AI Services and Token Capacity

How to read this chapter

Paying for tokens does not secure the capacity to use them. An AI service can stay inside its monthly budget and still have its requests refused on Monday morning, and the bill will not say why. The planning question is what the organization can deliver through its buying arrangements, at the required quality and within its budget. Eight parts walk through it in order.

1. Before any theory: the four ways to buy tokens — direct API, platforms, negotiated agreements and provisioned capacity, often in combination.
2. Token capacity is a different kind of capacity — how hosted token capacity differs from a server fleet, and the three forces that bound an AI service: scarcity, pricing and budgets.
3. The model product list and how to read it — what the names and prices mean, the effort and speed dials, and which models a subscription unlocks.
4. Sizing supply for the busy interval — the shared, priority and reserved rungs providers sell, and the supply and price tests a reservation must pass.
5. Aggregation: making a multi-model fleet governable — the platforms that unify many providers, and why direct consumption remains the alternative.
6. The reality of consumption at enterprise scale — subscriptions, application usage and separate buying entities alongside API traffic.
7. Restricted model access — the models not every organization can use, and the fallback they require.
8. Budgets and allowances, Measure the efficiency of the models, not just the spend, What the capacity team owns and Recommended practices — the controls, the efficiency measurement, the capacity record and the practices, followed by a Summary of the chapter.

Three lessons run through the chapter. Supply, price and budget are separate questions: a rate limit permits traffic without holding capacity for it, a spending commitment changes what supply costs without holding any, and a budget control decides what happens at a limit without making the busy interval feasible. The bill is the sum of many purchases whose rates move, so a forecast must separate price from workload. And tokens are what the provider counts, not what the business wants: the measure that justifies the spending is cost per accepted task.

Before any theory: the four ways to buy tokens

Prices can change during a planning cycle. On July 30, 2026, OpenAI cut GPT-5.6 Luna prices by eighty percent and Terra prices by twenty percent, attributing the reductions to improved serving efficiency.⁸ Keep the effective date with each rate.

Token capacity is a different kind of capacity

For the web pool in Chapter 2, we could connect sixty-four workers to measured request rates, headroom, failure groups and queue behavior. Hosted AI services still run on physical infrastructure, but the buyer usually plans against an API and a contract rather than an inspectable worker fleet. The commercial limits join the technical ones.

The FinOps community calls this territory token economics, or tokenomics, and Mike Fuller of the FinOps Foundation applies the term to the whole chain behind a token, from energy and silicon to pricing and margin.¹⁰ This book’s working definition is narrower: the relationship among AI workload design, consumption, supplied capacity, cost and accepted work.

Open full-size diagram ↗

The model product list and how to read it

Before any price, a capacity planner meets a product list: dozens of models from a dozen providers, each with its own rate card, and vendors that also sell their own judgment about how much thinking each task deserves.

A few dated examples show why the model list matters. OpenAI launched GPT-6.1 Sol on September 29, 2026, at $2 per million input tokens and $10 per million output tokens, with cached input at $0.10, which OpenAI positions as near-Astra results at one-fifth of Astra’s standard prices. Anthropic launched Opus 5.5 on September 22, 2026, at $4/$20, twenty percent below Opus 5’s corresponding rates; Anthropic also reported lower token consumption on its tested workloads.¹¹¹² Anthropic followed with Claude Sonnet 5.5 on September 28, 2026, at $2/$10, the same rates as Sonnet 5, reporting thirty percent faster output and up to thirty percent lower cost per task; ¹³¹⁴ Google announced Gemini 4 Argon on September 30, 2026, at an introductory rate of $2/$10 with cached input at 95 percent off, rising to $4/$20 after the introductory period; the introductory rate itself is part of the offering, and Google has since published its expiry: the introductory rates run through November 30, 2026.¹⁵ On October 7, 2026, Anthropic cut Sonnet 5.5’s cached-input rate to $0.10 per million tokens, half its launch rate, and launched Claude Haiku 5.5 at $0.10/$0.50 per million input/output tokens for prompts up to 100,000, rising to $0.50/$2.50 above them; Haiku 5.5 is the first Haiku-class model with an adjustable effort setting.¹⁶ OpenAI completed the GPT-6.1 family on October 26, 2026, launching Terra, the mid-tier model of the series, at $1/$5 per million input/output tokens with cached input at $0.05, standard rates from launch that OpenAI states are half those of GPT-5.6 Terra and GPT-6.1 Sol.¹⁷ Google announced Gemini 4 Carbon on October 28, 2026, at the same introductory rate as Argon, $2 per million input tokens and $10 per million output tokens with cached input at 95 percent off, rising to $4/$20 after the introductory period; Google has not published an expiry date for that window, and Carbon is not yet on any developer surface.¹⁸ Those rates are snapshots, not quotes for the reader’s next purchase, and the models are not interchangeable simply because both appear in a frontier category. A useful comparison preserves the model version, processing mode, context tier, cache treatment and outcome standard alongside the rate.

