform while holding open a pricing-blank for the flagship capability.

The SPP read

AWS introduced a commercially metered AI security agent priced on active working time rather than requests, seats, or outcomes, marking the first appearance of a task-hour unit in the observatory's harness. The metric sits at the outcomes end of the arc, binding cost directly to agent labor rather than to tokens consumed or a flat subscription. A free-trial window for new customers accompanies the launch, and the product was subsequently absorbed into a broader AWS bundle.

The SPP read

Amazon collapsed two request meters into one currency: Kiro's separate vibe and spec request pools became a single credit pool across a four-tier ladder, announced while usage was still free and armed two weeks later when billing switched on. The unit unification is the story; an incumbent shipped its agentic IDE's second pricing model before its first invoice, correcting the meter's architecture mid-preview rather than after customer bills made the flaw loud.

The SPP read

Crescendo has restructured when and whether it invoices by making billing contingent on measured outcome quality rather than on consumption of AI interactions. Interactions falling below a predictive quality threshold are excluded from billing entirely and credited back, while speed-to-go-live and baseline-beating quality conditions gate the start of invoicing. The move shifts the vendor's commercial meter from AI activity delivered to outcomes validated, placing it at the outcomes end of the pricing arc.

The SPP read

The raise landed August 16, 2026 at 16:00 UTC, exactly as the August 13 release note scheduled it. Flat per-token pricing became a peak/off-peak grid and every cell went up, with off-peak set at half of peak; the full rate card sits in the claim beside this read. The increase-warning footnote came off the page the day the increase arrived. This is the ledger's only prices-up reversal, and the mechanism it chose is the one it had rehearsed as a discount: the clock.

The SPP read

A week later the footnote dropped the mechanism and kept the direction: 'We plan to raise the overall pricing for DeepSeek API services in the near future, with a significant increase expected. Please plan your usage accordingly.' (SPP capture August 9, 2026; the South China Morning Post quoted the wording on August 6.) A vendor telling developers on the rate card itself to plan for a significant increase is the plainest prices-up signal in this ledger.

The SPP read

The reversal announces itself in a footnote. Between July 26 and August 2, 2026 (SPP captures), DeepSeek added a note to its live rate card: the API 'will soon adopt a peak/off-peak pricing policy,' peak prices at 2x regular across all billing items, peak hours 9:00-12:00 and 14:00-18:00 Beijing time daily. The mechanism that debuted in February 2025 as a nightly discount returns pointed the other way: the regular price becomes the floor and the peak window carries the raise. Listed prices unchanged; only the warning shipped.

The SPP read

The V3.2-Exp release cut API prices by more than half, effective immediately, and the vendor's stated reason was architectural: sparse attention cutting inference cost. Archived rate cards bracket it: V3.1-Terminus at $0.07/$0.56/$1.68 per million on September 22, V3.2-Exp at $0.028/$0.28/$0.42 on September 29. This is the floor of the arc; every later move is up from here.

The SPP read

Time-of-day metering enters this arc as a discount. On February 26, 2025 DeepSeek split its rate card into a standard price (UTC 00:30-16:30) and an off-peak price (16:30-00:30): V3 half off, R1 75% off, both landing on identical off-peak dollar rates, each request's completion timestamp deciding its tier. Utility-grade peak-load pricing applied to inference, eighteen months before DeepSeek reused the mechanism in the opposite direction. The archived vendor page is the record; the announcement post no longer exists on the vendor's site.

The SPP read

DeepSeek's R1 arrived priced, not previewed. The January 20, 2025 release note carried the rate card in the announcement itself: deepseek-reasoner at $0.14 per million input tokens on cache hit, against $0.55 on cache miss, the cache as the discount mechanism from day one. Archived captures bracket the debut: no reasoner row on the pricing page January 16, the full row by January 21.

The SPP read

The $2.19 output price is the number to read on R1's launch card. Reasoning models bill their thinking as output, and the card said so: chain-of-thought and final-answer tokens priced the same. A question that comes back as one short paragraph can still bill thousands of tokens, because the model reasons at length before it answers and the buyer pays for all of it. The expensive half of reasoning, the part nobody reads, went on a published flat rate from day one.

The SPP read

After a year of each generation entering above the last, the third-generation Flash included, this release turned the other way: output fell about a sixth while input held, and the model's own pitch is that tasks complete on materially fewer tokens. That makes the move two price cuts at once: one posted on the rate card, and a second inside the tokens a task consumes, which no rate card shows. When models differ in tokens per task, per-token prices stop being comparable across vendors, and the vendor decides how much of the invisible cut to pass through and how much to keep as margin and competitive room.

The SPP read

The bottom rung was cut nearly in half within months of reaching the home market, with the storage inside it doubled in the same season. The down-market rung is priced for reach, not margin, and every cut there resets what the rung above must justify.

