模型 API／推論成本
GPT-5.6 Sol Temporarily Drops to $20 per Million Output Tokens, While Entire Requests With Ultra-Long Prompts Still Incur Higher Rates
OpenAI has cut GPT-5.6 Sol API input and output prices by 20% and 33%, respectively, with the promotional rates lasting at least through November 21. Once input exceeds 272K tokens, the higher rates still apply to the entire request, so long-context agent costs cannot be estimated from standard rates alone.
OpenAI updated GPT-5.6 Sol pricing on August 21: input fell from $5 to $4 per million tokens, output dropped from $30 to $20, and cached input reads now cost $0.40. The company describes these as promotional prices lasting at least through November 21, 2026, rather than permanent price cuts. The original announcement also promises lower prices only for the API and credits, so teams should not assume that ChatGPT subscription allowances will increase accordingly. The [OpenAI update](https://openai.com/index/gpt-5-6/) details the reductions, while the [API model page](https://developers.openai.com/api/docs/models/gpt-5.6-sol) provides current rates and limits.
The tiered pricing rule is what truly affects agent cost models. Sol offers a 1,050,000-token context window and up to 128,000 output tokens, but when input exceeds 272K tokens, the entire request is billed at 2× the input rate and 1.5× the output rate—not just the portion above the threshold. At promotional rates, that works out to $8 per million input tokens and $30 per million output tokens. Cache writes are also billed at 1.25× the uncached input rate; only repeated reads receive the 90% discount. Large codebase indexes, long execution traces, and merged multi-agent contexts should therefore still use chunking, summarization, and explicit cache breakpoints instead of treating the 1M-token context window as free capacity.
The API supports Responses, Chat Completions, function calling, Structured Outputs, MCP, tool search, and computer use. The `gpt-5.6` alias currently routes to Sol, while reasoning effort ranges from `none` to `max`. Engineering teams should monitor post-promotion pricing, cache hit rates for long prompts, and changes to the version behind the alias. Deployments requiring reproducible evaluations should continue pinning a snapshot rather than only the product-family name. GitHub has also positioned Sol as a deep-reasoning option suited to large codebases and long-running agent workloads, but this is a use-case designation—not a guarantee of cost or accuracy. See the [GitHub model comparison](https://docs.github.com/en/copilot/reference/ai-models/model-comparison).