Effort changes consumption as well as latency. OpenAI, Anthropic and Google expose controls for reasoning or thinking on supported models, but the permitted settings and their effects differ. OpenAI bills reasoning tokens as output even when the answer does not show them, so an unchanged per-token rate does not imply an unchanged bill per task, and Anthropic invalidates prompt caching when top-level effort changes between requests.¹⁹ Artificial Analysis measured Gemini 3.8 Flash at $0.58 per task on its high-reasoning benchmark workload, versus $0.40 for its predecessor: a 45 percent increase, since (0.58 − 0.40) / 0.40 = 0.45. Its report attributes the difference to more output tokens and more agentic turns despite unchanged token prices.²⁰ That is evidence about the published test, not a forecast for our application, which needs the settings tested on representative work.

Processing speed may also be a separately priced choice. At GPT-6 Astra’s launch, OpenAI advertised Fast mode at twice the Standard price for up to twice the processing speed, and by October 3, 2026 its published API price list carried a second rung: an Ultrafast tier at six times the Standard rate, $60/$300 per million input/output tokens for Astra at short context.²¹²² Treat each rung as a vendor offering to test: faster generation does not necessarily halve end-to-end task time, which can include tools, queues and human decisions, and the sixfold rate applies to every token the tier processes, however much of the task’s time is spent waiting on something else.

Access depends on the product and agreement. A subscription can make models or features available that a lower tier lacks, while API access has its own account limits and billing. At its announcement, Gemini 4 Argon was reachable only through a vetted program, with paid API customers and subscribers named for a later phase and no published dates. That phase arrived on October 13, 2026, when Argon opened to paid API customers and to Google AI Ultra and Pro subscribers.²³ Do not transfer a subscription entitlement into an API plan without checking it.²⁴ A release can also split one model family by plan: on October 12, 2026, Meta put Muse Wonder on its Model API at $4/$20 per million input/output tokens with cached input at $0.20, while the stronger Muse Wonder Full went only to Muse Code subscribers, not to the API.²⁵

Open full-size diagram ↗

The figure’s cards restate the dated examples and the Gemini 3.8 Flash test above, and the Mythos 5.1 terms described later, with the same sources.

Sizing supply for the busy interval

An affordable month and a feasible busy interval require different measurements. When a support assistant hits its token limit on Monday morning, its requests wait, retry or fail, depending on how the application handles them.

What the organization can buy for Monday morning comes in rungs, and the major providers sell much the same ladder. Standard pay-per-token service runs on capacity shared across customers. OpenAI rejects requests within the per-minute limits when traffic ramps too fast and counts failed requests against the limit, so retries spend the headroom they wait for; Google treats a 429 on its standard tier as contention for a shared resource, not an exhausted quota. A priority tier, such as OpenAI’s Fast mode or Foundry’s priority processing, buys a better place in the shared queue at a premium per token, and Microsoft and Google both move priority requests back to standard processing when priority capacity runs short.²⁶ Reserved throughput holds supply: when OpenAI cannot serve all traffic, it rejects other traffic first.²⁷ Batch and flex processing sit below the ladder, trading delay for a lower price.

A service that cannot defer work needs enough throughput for its design peak, with a stated allowance for uncertainty; a batch process that can wait may use the same capacity over a longer interval. The worst observed minute is useful evidence but not automatically the purchasing target, and the interval that matters can be shorter than a minute: Google enforces provisioned throughput over windows as short as one to five seconds for its largest allocations, and warns that a burst can spill to pay-as-you-go while one-minute average utilization stays below 100 percent.²⁸

Foundry shows what holding supply involves. A provisioned throughput unit, or PTU, is not a fixed token rate: throughput per PTU varies by model, output consumes more of it than input, and cached input none, so a deployment holds a different amount of work when the prompt mix or cache hit rate changes. Approved quota permits a deployment but holds nothing. The deployment holds capacity and bills hourly, busy or idle; scaling it down or deleting it returns capacity to the region with no guarantee of getting it back. A reservation discounts the hourly charge without guaranteeing capacity and outlives a deleted deployment, so Microsoft recommends deploying first.²⁹³⁰

A reservation therefore has to pass two tests that are easy to merge: a supply test, whether it covers the busy interval at the required latency, and a price test, whether paying for it through every quiet hour costs less than buying the same work on the shared rungs. A reservation sized for Monday morning is paid for through Sunday night. OpenAI’s Reserved Tier is bought in dollars per minute and drawn down at list price, so a price change alters how many tokens it holds, and each minute’s excess is billed at pay-as-you-go rates.³¹ That spillover, which Foundry and Google also offer, supports reserving the steady base and buying the peak on a shared rung, with the peak’s risk returned to the shared pool.³²³³

The workload owner must decide how much waiting is acceptable. Some drafting work can queue; other drafting is part of a customer interaction with a short deadline. We can overbuy by reserving capacity for work that could wait, or underbuy by assuming that every request can wait. Neither answer follows from calling the application an assistant.