The SPP read

Top-up credits complete the credit system: the allotment seeds the habit, the top-up monetizes its overflow. Once a subscription's ceiling can be bought through in small denominations, the plan price is a floor, and the meter above it is the growth instrument.

The SPP read

The ceiling experiment corrected: the top rung cut, a mid rung inserted underneath, and the same event rotated consumer limits from prompts toward compute. A price cut wrapped around a metering change is the ledger's most repeated pairing; the visible number falls while the unit underneath grows more precise.

The SPP read

More inclusion at a held price is a cut that never touches the price page, the same instrument the model layer runs with capability. Inclusion increases are how a vendor defends a rung without repricing the ladder around it.

The SPP read

Google is moving Gemini API billing in AI Studio from postpaid invoices to prepaid credits, account by account: an assignment program running since March 2026 sets per-customer cutover deadlines, lower-usage tiers are required to switch while the highest tier can remain postpaid, and other Google Cloud services stay as they are. Rates do not move; what moves is when the money arrives, shifting collection risk and commitment onto the buyer, with unused credits expiring after 12 months. Breakage and float economics become the default billing posture at the largest vendor, framed as giving developers more control over consumption and costs.

The SPP read

The reversal beat: capability given away into every plan in 2025 came back as a paid access ladder in 2026, with a deadline on the promotional generosity. Bundling and re-monetization are one cycle, not two strategies; what the seat price absorbs can be re-fenced later at a price.

The SPP read

The down-market rung reached the home market with a credit allotment inside: the cheapest paid seat is also the most tightly metered one. At the bottom of the ladder the subscription is mostly a container for credits.

The SPP read

A new generation priced above its predecessor breaks the industry's own script that intelligence only falls in price. Each Gemini generation entered above the last at the API layer for a year, and the price of the frontier rose while the budget lanes advertised the fall.

The SPP read

Pricing images in output tokens pulls a new modality under the existing meter instead of minting a new unit for it. The token becomes the house currency: whatever the model produces, the bill speaks tokens. Extending one meter beats explaining two.

The SPP read

Doubling the monthly credit allotment without touching the price is a cut the price page never shows, and it is only possible because the credit is a surrogate unit the vendor controls. The exchange rate between the subscription and the underlying work moved; the number on the plan did not.

The SPP read

Every model family grows a budget lane, and the budget lane is doing pricing work: it catches the workloads the mid tier would lose to rivals on cost, and it anchors the family's floor while the flagship holds the ceiling. The ladder, not any one rung, is the pricing architecture.

The SPP read

Half price for workloads that can wait is a price on urgency: the vendor prices the buyer's flexibility instead of its own cost. Batch discounts sort traffic by latency tolerance, and the meter learns which work is time-critical, which is exactly the information outcome pricing will need.

The SPP read

The consumer ceiling rung arrived at work as a self-serve Workspace add-on, seat price unstated in the vendor's own post. When the top of the consumer ladder becomes a business SKU, the ladder itself has become the packaging architecture.

The SPP read

One repricing, two directions: output tokens cut while input tokens doubled, and the thinking premium folded into a single price. The input side is where agentic workloads live, because context in is what loops consume; pricing input up while headline output falls moves the bill toward the workloads that grew. The simplification story and the repricing story are the same announcement.

The SPP read

A $250 consumer rung is a statement about where the ceiling sits: Google priced its top plan an order above the default tier and fenced the agentic capability inside it, with concurrent agent tasks as the real gate. The premium rung is where agentic capacity gets rationed first; storage and models are the wrapping. The same post renamed the base plan from AI Premium to AI Pro, because Premium is a dead-end name once something must sit above it; the rename cleared the way for the ceiling tier it shipped with.

The SPP read

Collapsing the long-context price split is not a cut; it is a redistribution. The merged rate landed between the old pair: long-context work billed cheaper, while the short-context work that makes up most traffic paid more than its old floor. The simplification label carried a price increase for the majority of calls, and the same two-directional move returns later in this arc.

The SPP read

The same week Microsoft ran this play on consumers, Google ran it on business seats: the AI add-on retired, the capability folded into every Workspace plan, and the seat price raised behind it, roughly a sixth to a fifth higher by plan. Bundling converts an optional AI SKU nobody had to buy into a price increase nobody can decline.

The SPP read

The arc opens with a familiar instrument: a flat consumer subscription, AI folded into a storage bundle at a seat-shaped price. Every later move on this ledger is Google discovering what that one number could not hold.

The SPP read

Grok Bot is not sold on its own. Access rides with the top tier of xAI's own subscription and of the Cursor plans, so a buyer reaches the agent by moving up a plan rather than by buying a separate product or metering the work it does. For a team the entry point is a per seat team plan, which sets the smallest unit at which an organization can adopt the capability.

