What's Changed
Claude Opus 5 support — Opus 5 previously fell through to the legacy Claude model type, which sent temperature and top_p (rejected with a 400) and left reasoning, vision, tool use, and prompt caching disabled.
- Added Claude Opus 5 model type, ID detection (
claude-opus-5), and inference range. - Omitted sampling parameters and raised the default max tokens to 32000 so adaptive thinking has room alongside the response.
- Adaptive thinking is on by default; disabling it now sends
reasoning_config.type=disabledexplicitly, withxhigh/maxeffort clamped tohighsince Opus 5 rejects that combination. - Enabled vision, tool use, streaming tool use, prompt caching, document chat, and system prompts.
OpenAI GPT-5.6 Sol, Terra, and Luna — the new capability tiers are now selectable, served through the bedrock-mantle Responses API.
- Added the three tiers with full model ID matching ahead of GPT-5.5/5.4.
- Gave each tier its own token budget and reasoning effort baseline (Sol 32K/high, Terra 16K/medium, Luna 8K/low), replacing a hardcoded 8192/medium that also applied to GPT-5.5/5.4.
- Exposed
xhighandmaxreasoning effort levels, which GPT-5.6 supports. - Surfaced GPT-5.6 prompt caching from
input_tokens_details.
Region handling — Mantle model availability is now a single source of truth and is enforced on requests, not just when listing models. A conversation saved against a model no longer calls an endpoint that doesn't serve it after a region switch, and errors surface a readable message instead of a raw NSError.
Fixes
- Fixed Claude Sonnet 5 missing from the thinking toggle and MCP tool menu, where a
claude-sonnet-4substring check never matched it. - Fixed the Reasoning Effort picker rendering its label vertically.
Full Changelog: v1.4.9...v1.4.10