Open full-size diagram ↗

Aggregation: making a multi-model fleet governable

Even one application may run several models: a smaller one for retrieval queries, a more capable one for drafts and another for checks. Across teams and provider accounts, those choices become a portfolio of deployments and agreements. A gateway or aggregation platform can combine model access, usage accounting and routing; direct provider access remains the alternative where an existing agreement or feature matters. Compare actual quotes and fees rather than assuming either route is cheaper, and if the organization uses both, the cost view must include both.

Microsoft Foundry makes deployment type part of the decision: Standard, Provisioned and Batch serve different traffic patterns, with different processing-location options, and a resource’s location alone does not specify where inference is processed.³⁴ Volume II, Chapter 21 takes up what that means for data residency. For a customer-facing assistant and an overnight document job using the same model, I would compare their latency and placement requirements before assuming they should share one purchasing arrangement.

Google’s Gemini Enterprise Agent Platform lets a request header choose between provisioned throughput and pay-as-you-go. By default, a request beyond the remaining provisioned quota runs entirely as pay-as-you-go and is billed at that rate; with the dedicated-only header it is refused with a 429.³⁵ The default favors completing the work over controlling its cost.

OpenRouter combines provider access, usage accounting and routing behind one API and credit balance, passing provider prices through and charging a fee when credits are bought. Stripe announced an agreement to acquire it on August 19, 2026.³⁶

Verify team attribution, key limits, routing, data handling and failure behavior before choosing one as the place to enforce policy, and distinguish a gateway rejection from an upstream limit when investigating a failure. A gateway cannot observe direct calls, subscription sessions or cache details an upstream provider does not expose; Chapter 8 begins with that measurement boundary.

The reality of consumption at enterprise scale

Large organizations do not have one AI bill; they have several, and the same workload can touch all of them.

Consider a hypothetical organization with two thousand employees using AI subscriptions, a development team making a thousand API requests a day, contractors using separate accounts, and an experimental open-weight deployment on rented hardware. Seat counts, metered API usage and GPU hours describe different obligations. Subscription reporting need not match API token accounting, and a flat subscription price does not imply unlimited use, so check plan-specific limits and additional usage charges.³⁷

Start with a spending view reconciled to the invoices, then ask each program owner what work it supports. A useful API workflow does not establish that every subscription is useful, or that a contractor’s separate tool duplicates it.

Separate buying entities add another complication. Procurement may hold a negotiated contract while a business unit pays public rates through another account, and security approval may differ by workload, so the same model name does not establish the same price or permission.

Restricted model access

Some of the most capable models are not available to most organizations, because of the safety controls and government approvals attached to them.

Anthropic’s Mythos 5.1, introduced on September 1, 2026, is available only to vetted organizations through trusted access programs, currently a set of US organizations, and requires accepting 30-day data retention for safety monitoring by default. Access can also change from outside the contract. Anthropic posted on June 12, 2026 that Mythos 5 was unavailable, and on July 1 that export controls had been lifted and access restored for a set of US organizations following the US government’s approval.³⁸

Check access before designing a workload around a model: approval for one group or use case may not extend to another, and general availability does not settle the organization’s own security requirements. Name the fallback and test it before it is needed; a fallback found during an interruption of nearly three weeks is a migration under deadline.

Budgets and allowances

A budget states how much the organization intends to spend; enforcing it takes separate mechanisms. Choose each by the consequence of reaching it: the same cutoff can be sensible for an experiment and unacceptable for a customer-facing workflow.

• Prepaid credits fund a declining balance, and the cash leaves before the work is done. Check auto-recharge, expiry, refund terms and enforcement delay; OpenRouter, for example, can expire unused credits after a year and documents that a balance can go negative.³⁹ Decide what the application does when further requests are refused.
• Per-team and per-key allowances allocate spending authority. State whether crossing an allowance generates an alert, rejects new work or needs an exception.
• Provider spend caps can block further billable usage within a defined scope. Google announced its public preview in July 2026 and calls the cap non-destructive, meaning data and resources are not deleted and unrelated services keep running; it does not promise that the capped service keeps processing paid work. The block stays until someone lifts it, and it also pauses usage covered by Provisioned Throughput, which Google sells as a fixed-cost, fixed-term subscription, so a cap can stop work already paid for.⁴⁰
• Commitments and reservations can trade flexibility for a lower rate. Some products also reserve throughput, but a spending commitment alone does not: OpenAI’s annual commitments can be spent on any of its models and delivery mechanisms, which makes them money rather than held supply.⁴¹ Volume II, Chapter 20’s distinctions among price, supply, term and overage apply here.