The SPP read

Grok 4.5 lands at a fraction of the rates the incumbent frontier models charge, and its three-to-one output multiple is the mildest on the board. That is a challenger's price surface: make switching feel like found money and force the incumbents to defend their premium. For vendors building on model APIs, a widening spread between frontier price floors is margin room, but only for architectures that can route work across models without repricing the customer.

The SPP read

The budget lane closed the way few price increases are ever announced: as a migration. Eight models retired, and the retired endpoints redirect to a model priced several times higher, so a budget buyer's bill rose without any price on any page changing. Redirect-repricing is the meter's quietest lever, and it only works because API buyers bind to model names rather than to rates. Watch for it wherever a vendor prunes a model list: the retirement notice is also the price sheet.

The SPP read

Beside the token meter, a second meter appeared denominated in something new: successful tool invocations, billed per thousand calls that worked. Charging only on success moves a unit of risk from buyer to vendor, and it is the first step on this arc that prices what the model accomplished rather than what it consumed. Small in revenue terms, large in signal: a vendor experimenting with success-denominated metering is pricing something other than tokens, which measure effort, not value.

The SPP read

A budget lane opened an order of magnitude below the flagship, and the vendor's own framing priced intelligence rather than tokens: near-flagship benchmark performance at a small fraction of the rate. Two prices for the same tokens depending on context length made the meter itself tiered. And when the same benchmark performance sells at token rates an order of magnitude apart, the token has stopped being a consistent unit of the thing the buyer is buying. The vendor's framing prices intelligence while its meter still prices tokens, and the distance between those two is where this arc's later moves happen.

The SPP read

A late API entrant priced its flagship at the going flagship rate, three dollars in and fifteen out, matching the incumbent workhorse to the dollar. Entering at parity rather than under it says the vendor is selling capability membership, not a discount, and it marks the moment this arc joins the metered market it will spend the next year renegotiating.

The SPP read

Hours after the flagship model shipped, the subscription tier carrying it went from $22 to $40 a month. Pricing the carrier rather than the model kept the model's own price page clean while nearly doubling what access costs, and the same launch window brought a standalone Grok subscription, reported at the time rather than posted, which marks the start of Grok's commercial life outside the platform. When a model's access price lives in another product's tier, every model launch is a repricing option on that tier, and this vendor exercised it within hours.

The SPP read

The arc opens with a flip that looks like generosity and functions as instrumentation: a paid-only assistant went free for every user on the platform, with a bounded allowance metering how much free anyone gets and the paid tiers keeping the higher limits. A free tier with an explicit bound is pricing already at work. The bound converts free users into a measured population, and the allowance boundary becomes the upsell line that every later move on this arc renegotiates.

The SPP read

One announcement, three prices. On July 30, 2026 OpenAI cut GPT-5.6 Luna by 80 percent, to $0.20 per million input tokens and $1.20 per million output, trimmed GPT-5.6 Terra 20 percent to $2 and $12, and added a Fast mode for Sol. The headline write-ups carried Luna's number; the card carried all three. The cheap tier got radically cheaper while the flagship edged down, which is what a price-performance frontier looks like when a vendor moves the whole curve at once.

The SPP read

GPT-5.6 launched July 9 with Sol as the new flagship at $5 per million input and $30 output on short context, the same list price as the GPT-5.5 it superseded; OpenAI positioned Terra, at half that, as the 5.5 analog. The August 22 move is what makes this an arc event: Sol's price went on a clock. The developer pricing page now lists $4 and $20 with the line 'GPT-5.6 Sol's promotional pricing is available at least through November 21, 2026.' A discount with a published end date is a promotion, not a price cut, and its withdrawal is an increase nobody has to announce. Google shipped the same instrument in the same window, with Gemini prices that revert on January 1, 2027.

The SPP read

One announcement, three named rungs: a small tier at a fraction of the mid, a mid at half the top, and the top holding the flagship rate. Model families used to ship as one price with a mini beside it; this ships as a designed ladder, and the design is the point. Named rungs give the meter a packaging layer: buyers self-sort by workload instead of negotiating, and the vendor gains rungs it can move independently later. The ledger shows that option getting exercised within weeks, when the bottom rung was cut by four fifths and the middle trimmed while the top held. A ladder is the set of places a future price change can land without touching the others, and this vendor built one on purpose.

The SPP read

A six-to-one output premium puts the price on generation, not context. For application vendors pricing on top of the API, the asymmetry matters more than the headline rate, because features that read long and write short carry a different unit cost than features that write long. Blended per-request pricing built on an average mix breaks the day the mix shifts. Price the direction your workload actually spends.

