AI Model API Price Changes, Q3 2026: Five Labs, One Table

Between July 1 and September 30, 2026, Anthropic, OpenAI, Google, DeepSeek and xAI made at least 28 changes to the prices of their text-model APIs, counting new models and their launch prices. Most new flagships arrived at or below the price of the model they replaced, one vendor raised its prices, and several prices now carry an end date. Every row below is dated from the vendor’s own changelog or pricing page.

Prices are US dollars per million input and output tokens at the standard tier, for prompts under any long-context threshold. Where a vendor’s current page no longer shows an old price, it was read from an Internet Archive copy of that vendor’s own pricing page.

The chart ranks the quarter’s new text models by their output price. The spread runs from $0.50 to $50 per million, a hundredfold gap.

Successors at the same price or lower. Anthropic’s release notes put Opus 5 at Opus 4.8’s $5 / $25, then Opus 5.5 at $4 / $20 two months later. Sonnet 5.5 kept Sonnet 5’s $2 / $10, and xAI held Grok at $2 / $6 through two releases.

Cheaper cache reads. Fable 5.1 charges 0.025 times its input price for a cache read and Opus 5.5 charges 0.05 times, against the 0.1 times Anthropic charges on its other models. OpenAI’s changelog lists GPT-6.1 Sol’s cached input at $0.10, half of GPT-6 Sol’s. For agents that resend long context on every turn, these cuts can matter more than the headline rates.

Cuts soon after launch. OpenAI cut GPT-5.6 Luna by 80% and Terra by 20% on July 30, three weeks after the family launched, then put GPT-5.6 Sol on a promotional $4 / $20 on August 21, down from $5 / $30.

Time-of-day pricing. DeepSeek moved to peak and off-peak rates on August 16, with off-peak at half the peak rate; both were above the flat price they replaced. Its pricing page sets peak hours at 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, excluding Chinese public holidays. Our note on off-peak LLM pricing covers how to schedule around it.

Several of the quarter’s prices are temporary. Google’s Gemini API pricing lists Gemini 3.6, 3.7 and 3.8 Flash at introductory rates through December 31, 2026, doubling on January 1, 2027. OpenAI says GPT-5.6 Sol’s promotional price is available at least through November 21, 2026. One temporary price went the other way: Anthropic made Sonnet 5’s introductory $2 / $10 permanent on August 10 and cancelled the rise it had scheduled for September 1.

For current prices across every surface, not just changes, use our maintained frontier model API price index. The August cuts and promotions tracker and the Q2 price tracker give earlier context. Teams that want model costs modelled against their real workloads can work with our AI transformation team.

The conclusion says to re-run last month’s token counts against the current prices. The table makes that a mechanical job if it is done in the right order, and a misleading one if it is not. Start from the usage record, not from the table: for each model you ran, pull the month’s uncached input, cached input and output token totals from the provider’s own billing export. Those three totals are the workload; the table only supplies rates.

Then find each model’s current row and, separately, the row for its successor if the quarter produced one. Opus 5 users have Opus 5.5 at $4 / $20; GPT-6 Sol users have GPT-6.1 Sol with cached input at half the price. Apply both rows to the same three totals. The difference is the saving the successor would have produced on last month’s work, before any change in quality, retries or review time, which this table does not measure and which our guide to review cost on cheaper models covers.

Finally, check every model in the plan for an end date in the table’s Change column. A row priced at an introductory rate is re-priced a second time at the rate it will carry on the day the promotion ends, where that rate is published; for GPT-5.6 Sol, assume its pre-promotion $5 / $30 unless OpenAI says otherwise. The plan keeps both numbers. A budget that carries only the introductory figure is correct until a date the table already knows.

Keep the re-pricing in the same document as the usage export, with the table’s as-of date written at the top. When the Q4 edition of this table arrives, the exercise is repeated against the same usage record, which is what turns a one-off check into a quarterly habit: the workload stays constant and only the rate column moves, so the difference between two quarters is a clean measure of what the vendors’ price changes did to your bill, as long as the model mix is held constant.

A dated table of list prices is a good instrument that answers one question. The illustrative mistakes below are each a case of asking it a different one.

The headline-only comparison: two models are ranked on their output price and the cache-read price is ignored, when for an agent that resends long context every turn the cache rate can decide the bill, as the patterns section shows for Fable 5.1 and Opus 5.5. The frozen promotion: a 2027 forecast is built on Gemini 3.7 Flash at $0.75 / $3.75 without the January 1 doubling the pricing page already states. The flat-rate assumption: a DeepSeek workload is costed at one price when the vendor has billed peak and off-peak rates since August 16, and the plan did not say which hours the jobs run in. The successor as a free lunch: a cheaper successor is adopted on its list price, and the review work its output needs is never measured, so the saving on the bill is real and the saving on the task is unknown.

The table answers the first three in its Change column, which records cache-read prices, end dates and DeepSeek’s peak and off-peak split, provided it is read. The fourth it cannot answer, and the method rows below say so.

A fifth, smaller misreading is worth naming: taking a list price for a price. Almost every row is a standard-tier list rate for short prompts, as the first section says; the exception is OpenAI’s Fast mode row. Volume and enterprise agreements, regional uplifts and other premium speed tiers are excluded by the method, so a team on negotiated terms should treat the table as the public reference its own rates are measured against, not as the figure on its invoice, and should write its own negotiated rate beside each row it uses.