Azure API Management provides one concrete enforcement point. Its llm-token-limit policy applies a token rate, a period quota or both to a counter, best keyed on something the caller cannot simply assert, such as the API Management subscription issued to each team:

<llm-token-limit counter-key="@(context.Subscription.Id)"
    tokens-per-minute="50000"
    token-quota="20000000" token-quota-period="Monthly"
    estimate-prompt-tokens="false" />

The numbers are illustrative. An exceeded rate returns HTTP 429 and an exceeded quota 403; counters are local to each gateway, streaming counts are estimated, and concurrent requests can overshoot before the limit takes effect.⁴² That is control over one gateway’s traffic, not an enterprise-wide dollar ceiling: prices vary by model and token category, and tool charges sit outside the counter. Test the refusal and retry behavior as part of the application, reading the error code as well as the status: at OpenAI, an exhausted credit balance and a hard spend limit also return 429.⁴³

Anthropic’s task budgets give a model an advisory token allowance for a full agentic task, which the documentation distinguishes from an enforced limit.⁴⁴ A hard spending ceiling still needs controls outside the model, including limits on repeated calls and tools, that survive a change of provider.

Routine consumption inside allowances should proceed without ceremony; this is Volume II, Chapter 17’s relationship discipline applied to a new resource. Exceptional decisions — a new commitment, a tier change, a routing policy with acceptance risk — should reach somebody able to weigh them, and the review in Volume II, Chapter 22 should see consumption, attribution and the peak together. Ordinary use must not become exhausting: the surest way to lose visibility is to make people avoid the process that provides it.

Measure the efficiency of the models, not just the spend

Token consumption alone cannot tell whether spending is going well. A bill that rose because the workload grew looks identical to one that rose because the work became less efficient, and Fuller’s example shows a bill rising while efficiency improves.⁴⁵ The book’s measurement for an enterprise deployment is deliberately plain:

Open full-size diagram ↗

Its reciprocal, tokens per accepted task, carries the same information read as intensity, and in that form it joins the bill. For one task class over one period, with failed attempts included in the tokens,

$$\text{token cost} = \text{accepted tasks} \times \frac{\text{tokens}}{\text{accepted task}} \times \frac{\text{token cost}}{\text{token}}.$$

The three factors separate what a bill conflates: more accepted work is volume, more tokens per accepted task is intensity, and a higher blended cost per token is price, including a shift in the input and output mix or the cache hit rate. If intensity rises while accepted work stays flat, investigate prompt growth, retries, tool behavior, cache treatment, model changes and shifts in task difficulty or acceptance criteria.

Cost per accepted task divides the total cost of running the workload by its accepted tasks, keeping failed attempts in the numerator and adding what the identity leaves out: attributable subscriptions, committed capacity and other fixed charges. Use the period’s amortized share of a commitment, Chapter 6’s effective cost, or a prepaid reservation will make one month look ruinous and the next free.

Keep the expensive tail visible alongside the average; an occasional task with repeated calls can matter to the budget and the busy interval even when most tasks are routine. These measures describe how well work is done inside the current design, not whether the design, or the work, should exist. Chapter 8 takes up that question, along with the call-level model and a benchmark-weighted index that does not convert tokens into accepted work.

What the capacity team owns

The capacity team connects commitments, measured demand and accepted work, with Procurement, Finance, engineering and the business owners, and gets each tradeoff decided by somebody authorized to make it. A single subscription line is too little evidence; the decision needs a record for each workload that keeps supply, price and budget in separate rows:

Before recommending the next stage of adoption, I would want this record complete for the workloads involved and every open item owned. Chapter 8 supplies the measurements behind its last rows.

Recommended practices

I would carry the following practices into the review:

Treat AI consumption as many programs, not one. Build the reconciled view of subscriptions, API traffic, negotiated agreements and contractor seats before planning against any one program’s numbers.

Watch the model releases the way you watch the bills. A new version can change the bill and the answer with no event on the cost side. Anthropic’s Opus 5.5 defaults to medium effort where earlier Opus models default to high, and Sonnet 5.5 recalibrated its levels, so an unchanged request can buy different thinking after an upgrade.⁴⁶ Re-baseline the forecast when a release changes something it assumed.

Match the benchmark to the step. Read its methodology, test conditions and funding disclosures, as Chapter 8 describes; independence alone does not establish applicability.