The SPP read

An 80% cut on the reasoning model, announced the same day a pro variant arrived at ten times the new rate, is one decision expressed as two prices. The cut moves the base model toward commodity territory and the simultaneous premium launch preserves somewhere for the spend to go. Vendors under capability deflation keep doing this: the price of last year's frontier falls fast, and a new rung appears above it so the falling floor never takes the ceiling down with it. Watch the gap between rungs rather than either price alone; the gap is the actual product strategy.

The SPP read

Microsoft matched OpenAI's July 30 price cuts in Azure Foundry effective August 1: GPT-5.6 Luna fell 80% to $0.20 per million input tokens and $1.20 per million output, and Terra fell 20% to $2.00 and $12.00. The launch prices had held for three weeks. What moved the Azure price was not Azure's costs but its parity posture: when the model vendor cuts, the platform reprices in step, which makes upstream rate volatility a direct term of the downstream contract.

The SPP read

The same migration, a third time in one earnings call, and here with a consumption number attached. Customer service leads because that is where work shifts from seated humans to metered agents most directly. The seat count stops tracking the work volume, and the meter bills the work that no longer arrives with a human attached; that is why the model is seats plus consumption rather than a conversion.

The SPP read

Three rungs published in one pass by the cloud reselling the models rather than the lab that made them. Sol sits at five times Luna's input rate with Terra between them, and the spacing is the product decision: a buyer choosing a tier is choosing how much reasoning to pay for, not which vendor to use. A rate card printed by a reseller also says the model has become a component with a distribution margin on it rather than a service bought from its maker.

The SPP read

A new product carrying a consumption meter at general availability, rather than starting per-seat and converting later. Vendors that meter from the start never have to run the conversion that breaks installed bases.

The SPP read

Free distribution reversed exactly where it created licensing pressure. The in-app Copilot Chat access Microsoft rolled out to unlicensed users by September 2025 was withdrawn from Word, Excel, PowerPoint, and OneNote starting April 15, 2026 (delayed to May 16 for some customers), for organizations above 2,000 seats, via admin-channel Message Center posts rather than a public announcement. Smaller tenants kept a throttled version. Free access at enterprise scale had become the alternative to buying licenses, and the retraction drew the line at the population where that substitution costs the most.

The SPP read

The commercial echo of the consumer move eleven months earlier: capabilities in, price up. Read together with the meter that arrived the same month the increase took effect, the design uses both instruments at once: the per-seat price absorbs the AI everyone gets, and the meter prices the usage that varies.

The SPP read

The free ceiling lifted where it markets best and held where it costs most. Microsoft's October 1, 2025 announcement gave Personal and Family subscribers unlimited image generation and unlimited Voice, in the Copilot app specifically, while the 60-credit monthly pool stayed in place for other Copilot actions inside the Office apps. Unlimited in the showcase surface, metered in the working surface: the allowance moved where usage is promotional, not where it is habitual.

The SPP read

The $19.99 tier reads as an addition and works as a consolidation. Microsoft 365 Premium launched October 1, 2025 above Personal at $9.99 and Family at $12.99, and the same announcement stopped selling Copilot Pro to new subscribers. The standalone AI add-on folded into the top of the plan ladder: one price now carries the productivity suite and the AI allowance that used to be bought separately, and the January 2025 price increase underneath it stayed.

The SPP read

The currency changed and the price did not. Effective September 1, 2025, Microsoft renamed the consumption unit for Copilot Studio agent billing from messages to Copilot Credits, with feature-based counting, while pack quantities and the $0.01 pay-as-you-go rate stayed exactly where they were. A unit rename with the math held still is the cleanest minted-currency specimen in the ledger: what changed is the language the meter speaks, and language is where conversion tables live.

The SPP read

The consumer surface ran the credits play the same week the commercial surface metered agents: a monthly allotment burned across apps. The credit is a surrogate unit: it lets the vendor tune the exchange rate between the subscription price and the underlying model cost without reprinting either.

The SPP read

The other half of the credits move: the first price increase in the product history, paired with AI now included. Bundling the model into the subscription converts an unpriceable cost into a priced one, and the increase funds the absorption.

The SPP read

Eighteen months before the seat-plus-meter earnings call the pattern was already set, and the line is not where the announcement's language suggests. Microsoft's own documentation splits agents by GROUNDING rather than by agency: an agent built on instructions and public web content is included, and an agent reaching SharePoint, tenant files or Graph connectors is metered. So a fully autonomous agent working public sources costs nothing, while a passive one reading your own documents bills. The meter tracks access to proprietary data, which is the input Microsoft pays to index and serve, not the independence of the thing consuming it.

The SPP read

The free half of the same announcement: chat for every commercial customer at no charge. Free access is not generosity; it seeds the habit that the metered agents monetize. The giveaway and the meter are one design.

