The reasoning-model era has a cost problem

TL;DR

• OpenAI, Anthropic and Google each state that the tokens a model spends reasoning are billed as output tokens, at the full output rate, even when the API returns only a summary of that reasoning or nothing at all.
• The control surface has moved from fixed token budgets to named effort levels. Anthropic’s budget_tokens is deprecated on Claude 4.6 and rejected with an error on Claude 4.7 and later; OpenAI, Google and DeepSeek all expose level-style settings.
• Third-party measurement shows how wide the spread is: Artificial Analysis lists the same model, GPT-6 Sol, at $0.13 per evaluation task at low effort and $1.06 at max, for index scores of 34 and 48.
• The per-token price of a fixed capability keeps falling, by about 47 percent a quarter according to Epoch AI, but the price of a single request is now set by how hard the question is and how much thinking the buyer permits.

The price of a model call used to be a function of two numbers: how many tokens went in and how many came out. Reasoning models added a third, the tokens a model generates while working out its answer, and every major provider bills those tokens as output. The result, visible across the pricing and documentation pages fetched for this article on 27 September 2026, is that the cost of a request now varies with the difficulty of the question and with a setting the buyer chooses. This piece reports what the vendors’ pages say about that billing, what controls they expose, and what independent measurement shows the spread to be.

Reasoning tokens on the bill

A reasoning model produces an intermediate chain of tokens before its final reply. Providers differ in what they show of that chain, but they agree on how they charge for it.

OpenAI’s reasoning guide puts it in one sentence: “While reasoning tokens are not visible via the API, they still occupy space in the model’s context window and are billed as output tokens.” The count appears in the response under output_tokens_details.reasoning_tokens. OpenAI also tells developers to reserve “at least 25,000 tokens for reasoning and outputs” when they start with these models, and notes that max_output_tokens limits the total generated, “including reasoning tokens, visible output tokens, and non-visible formatting tokens”.

Anthropic’s steering and cost page lists thinking charges as “Tokens Claude uses while thinking (billed as output tokens)” and adds a warning: “You are billed for the full thinking process, not the thinking content visible in the response.” Whether a developer asks for summarised thinking or omits it entirely, the billed count is the same, and the breakdown is exposed in usage.output_tokens_details.thinking_tokens. The same page notes that thinking blocks from earlier turns that remain in context are billed again as input on models that keep them.

Google’s Gemini thinking documentation says: “When thinking is turned on, response pricing is the sum of output tokens and thinking tokens.” It goes further on the summary question: “Pricing is based on the full thought tokens the model needs to generate, despite only the summary being output from the API.” The thought count is reported as total_thought_tokens.

DeepSeek’s thinking mode guide states that “Thinking mode is enabled by default” and that the chain of thought “is returned via the reasoning_content parameter, at the same level as content”. Its pricing page lists a single output price per model and does not carry a separate note on how reasoning content is metered.

Current list prices

The following per-million-token list prices were shown on the vendors’ pages at the time of writing. All are US dollars for standard, non-batch access.

Two footnotes matter for anyone forecasting a bill. Anthropic’s pricing page says Claude 4.7 and later models use a newer tokenizer that “produces approximately 30% more tokens for the same text”, so a like-for-like price comparison against Claude 4.6 understates the difference. It also records that Sonnet 5’s $2/$10 introductory price, due to rise to $3/$15 on 1 September 2026, “is now the standard price” and the increase “will not occur”. Google’s pricing page schedules a doubling of the Gemini 3.8, 3.7 and 3.6 Flash prices on 1 January 2027.

Controls by provider

A year ago the main lever was a token budget. Today the vendors expose named levels, and one has withdrawn its budget parameter from new models. The table summarises what the documentation fetched for this piece describes.

Availability differs by model even within a family. OpenAI’s GPT-6 Sol page lists all levels from none to max, while the reasoning guide notes that “GPT-6 Astra does not support none reasoning effort”.

Anthropic’s effort page describes effort as “a behavioral signal, not a strict token budget” and says it applies to “all tokens in the response”, including text, tool calls and thinking. On Claude Opus 5.5, “Adaptive thinking is always on and can’t be turned off”. The extended thinking page confirms that the older budget_tokens mode is “deprecated on the Claude 4.6 models” and that “Claude 4.7 and later models do not support it and reject requests that use it”. Anthropic’s steering page describes the replacement: “Claude’s thinking is adaptive: the model evaluates each request and decides for itself whether to think and how much.”

Google’s page makes the same claim for its models: “Gemini models engage in dynamic thinking by default, automatically adjusting the amount of reasoning effort based on the complexity of the request.” The page describes thinking_level only, with no token-budget parameter for any current model.