Audit what the work is for, at least once a year. Chapter 8’s purpose questions apply to a year-old API program as much as to a new one.

Summary of the chapter

The next purchase needs a joined view of demand, access, supply, price and accepted work. The practical checks are:

• Reconcile the buying paths before planning against any one of them.
• Keep the date and model version with every rate.
• Test the busy interval against held supply, and any reservation against price.
• Name each budget control’s behavior at its limit.
• Measure accepted work, with failed attempts left in the cost.

Chapter 8 follows this bill call by call to the tasks it served, Chapter 9 follows the service past the model call into memory, tools and harnesses, and Chapter 10 takes up the serving, profiling and training questions that belong to hardware.

---

Notes and sources

1. OpenRouter, FAQ, https://openrouter.ai/docs/faq, and “API Credit & Rate Limits,” https://openrouter.ai/docs/api_reference/limits, checked October 4, 2026; Stripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage,” August 19, 2026, https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter. Product terms and transaction announcement; no transaction price is inferred.↩︎

2. OpenAI, “GPT-6 Astra: A new generation of intelligence,” September 3, 2026, https://openai.com/index/gpt-6-astra/. Launch Standard rates $10/$50 per million input/output tokens; Fast mode twice the Standard price and advertised at up to twice the speed. On OpenRouter, which reports weekly route traffic, Astra was among the most-used routes through early October 2026 and processed about 98.6 billion tokens on October 9, 2026, moving down the popularity list.↩︎

3. Microsoft, “Understanding deployment types in Microsoft Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types, checked October 3, 2026. Deployment categories, supported processing boundaries and model-specific availability; managed-compute deployments are a separate option.↩︎

4. Google Cloud, “Use Provisioned Throughput,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/use-provisioned-throughput, checked October 4, 2026. Default spillover, the dedicated-only 429, enforcement windows and spillover hidden by one-minute averages. Renamed from Vertex AI on April 22, 2026, per the release notes: https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes.↩︎

5. OpenRouter, FAQ, https://openrouter.ai/docs/faq, and “API Credit & Rate Limits,” https://openrouter.ai/docs/api_reference/limits, checked October 4, 2026; Stripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage,” August 19, 2026, https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter. Product terms and transaction announcement; no transaction price is inferred.↩︎

6. Microsoft, “Provisioned throughput billing and cost management,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput-billing, checked October 4, 2026. Deployed throughput incurs hourly charges whether or not it is used; a reservation is a discount on those charges and does not guarantee capacity.↩︎

7. Google Cloud, “Use Provisioned Throughput,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/use-provisioned-throughput, checked October 4, 2026. Default spillover, the dedicated-only 429, enforcement windows and spillover hidden by one-minute averages. Renamed from Vertex AI on April 22, 2026, per the release notes: https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes.↩︎

8. OpenAI, “Advancing the price-performance frontier with GPT-5.6,” July 30, 2026, https://openai.com/index/advancing-the-price-performance-frontier-with-gpt-5-6/. Vendor announcement of the Luna and Terra price cuts.↩︎

9. OpenAI, “Rate limits,” https://developers.openai.com/api/docs/guides/rate-limits, and “Error codes,” https://developers.openai.com/api/docs/guides/error-codes; Google Cloud, “Standard PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/standard-paygo, and “Priority PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/priority-paygo; Microsoft, “Enable priority processing for Microsoft Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/priority-processing; AWS, “Service tiers for optimizing performance and cost,” https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html; checked October 3, 2026. Documented ramp, retry, spend-limit and downgrade behavior; shared pool sizes are not published.↩︎

10. Mike Fuller, “FinOps and Tokenomics: How do they fit together?”, FinOps Foundation, August 25, 2026, https://www.finops.org/insights/finops-tokenomics-fit-together/, checked October 4, 2026. Tokenomics spans the production chain and complements FinOps for AI; its example bill rises while tokens per resolved outcome fall.↩︎

11. OpenAI, “GPT-6 Astra: A new generation of intelligence,” September 3, 2026, https://openai.com/index/gpt-6-astra/. Launch Standard rates $10/$50 per million input/output tokens; Fast mode twice the Standard price and advertised at up to twice the speed. On OpenRouter, which reports weekly route traffic, Astra was among the most-used routes through early October 2026 and processed about 98.6 billion tokens on October 9, 2026, moving down the popularity list.↩︎

12. Anthropic, “Introducing Claude Opus 5.5,” September 22, 2026, https://www.anthropic.com/claude-opus-5-5. Launch input/output rates of $4/$20 per million tokens, versus Opus 5 at $5/$25; lower task-cost results are Anthropic’s reported tests.↩︎