The SPP read

The included allowance became a purchasable pool: Figma announced a credits subscription selling teams shared credits at a better rate, beside a dated commitment to start enforcing every seat's credit limit. Enforcement is what turns a denominated currency into a billable one, and the subscription builds the expansion path before the limits bite: the free-inside-the-seat stage ends, and the credit pool becomes a line a team budgets.

The SPP read

The meter got a buyer-side brake: pay-as-you-go billing for AI credits, capped at a spending limit the team sets itself, announced to run alongside the credits subscription or on its own. An unbounded meter with a buyer-set bound is the arc's compromise position: the vendor keeps consumption billing while the customer holds the cap, and the pairing moves Figma from allocating credits to operating a working usage market inside its plans.

The SPP read

The credit currency arrived inside the seat: when Figma's AI features left beta, every seat gained a monthly allowance of AI credits scaled by plan, with enforcement deliberately deferred for full seats. Bundling is the quiet first stage of a credits arc: the usage currency gets denominated and allocated while the invoice still says seats, so customers learn the meter's units before the meter ever bills.

The SPP read

The seat split into classes, but the classes differ on volume alone: Premium is the standard seat with a larger allowance, the meter collapsed into the packaging. A volume limit is a licensing-model construct, so this edition split spends the packaging layer doing the licensing layer's work and leaves packaging's own levers unpriced: priority access, longer-running sessions, the capability boundaries a team would actually pay to cross. What Premium sells is more of the same unit; what it could sell is different capabilities.

The SPP read

A seat became a meter: Bugbot moved from a flat per-seat subscription to usage-based billing for teams. The direction runs against the market's comfort story, since product lines inside one vendor can walk the arc both ways, and a bounded seat fee became an unbounded meter on exactly the workload that agents multiply, code review.

The SPP read

Two meters moved at once, and both on renewal rather than on the announcement. Teams agent usage left fixed per-request pricing for variable credits that rise with the work an agent does, the unit individual plans already carried, so one billing unit now spans every tier. Auto for individuals lost its unlimited status and started drawing on included monthly usage at token rates. An allowance that stops being unlimited is a bounded meter, and a bounded meter is a price.

Related: When usage-based pricing backfires: a field guide to AI metering gone wrong

The SPP read

Three weeks after the unit changed, the vendor paid for the transition itself: a public apology naming the failure precisely, unlimited had quietly meant only one mode, and full refunds for every unexpected charge in the window. Reaction events are part of the arc, and this one prices the cost of moving a meter faster than customers can re-learn it: the refund is the vendor buying back the variance it had shifted onto buyers without telling them.

The SPP read

The unit dropped beneath the request: five hundred countable fast requests became a pool of included frontier usage metered on compute, and the plan price stayed put while the thing it bought changed shape. The same announcement grew the ladder a $200 rung, Ultra at twenty times the included usage, priced as much for not having to watch the meter as for the capacity. A request was a unit the buyer could count; included-usage dollars are a unit the buyer must trust the conversion on. The rollout drew a public apology and refunds within three weeks, the price of changing the unit faster than customers could re-learn the meter.

The SPP read

On October 20, 2025 Box pushed its AI unit down the price list and out to the whole customer base. Business, Business Plus and Enterprise gained access to the Box AI APIs, Enterprise began receiving 1,000 AI Units with the plan, and AI Units became purchasable by any customer as one-time or recurring SKUs. The top two tiers changed not at all: Box says plainly that Enterprise Plus and Enterprise Advanced continue to include their 2,000 and 20,000 units. So the move is distribution, not repricing. A surrogate unit that had been an enterprise construct became the currency every Box customer can buy, which is how a unit stops being a feature of the top tier and starts being the thing the product is metered in. The mechanics underneath make that consequential: units are allotted per entitlement period, pooled across every chargeable AI feature at the enterprise level, and expire at period end with no rollover.

The SPP read

SAP split its AI portfolio's price in two, and Joule is the flagship instance. Base AI ships inside every SAP Cloud subscription at no additional cost and without limits; Joule Base went to a flat USD 0.00 covering navigational, informational, transactional, and simple analytical use (announced at Sapphire on May 20, 2025, with the formal bundling placed in July 2025). Premium AI monetizes through AI Units, which SAP's own FAQ calls a virtual currency, purchased as a single SKU and pooled across the landscape. The currency then splits by usage pattern: capabilities assigned to named users price per user per month in AI Units, tiered from 8 down to 1, while system-triggered capabilities meter on consumption in predefined per-service metrics, such as Document Grounding at 0.005 AI Units per record. Documents and records are what get counted; AI Units are what get billed; the conversion factors between them decide what anything costs while published prices hold still. The human seat is priced; the system's own activity is metered.