Measured cost per task

List prices describe the rate, not the bill. Two independent groups publish measurements that connect the two.

Artificial Analysis reports a “cost per Intelligence Index task”, defined as the “weighted average cost (USD) per Artificial Analysis Intelligence Index task, segmented by token type”, across the ten evaluations in version 4.3.2 of its index. Its GPT-6 Sol page lists one model at six settings.

Same model, same per-token price, an eightfold spread in cost. The non-reasoning row is shown costing more per task than the low setting, which the page does not explain; the figures are reproduced as displayed. In its 22 September 2026 note on the GPT-6 launch, the firm reports that Sol’s list price fell from $4/$20 to $2/$10 and its cost per task at max from $1.99 to $1.06, while GPT-6 Luna fell to $0.07 per task from $0.18. It also reports that Sol used more output tokens on the index than its predecessor, 31k against 29k, so part of the price cut was absorbed by longer reasoning.

Epoch AI’s measurements point the same way over a longer window. In a February 2026 note, Jean-Stanislas Denain writes that on one benchmark Epoch tracks, “reaching roughly 27% accuracy required about 43 million output tokens with o4-mini with high reasoning effort in April 2025, but only about 5 million tokens with GPT-5.2 with low reasoning effort in December 2025.” In a September 2026 report, Luke Emberson and David Roodman estimate that “the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year”, and give a concrete case: 75 percent accuracy on GPQA Diamond cost $0.30 per question in January 2025 and $0.0004 eighteen months later, “a 725-fold drop”. The decline is uneven: “cost falls 66% per quarter (75× per year) for performance that has just debuted as SOTA. Two years later, prices fall half as fast, at 32% per quarter (4.7× per year).”

The research behind the trade

The reason vendors sell thinking at all comes from work on test-time compute, the idea that spending more computation when a model answers can substitute for a bigger model. Snell, Lee, Xu and Kumar, in a paper posted in August 2024, reported that a compute-optimal allocation strategy “can improve the efficiency of test-time compute scaling by more than 4x compared to a best-of-N baseline”, and that “on problems where a smaller base model attains somewhat non-trivial success rates, test-time compute can be used to outperform a 14x larger model.” Epoch’s September report describes the commercial consequence: “Reasoning models can productively consume far more tokens, but as a result extract good performance from an underlying LLM that is smaller and cheaper to run per token.” Our explainer on test-time compute covers the mechanism in more detail.

Same model, same per-token price, an eightfold spread in cost per task between the lowest and highest effort settings.

Buyer responses in the documentation

No survey of adoption is cited here. What follows are the practices the vendors themselves document as the way to manage the new variable.

Set effort per workload, not per model. Anthropic’s effort page recommends low for “Simpler tasks that need the best speed and lowest costs, such as subagents”, and lists “Consider dynamic effort: Adjust effort based on task complexity” among its best practices. For Opus 5.5 it tells developers to “Run an effort sweep on your own evals rather than carrying settings over from an earlier model.”

Cap the request, not just the thinking. Because effort is guidance rather than a limit, both OpenAI and Anthropic point to the total output cap as the hard boundary. Anthropic warns that “A max_tokens sized for a response with no thinking is often too small once Claude starts thinking on hard requests”. OpenAI’s 25,000-token reservation is the same advice from the other direction.

Mind the cache. Anthropic documents that changing the top-level effort between requests “invalidates cache breakpoints”, because the effort value is rendered into the prompt. Its answer, in beta behind the header mid-conversation-output-config-2026-07-01, is a per-message effort change that leaves the cached prefix intact on Fable 5.1, Mythos 5.1, Opus 5.5 and Opus 5.

Measure before shipping. Anthropic’s steering page says to “Run a representative sample of your traffic with and without the guidance, and compare how often thinking triggers”, along with output tokens, latency and answer quality. That is the discipline our comparison of evaluation frameworks describes, applied to a cost setting. Where routing between models or effort levels happens outside the application, it tends to live in a proxy layer; our field notes on running an AI gateway cover per-key budgets and routing rules, including the failure mode where a cheaper route silently degrades quality.

Limits of the evidence

Everything above is list price and public benchmark. Negotiated rates, batch discounts (Anthropic lists 50 percent) and prompt caching all change the arithmetic, and no vendor publishes the distribution of reasoning tokens across real customer traffic. The Artificial Analysis figures cover one index and one model family. Epoch’s cost-decline figures hold performance fixed, so they say nothing about what a buyer pays for the newest capability. DeepSeek’s pages fetched for this article do not state how reasoning content is metered, and Google’s page does not describe what each thinking_level does to token counts. Prices quoted here were current on 27 September 2026 and several carry scheduled changes.