OpenAI and Anthropic both cut prices on 22 September. One of them costs you a refactor.

OpenAI released GPT-6 Sol and GPT-6 Luna on 22 September. Both accept text and image input and return text, through the Responses and Chat Completions APIs. Standard pricing per million tokens, for prompts up to 272K input tokens, is $2 input, $0.20 cached input and $10 output for Sol, and $0.10 input, $0.01 cached input and $0.50 output for Luna.

The same day Anthropic released Claude Opus 5.5 at $4 and $20 per million tokens, against Opus 5 at $5 and $25. It comes with a 1M token context window by default, 128k maximum output tokens, and always-on adaptive thinking. Anthropic says the model will cost 40 percent less than Opus 5 on typical workloads at default settings, a larger figure than the 20 percent list price cut because it also claims token efficiency gains, and that it generates output more than 30 percent faster.

One detail in the Anthropic pricing is easy to miss and worth real money on agentic workloads. Cache hits on Opus 5.5 are priced at 0.05x the base input price rather than the usual 0.1x multiplier, which puts a cache hit at $0.20 per million tokens. Anthropic describes that as 60 percent less than Opus 5.

The GPT-6 prices above are the short context rates. Above 272K input tokens, different numbers apply: GPT-6 Sol moves to $4 input, $0.40 cached and $15 output, and GPT-6 Luna to $0.20, $0.02 and $0.75.

Note the shape of that change. Input and cached input double, but output rises by half, not double. It is a common mistake to describe long context pricing as a doubling across the board, and it produces a budget that is wrong in both directions depending on how output-heavy your workload is.

If your prompts sit anywhere near 272K, this is worth a guard in code rather than a note in a spreadsheet. Count tokens before the call and decide deliberately whether a given request is worth crossing the threshold.

The Anthropic cut is not a model id swap. The release notes list three request shapes that now fail outright, and a failure here is an HTTP 400, not degraded output.

First, thinking is always on, and the thinking field is rejected in both its forms. Sending thinking with type disabled returns a 400, and so does sending it with type enabled. The fix is to omit the field entirely and control thinking depth with the effort parameter instead.

Second, tool_choice types any and tool return a 400, as they already did on Claude Fable 5.1. Anthropic points you at auto with strict tool use. If you currently force a tool call to guarantee a structured response, that is a real refactor rather than a config change, because you are replacing a hard guarantee with a softer one and need to handle the case where the model answers in prose.

Third, computer use on this model requires the computer_toolset_20260801 toolset on the Claude API and on Google Cloud, where the earlier computer_20251124 tool returns a 400. On Amazon Bedrock, computer_20251124 keeps working, so the same code can pass on one platform and fail on another.

Read what each provider commits to. OpenAI explicitly labels GPT-5.6 Sol's price as promotional, available at least through 21 November 2026. That phrasing is a floor, not a ceiling, and it tells you the number is a decision the provider can revisit. Nothing on OpenAI's changelog or pricing page describes GPT-6 Sol or Luna pricing as promotional or introductory.

Google's pricing page does something different again. It carries no word like introductory anywhere, and simply lists two prices with a date between them: Gemini 3.6, 3.7 and 3.8 Flash go from $0.75 to $1.50 input and $3.75 to $7.50 output on 1 January 2027. A scheduled increase stated as plainly as that is arguably more honest than a promotional label, and it is certainly easier to budget against.

The broader point for anyone building a business on tokens: a price with a stated end date is a different product from a price without one. Treat the first as a promotion you are borrowing and the second as a number you can plan on, and check which kind you are standing on before you quote a customer a margin.

Two frontier cuts in one day is unusual and it is tempting to reroute everything immediately. The cheaper move is to treat a price change as a prompt to run the eval you already have, not as a reason to rebuild routing.