The SPP read

Nine months after minting AI Units, SAP started dealing them in. RISE with SAP premium commercial packages began carrying bundled AI Units as pilot credit for delivered AI innovations (Cash Application, Business Integrity Screening, Field Service Management), and Joule availability was extended to all RISE with SAP customers. Bundled credits are the distribution move in a minted-currency arc: the unit stops being something a customer decides to buy and becomes something their platform contract already holds. Adoption of the currency precedes consumption pricing on it.

The SPP read

SAP minted a currency in a document title. The AI Units supplement (October 2023) never defines an 'AI Unit': its operative terms are 'Capacity Unit' and 'Capacity Unit Value', and its usage-metrics clause names Capacity Unit as the contract's metric. The unit that gets counted is a multiplication: Units of Measure consumed times the per-service Capacity Unit Value carried in the companion AI Services List (December 2023), where Joule meters in Messages, one request plus its response, sold in 10,000-Message blocks at factor 7. Unused Capacity Units forfeit at the end of each Contract Year, and SAP can extend the rate table service by service as new AI Services ship. So 'AI Units' is the marketed name for a quantity the contract computes in Capacity Units through a rate table, the conversion factor is where the price actually lives, and every later Joule repricing runs through that table rather than through a price list.

The SPP read

The low end flipped to bundled-free while the top end went metered: Agentforce built into the entry Suites at no added cost, while enterprise editions carry credit meters and flat-fee agreements above them. Free at the bottom of a ladder is an acquisition instrument; the meter waits upstream.

The SPP read

Three purchase models for one product, available simultaneously with no forced migration: a consumption pool, per-profile SKUs, and the shared Flex Credit pool. Three meters for one product means the buyer, not the vendor, picks which unit the bill is denominated in.

The SPP read

The first modality-differentiated action rate to take effect: the same agent action costs fifty percent more when it is spoken. Cost-based rate differentiation inside a value-denominated currency tells you which way the meter is actually indexed.

The SPP read

The flat-fee agreement is the escape hatch from their own consumption model: when per-conversation and per-action pricing stalls enterprise commitment, a ceiling-priced multi-year restores it. The shared-risk framing does not answer the metering objection; it moves where the risk sits.

The SPP read

For the installed base the renewal-transition date is the real effective date. AI moved from a paid add-on into the higher editions, the standalone add-on was retired from sale, and the edition price rose twenty percent: a feature-level meter absorbed back into the edition, and the edition repriced for it.

The SPP read

A list action carrying an AI story, and the AI story is the one that matters. Six percent landed on the premium editions with AI named as the justification. Alongside it, a $125 per-user Agentforce add-on with UNMETERED employee agent usage, sold next to the per-action credit meter running elsewhere in the same catalog. Unmetered per-user agents beside per-action credits is a pricing surface admitting the metric question is not settled. The word doing the work is unmetered, and it is narrower than it sounds: the grant covers employee-facing workflows, and it is governed by fair use rather than by a meter: standard governor limits, API thresholds and platform performance guardrails still bound it. Unmetered here names a boundedness decision, not the absence of one. A seat price only holds when something bounds what the seat can consume, and here the bound moved from a counter the buyer can see to a policy the vendor administers.

The SPP read

Eight months after putting a price on the conversation, Salesforce put a price back on the seat. The Digital Labor add-ons and the Agentforce 1 edition introduce no new unit: they are per-user wrappers sold around the same meter, priced for buyers who could not underwrite a variable bill. The meter did not go away when the wrapper arrived. It moved behind a number the buyer could put in a budget.

The SPP read

Seven months after pricing the conversation, the meter moved down a level: actions at a dime, denominated in credits sold in hundred-thousand blocks, with licenses convertible into credits on demand. A finer unit gives the vendor pricing precision and gives the buyer a forecasting problem; the conversion bridge keeps both instruments live, licenses and credits, with a rate between them.

The SPP read

The arc opens with a price on conversation itself: agentic AI at $2 per conversation at general availability, volume discounts standard. A conversation is a surrogate for work the buyer cannot forecast, and every later Salesforce move on this ledger renegotiates what the meter counts: a conversation, an action, a credit, a resolution. The unit kept changing; what a customer is paying for never settled.

The SPP read

One year after minting it, Cognition took the ACU off the retail shelf. The April 2026 restructure retired Core and the $500 Team plan for a Free / $20 Pro / $200 Max / Teams-at-$80-minimum ladder where included usage counts against quota and overage bills in dollars, not ACUs. The minted unit didn't die (Enterprise still transacts in it); it retreated up-market to contracts where a normalized unit earns its complexity, while the team entry point fell from $500 to $80 and formerly free products (Ask Devin, DeepWiki, Devin Review) started metering the compute they consume.