13. Anthropic, “Introducing Claude Sonnet 5.5,” September 28, 2026, https://www.anthropic.com/claude-sonnet-5-5. Priced the same as Sonnet 5 at $2/$10 per million input/output tokens; cache reads $0.20 and cache writes $2.50 per million; reads were halved to $0.10 per million on October 7, 2026 (see the Haiku 5.5 footnote). The 30-percent speed and cost figures are Anthropic’s reported tests. Vendor announcement; re-verify before quoting a rate.↩︎

14. OpenAI, “Introducing GPT-6.1 Sol,” September 29, 2026, https://openai.com/index/introducing-gpt-6-1-sol/. Introductory API rates at announcement, held as Standard through the checked window: $2/$10 per million input/output tokens, cached input $0.10 — 95 percent off standard input and half of GPT-6 Sol’s cached rate. Available in ChatGPT Work, Codex and the API (gpt-6.1-sol); not yet in Chat. The “Ultrafast” option announced that week, advertised at up to 8x faster generation in Codex, appeared in OpenAI’s published API price list by October 3, 2026 at six times the Standard rate ($60/$300 per million input/output tokens for GPT-6 Astra at short context, see the GPT-6 Astra footnote). Speed figures remain vendor-advertised. Vendor announcement; re-verify before quoting a rate.↩︎

15. Google, “Gemini 4 Argon: our next era of frontier intelligence,” September 30, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-4-argon. Announced introductory rates of $2/$10 per million input/output tokens with cached input at 95 percent off, and $4/$20 after the introductory period; output-token limit raised from 64,000 to 1,000,000. Availability at announcement limited to a vetted program, with paid API customers and Google AI Ultra subscribers named for a later phase without dates. Google published the expiry on October 13, 2026 in its availability announcement (see the Argon general-availability footnote): introductory rates through November 30, 2026, then $4/$20. Vendor announcement; dates and terms to re-verify before quoting a rate.↩︎

16. Anthropic, “Introducing Claude Haiku 5.5,” October 7, 2026, https://www.anthropic.com/claude-haiku-5-5. Announced at $0.10/$0.50 per million input/output tokens for prompts up to 100,000 tokens and $0.50/$2.50 above them; cache reads $0.01, $0.05 over 100,000; cache writes $0.125 per million. The same announcement halved Sonnet 5.5’s cache reads to $0.10 per million and added a monthly API credit for Max and Team subscribers. First Haiku-class model with an adjustable effort setting. Vendor announcement; re-verify before quoting a rate.↩︎

17. OpenAI, “Introducing GPT-6.1 Terra,” October 26, 2026, https://openai.com/index/introducing-gpt-6-1-terra/. Standard API rates of $1 per million input tokens, $0.05 per million cached input tokens and $5 per million output tokens, which the announcement states are 50 percent lower than GPT-5.6 Terra’s and GPT-6.1 Sol’s; the OpenAI API price list also carries a $1.25 per million cache-write rate. Available in ChatGPT Work and Codex to Plus, Pro, Enterprise and Edu plans and through the API as gpt-6.1-terra; no introductory window is stated. OpenRouter listed the route, a :batch variant and a Pro reasoning-mode variant the same day. Vendor announcement; re-verify before quoting a rate.↩︎

18. Google, “Introducing Gemini 4 Carbon,” October 28, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-carbon. Announced introductory rates of $2/$10 per million input/output tokens — the same as Argon — with cached input at 95 percent off the input price, and $4/$20 after the introductory period; no expiry date is published. Access at announcement went to a subset of cyber defenders and trusted partners through Google’s Fairwind Program, with availability to developers, businesses and consumers promised but not dated. Carbon was not listed on OpenRouter as of October 30, 2026. Vendor announcement; re-verify before quoting a rate.↩︎

19. OpenAI, “Reasoning models,” https://developers.openai.com/api/docs/guides/reasoning; Anthropic, “Effort,” https://platform.claude.com/docs/en/build-with-claude/effort; Google, “Gemini thinking,” https://ai.google.dev/gemini-api/docs/thinking, checked October 4, 2026. Model-specific controls and token accounting; Anthropic’s per-model defaults and cache invalidation on top-level effort changes.↩︎

20. Artificial Analysis, “Google has released Gemini 3.8 Flash,” September 2, 2026, https://artificialanalysis.ai/articles/gemini-3-8-flash. Reports $0.58 versus $0.40 per task, about 30% more output tokens and more agentic turns on its benchmark. The rounded dollar figures imply 45% growth; the article calls it approximately 40%.↩︎

21. OpenAI, “GPT-6 Astra: A new generation of intelligence,” September 3, 2026, https://openai.com/index/gpt-6-astra/. Launch Standard rates $10/$50 per million input/output tokens; Fast mode twice the Standard price and advertised at up to twice the speed. On OpenRouter, which reports weekly route traffic, Astra was among the most-used routes through early October 2026 and processed about 98.6 billion tokens on October 9, 2026, moving down the popularity list.↩︎

