GPT-5.6 Sol Gets Faster API Mode as Bedrock Adds Explicit Caching

OpenAI's July 30 update changes the operating economics around the GPT-5.6 family rather than describing a new model. Sol receives a Fast API mode, while Terra and Luna receive lower prices. OpenAI says Fast mode can deliver up to 2.5 times the speed of Standard processing for Sol at twice the price, with no change in intelligence, and that existing requests tagged `priority` remain compatible. For teams already using Sol, the immediate decision is therefore whether latency is worth a predictable premium on selected paths, not whether to replace their model integration.

The price reductions are deliberately uneven. OpenAI lists Terra at $2 per million input tokens and $12 per million output tokens, and Luna at $0.20 and $1.20 respectively, starting July 30; Sol pricing is unchanged. That makes the three-tier split more operationally consequential: Sol can be reserved for difficult planning or debugging steps, while Terra and Luna handle routine work where the cost of a wrong answer or slow response is lower. Those are vendor pricing claims, not a universal quality ranking, so an application should test task success and latency before moving traffic between tiers.

AWS adds a concrete deployment path for the same family. Its Bedrock guidance says Sol, Terra, and Luna are generally available through the OpenAI-compatible Responses API on the `bedrock-mantle` endpoint, with Sol positioned for autonomous coding, security research, scientific analysis, and deep multi-step reasoning. AWS also lists a 272K-token context window and the same reasoning-effort settings for all three models. For an existing Responses API application, that makes a Bedrock trial comparatively bounded, but region availability, quotas, and data-retention settings still need to be checked against the deployment's own requirements.

The more distinctive AWS change is explicit prompt caching. Cache reads are billed at a 90 percent discount relative to uncached input, but writes cost 1.25 times the uncached input rate; a reusable prefix must contain at least 1,024 tokens and remains available for at least 30 minutes. That favors write-once, read-many agent runs with stable system instructions, tool definitions, or reference material. AWS recommends logging `cached_tokens` and `cache_write_tokens`, which is an important limitation: the advertised discount is not a substitute for measuring whether a workload actually reuses enough context to offset cache writes.

OpenAI's accompanying ARC-AGI-3 analysis reinforces why model comparisons should include the harness. It reports that retaining reasoning and using compaction raised GPT-5.6 Sol's score on its public task set from 13.3 percent to 38.3 percent while reducing output tokens sixfold. The result is OpenAI's own evaluation, not an independent leaderboard result, but the practical lesson travels: context retention, truncation policy, and API settings can change both cost and measured capability. A careful GPT-5.6 rollout should therefore compare full agent configurations, including cache behavior and reasoning retention, against a current production baseline rather than treating a model name or token price as the whole upgrade.