The SPP read

The retreat from credits to quotas is the market's most repeated correction: a synthetic currency buyers could not translate, replaced by an allotment they can. Note that the seat did not disappear. It came back carrying a quota.

The SPP read

With Devin 2.0 the ACU went retail. A $20 entry minimum bought agent time pay-as-you-go at $2.25 per ACU, against $2.00 inside the $500 subscription's 250 included units, a committed-use spread priced like cloud capacity, not like a seat. The ACU is the market's only time-denominated minted unit: by the vendor's own definition one ACU is roughly 15 minutes of active Devin work, so the $20 entry buys about 2.25 hours of an engineer that isn't a person. Time-and-materials came back as a pricing model; the materials are compute, and the time belongs to the agent.

The SPP read

Cognition priced an autonomous engineer the way software used to be priced: one flat number. Devin went generally available in December 2024 at $500 a month for engineering teams, no seat limits, a monthly allotment of Agent Compute Units inside the subscription, custom Enterprise above it. The unit in the fine print is the story: an ACU is a normalized measure of the resources Devin consumes, so the flat subscription was a prepaid block of agent work from day one. The seat was never the axis (before this there was no public price at all, only request access), and what arrived in its place was work, denominated in a unit the vendor mints.

The SPP read

The charge moved onto the resolution: from this announcement Zendesk billed only where an AI agent closed a customer issue end to end, and not where a human had to finish it. Zendesk framed the move as displacing traditional pricing in a market it said should be measured by outcomes, and claimed the position first in its category. It is the opening beat of the arc, the moment a service vendor stopped selling the capacity to answer and started selling the answer.

The SPP read

Handing the buyer the definition of qualified is the opposite of the resolution meter, and it is the more durable choice. A metric the customer defines is a metric the customer cannot dispute at renewal.

The SPP read

A metric migration inside the same product. Resolution was legible to a buyer, outcome is broader and vaguer, so watch who holds the definition. The party that defines the outcome is the party that controls the invoice.

The SPP read

The first credible outcome meter in software: a price that fires only when the work succeeds, with a documented zero-charge path when it does not. The free-when-it-fails half is what made it defensible, and it is the half most imitators leave out.

The SPP read

A week after switching on its first live AI meter, Atlassian raised the price of the seat. Cloud list prices moved on October 15, 2025, notified to customers August 19: Jira, Confluence, Jira Service Management and the Teamwork Collection at 5 percent on Standard, 7.5 on Premium, 7.5 to 10 on Enterprise, with Bitbucket at a flat 10. Jira Standard went from USD 8.60 to USD 9.05 per user per month; Jira Enterprise at a thousand users from USD 155,000 to USD 166,000 a year. The stated reason is not cost. Atlassian's own FAQ answers 'why are you increasing the price' with 'the latest innovations now available,' and calls the result value-based price increases, listing AI features among the things delivered rather than among the things expensive. This is what absorption looks like when it comes due: the cost of serving generative AI arrives in the seat price, justified as value and never named as cost.

The SPP read

The seat price Atlassian retired came back with a meter attached. Rovo Dev, the coding-agent surface, went GA in October 2025 as its own SKU at USD 20 per user per month (the figure standalone Rovo gave up that April), but this one includes 2,000 monthly Rovo Dev credits and bills additional usage at USD 0.01 per credit on the next invoice. The beta had metered tokens and billed nothing; GA renamed the unit to credits and switched the bill on. The one place in the Rovo architecture where overage is charged today is the surface where the work is agentic rather than assistive: the human seats got absorption, the coding agent got a live meter, and the free tier simply stops at its limit rather than billing past it.

The SPP read

Six months after selling Rovo as a seat, Atlassian stopped selling the seat. The Team '25 announcement (April 9, 2025) folded Rovo into paid Jira, Confluence, and Jira Service Management subscriptions, Premium and Enterprise first, and the standalone pricing page went dark within weeks. What remained of the price is the meter that arrived in the same move: the support doc that counted Requests in January 2025 counted 'AI credits' by late April (70 per user per month on Premium, 150 on Enterprise, pooled across the organization), with overage billing named as a future event, not a current one. The add-on's price did not drop to zero; it moved into the tier. The credit allotment is what makes that possible: the seat price carries the AI cost, and the allotment defines how much it carries.

The SPP read

Atlassian launched Rovo the way the last decade priced software: a seat. The Team '24 Europe announcement (October 2024) made Rovo generally available as its own per-user subscription, and the pricing page's embedded rate plans show the shape precisely: USD 24.00 per user monthly for the first hundred users, graduating to USD 12.00 past twenty-five thousand, with annual prepaid bands working out to USD 20.00 per user per month through a hundred users. The consumption future was declared on day one: the same page's FAQ says usage beyond stated quotas will, in future, incur charges based on consumption. The debut was a seat price wrapped around a metered intent: the unit customers signed was per-user; the unit Atlassian reserved was per-use.