22. OpenAI, “Pricing,” https://developers.openai.com/api/docs/pricing, checked October 3, 2026. Fast at twice Standard; Ultrafast, listed for GPT-6 Astra only, at six times; notes Priority processing renamed Fast mode on July 30, 2026.↩︎

23. Google, “Gemini 4 Argon: Built to solve your hardest problems,” October 13, 2026, https://blog.google/innovation-and-ai/models-and-research/gemini-models/introducing-gemini-4-argon. Argon opened to paid API customers in AI Studio and the Gemini platform (model ID gemini-4-argon), to Google AI Ultra subscribers in Antigravity, and to Ultra and Pro subscribers in the Gemini app and AI Mode in Search. The post prices the launch at $2/$10 per million input/output tokens with cached input at 95 percent off, states in a footnote that the introductory period expires on November 30, 2026 with $4/$20 applying after it, and describes a 1M unified input-output token limit via the Gemini API. Google Cloud announced the same model ID and rates the same day. Vendor announcement; re-verify before quoting a rate.↩︎

24. OpenAI, business pricing and plan comparison, https://openai.com/business/pricing/, checked October 3, 2026. Seat types carry different usage allowances, and more access can be bought as credits; subscription billing is distinct from API token accounting.↩︎

25. Meta, “Muse Wonder: A Frontier Model For Personal & Professional AI Agents,” October 12, 2026, https://research.meta.ai/blog/muse-wonder/. Muse Wonder (identifier muse-wonder-1.0) is on the Meta Model API at $4/$20 per million input/output tokens, cached input $0.20, with web search at $2.50 per 1,000 queries; Muse Wonder Full is available only to Muse Code subscribers and is not offered through the API. The same rates appear on OpenRouter’s meta/muse-wonder route. Vendor announcement; re-verify before quoting a rate.↩︎

26. OpenAI, “Rate limits,” https://developers.openai.com/api/docs/guides/rate-limits, and “Error codes,” https://developers.openai.com/api/docs/guides/error-codes; Google Cloud, “Standard PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/standard-paygo, and “Priority PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/priority-paygo; Microsoft, “Enable priority processing for Microsoft Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/priority-processing; AWS, “Service tiers for optimizing performance and cost,” https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html; checked October 3, 2026. Documented ramp, retry, spend-limit and downgrade behavior; shared pool sizes are not published.↩︎

27. OpenAI, “Reserved Tier for API Customers,” https://openai.com/api-reserved-tier/, and “Scale Tier for API Customers,” https://openai.com/api-scale-tier/, checked October 3, 2026. Per-model dollars-per-minute reservations drawn at list price, reserved traffic rejected last, annual commitments usable on any model. Vendor pages; contract terms are unpublished.↩︎

28. Google Cloud, “Use Provisioned Throughput,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/use-provisioned-throughput, checked October 4, 2026. Default spillover, the dedicated-only 429, enforcement windows and spillover hidden by one-minute averages. Renamed from Vertex AI on April 22, 2026, per the release notes: https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes.↩︎

29. Microsoft, “Provisioned throughput billing and cost management,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput-billing, checked October 4, 2026. Deployed throughput incurs hourly charges whether or not it is used; a reservation is a discount on those charges and does not guarantee capacity.↩︎

30. Microsoft, “Provisioned throughput for Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput, checked October 4, 2026. PTU throughput by model, output weighting and cache treatment; quota, capacity and reservation discounts as distinct; released capacity may not return; optional spillover.↩︎

31. OpenAI, “Reserved Tier for API Customers,” https://openai.com/api-reserved-tier/, and “Scale Tier for API Customers,” https://openai.com/api-scale-tier/, checked October 3, 2026. Per-model dollars-per-minute reservations drawn at list price, reserved traffic rejected last, annual commitments usable on any model. Vendor pages; contract terms are unpublished.↩︎

32. Microsoft, “Provisioned throughput for Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/provisioned-throughput, checked October 4, 2026. PTU throughput by model, output weighting and cache treatment; quota, capacity and reservation discounts as distinct; released capacity may not return; optional spillover.↩︎

33. Google Cloud, “Use Provisioned Throughput,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/use-provisioned-throughput, checked October 4, 2026. Default spillover, the dedicated-only 429, enforcement windows and spillover hidden by one-minute averages. Renamed from Vertex AI on April 22, 2026, per the release notes: https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes.↩︎

34. Microsoft, “Understanding deployment types in Microsoft Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/foundry-models/concepts/deployment-types, checked October 3, 2026. Deployment categories, supported processing boundaries and model-specific availability; managed-compute deployments are a separate option.↩︎

