2.1.1

2026-09-08

Added

• gpt-6-astra on the OpenAI engine: Responses API, 1.05M context, reasoning efforts low–max, and explicit prompt-cache breakpoints (writes bill at 1.25x input, reads at 0.1x).
• complete() pass-through turn on the Anthropic engine: posts Messages requests as written — system content blocks with cache_control, multi-turn tool_use/tool_result history — and reassembles streamed replies with original block order, thinking signatures, and redacted_thinking, so they can be echoed back next turn.
• cache_write rate on ModelPricing for providers that bill cache writes above the input rate.
• Anthropic engine test suite: 42 mock tests plus live Opus 5 thinking-signature and tool-call roundtrips.

Changed

• Removed the long_context_1m wrapper on the Anthropic engine; 1M context is the model default, and the retired beta header can still be passed via extra_headers.
• System content blocks (with cache_control) are posted as structured blocks instead of being stringified.
• OpenAI engine validates reasoning.effort locally against the model spec instead of round-tripping an HTTP 400.

Full Changelog: 2.1.0...2.1.1