The SPP read

The enforcement date is its own move: the free window closed on schedule and the meter became mandatory, with one carve-out keeping early Prospecting activators free into 2026. That is staged conversion of an installed base, cohort by cohort, the same discipline GitHub skipped.

The SPP read

Once the currency exists, everything meterable migrates into it: five features began drawing on the same credit pool in a single announcement, each at its own exchange rate. Per-feature rates inside one shared unit is how a vendor tunes price without touching list, and the grandfather window is the rollout instrument, the meter arriving before the bill.

The SPP read

For the first time the included-credit allotment itself became the packaging differentiator: same hub list price, double the included credits at the platform tier. When the meter is the product, allotments start doing the work editions used to do.

The SPP read

The counter-move in the same keynote: the assistant and enrichment moved into the seat while agent work moved onto the meter. The boundary being drawn, assistance included and agent work metered, is the architecture decision here; the rates are secondary.

The SPP read

The mechanism is a currency migration of the installed base: new customers priced in credits from day one, the installed base following sixty days later with included allotments sized to the edition. The allotment is the cushion that lets a metering change land as packaging rather than a price increase.

The SPP read

Twenty-four days after pricing the effort, the vendor priced its own rollout: an acknowledgment that the transition 'did not meet our standards,' auto-refunds for a day of mis-computed checkpoint charges, and blanket credits for everyone on the new model. The remediation ran entirely through the price lever, credits and refunds, while the meter itself stood, so the allocation that caused the anger, customers carrying the cost of failed attempts, survived its own apology intact.

The SPP read

Replit priced the effort itself: a flat rate per checkpoint became a variable one that scales with how hard the agent worked, simple tasks below the old flat price and complex ones above it. Every other meter on this ledger counts inputs consumed or results delivered; this one prices the difficulty in between, the unit closest yet to the agent's own judgment of how much work the work required. Reception made the allocation visible: users found themselves billed for runs that failed or looped on the agent's own errors, and within a month the vendor said publicly that the transition had fallen short of its standards.

The SPP read

The meter dropped beneath the message: v0 stopped counting prompts and started metering the tokens inside them, converted to credits. Each plan fee became an included credit allowance, so the subscription now sets the floor and the meter charges past it. A message was a unit a buyer could count; a credit is a unit the vendor prices, and the conversion rate between them belongs to the vendor.

The SPP read

Adobe introduced two standalone Firefly subscription tiers priced on a generative credit unit that meters video and audio output, leaving image and vector features outside the credit meter entirely. The structure separates premium AI output types from unlimited-access features within the same plan, a bounded-credit model new to Adobe's Firefly commercial lineup. A higher-volume tier was announced as forthcoming, signaling further movement along this credit arc.

The SPP read

Adobe switched the generative credit meter on in two steps: the counters started running on November 1, 2023 with the limits waived for a limited time, and enforcement on select plans began January 17, 2024. A grace window between allocation and enforcement is a deliberate mechanism, not hesitation: it lets subscribers see the counter move, learn what a generation costs, and adjust before a cap can interrupt work, and it lets the vendor watch the consumption distribution before the first cap lands. The commercial shape did not change on either date; the September allocation, the slow-speed fallback and the purchasable pack were all in place. What changed was when the buyer could feel it.

The SPP read

Adobe's Text Effects feature, previously available at no credit cost within Firefly's generative credits system, shifted to a per-use credit charge on the stated date. The move converts a zero-cost allowance into a billable consumption event, extending the generative credits meter to cover a feature that had been effectively free. This marks the first metered boundary applied to Text Effects on Adobe's credits arc.

The SPP read

Two acts landed on the same day, and the arc should hold them apart even where they coincide. The first is the minting of a new currency: Generative Credits, a vendor-controlled unit that stands between the subscriber and the metered generation underneath, priced and redefined at Adobe's discretion. The second is the seeding: that currency was allocated into every existing paid Creative Cloud, Firefly and Express plan at once, with slower generation when the allocation runs out and a purchasable pack from November. Adobe raised list prices the same day and said why, new costs associated with generative AI content creation, which makes the seeding a repricing and not a gift. Other vendors have done these in sequence, minting first and seeding later; here the mint and the seed are one announcement, which is exactly what makes the later enforcement date (row 518) its own move rather than part of this one.

The SPP read

Adobe introduced a published credit rate for four generative AI capabilities in its Firefly toolset, tying consumption to a per-use credit meter rather than bundling usage invisibly into subscription tiers. The credit unit is bounded by an image-size threshold, establishing a two-tier meter structure from the outset. This is Adobe's first public articulation of how generative credits are spent against specific features, placing it on the tokens-to-credits arc for the first time.