35. Google Cloud, “Use Provisioned Throughput,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/use-provisioned-throughput, checked October 4, 2026. Default spillover, the dedicated-only 429, enforcement windows and spillover hidden by one-minute averages. Renamed from Vertex AI on April 22, 2026, per the release notes: https://docs.cloud.google.com/gemini-enterprise-agent-platform/release-notes.↩︎

36. OpenRouter, FAQ, https://openrouter.ai/docs/faq, and “API Credit & Rate Limits,” https://openrouter.ai/docs/api_reference/limits, checked October 4, 2026; Stripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage,” August 19, 2026, https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter. Product terms and transaction announcement; no transaction price is inferred.↩︎

37. OpenAI, business pricing and plan comparison, https://openai.com/business/pricing/, checked October 3, 2026. Seat types carry different usage allowances, and more access can be bought as credits; subscription billing is distinct from API token accounting.↩︎

38. Anthropic, “Claude Mythos,” https://www.anthropic.com/claude/mythos, checked October 4, 2026. Vetted, currently US-only Mythos 5.1 access with default 30-day retention; dated notices of Mythos 5’s June 12 unavailability and July 1 restoration under lifted export controls.↩︎

39. OpenRouter, FAQ, https://openrouter.ai/docs/faq, and “API Credit & Rate Limits,” https://openrouter.ai/docs/api_reference/limits, checked October 4, 2026; Stripe, “Stripe agrees to acquire OpenRouter to help businesses optimize token routing and usage,” August 19, 2026, https://stripe.com/newsroom/news/stripe-agrees-to-acquire-openrouter. Product terms and transaction announcement; no transaction price is inferred.↩︎

40. Google Cloud, “New early anomalies and spend caps on Google Cloud Budgets,” July 28, 2026, https://cloud.google.com/blog/topics/cost-management/new-early-anomalies-and-spend-caps-on-google-cloud-budgets/, “Manage spend cap budgets,” https://docs.cloud.google.com/billing/docs/how-to/budgets-spend-caps, and “Provisioned Throughput overview,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/provisioned-throughput/overview, checked October 3, 2026. Preview caps block new usage, including usage covered by Provisioned Throughput, until lifted manually; enforcement uses estimated costs and is not instant.↩︎

41. OpenAI, “Reserved Tier for API Customers,” https://openai.com/api-reserved-tier/, and “Scale Tier for API Customers,” https://openai.com/api-scale-tier/, checked October 3, 2026. Per-model dollars-per-minute reservations drawn at list price, reserved traffic rejected last, annual commitments usable on any model. Vendor pages; contract terms are unpublished.↩︎

42. Microsoft, “Azure API Management policy reference - llm-token-limit,” https://learn.microsoft.com/en-us/azure/api-management/llm-token-limit-policy, checked October 4, 2026. Attributes, rate/quota responses, per-gateway counters, estimated streaming usage and concurrent overshoot. Verify the selected gateway tier and API schema.↩︎

43. OpenAI, “Rate limits,” https://developers.openai.com/api/docs/guides/rate-limits, and “Error codes,” https://developers.openai.com/api/docs/guides/error-codes; Google Cloud, “Standard PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/standard-paygo, and “Priority PayGo,” https://docs.cloud.google.com/gemini-enterprise-agent-platform/models/deploy/priority-paygo; Microsoft, “Enable priority processing for Microsoft Foundry Models,” https://learn.microsoft.com/en-us/azure/foundry/openai/concepts/priority-processing; AWS, “Service tiers for optimizing performance and cost,” https://docs.aws.amazon.com/bedrock/latest/userguide/service-tiers-inference.html; checked October 3, 2026. Documented ramp, retry, spend-limit and downgrade behavior; shared pool sizes are not published.↩︎

44. Anthropic, “Task budgets,” https://platform.claude.com/docs/en/build-with-claude/task-budgets, checked October 4, 2026. Advisory task-wide allowance on supported models; per-request output limits remain distinct from a total application spending cap.↩︎

45. Mike Fuller, “FinOps and Tokenomics: How do they fit together?”, FinOps Foundation, August 25, 2026, https://www.finops.org/insights/finops-tokenomics-fit-together/, checked October 4, 2026. Tokenomics spans the production chain and complements FinOps for AI; its example bill rises while tokens per resolved outcome fall.↩︎

46. OpenAI, “Reasoning models,” https://developers.openai.com/api/docs/guides/reasoning; Anthropic, “Effort,” https://platform.claude.com/docs/en/build-with-claude/effort; Google, “Gemini thinking,” https://ai.google.dev/gemini-api/docs/thinking, checked October 4, 2026. Model-specific controls and token accounting; Anthropic’s per-model defaults and cache invalidation on top-level effort changes.↩︎