tool_choice types any and tool now return 400 on Fable 5.1 because always-on adaptive thinking makes forced tool selection incompatible with the generation order. Update to tool_choice: {"type": "auto"} plus structured outputs before you switch model versions.any and tool return 400; update to auto + structured outputs now or pin to Fable 5git clone https://github.com/every-app/open-seo
cd open-seo && cp .env.example .env # add DATAFORSEO_LOGIN and DATAFORSEO_PASSWORD
docker compose up -d
tool_choice: {"type": "any"} and tool_choice: {"type": "tool", "name": "..."} now return a 400 error when used with claude-fable-5-1 or claude-mythos-5-1. auto and none are unchanged.auto + strict schemas or structured outputs instead.any/tool) conflicts with the generation order because the thinking block must complete first. Anthropic's documented fix is: (1) set tool_choice: {"type": "auto"}; (2) define your tool schemas with strict: true to guarantee schema-conformant inputs; (3) if you truly need the model to output a specific structure, switch to structured outputs (response_format: {"type": "json_schema", ...}). If you cannot change your calling code immediately, pin to claude-fable-5 while you migrate — it still supports any/tool. The migration guide is at platform.claude.com/docs/en/models/fable-5-1/migration-guide.tool_choice.type set to "any" or "tool" — common in agentic frameworks that force the model into a specific tool step.EffortModerate — requires updating tool_choice parameters and potentially switching to strict schemas or structured outputs.thinking-binding-controls-2026-08-01 gives you control over the mismatch behavior: thinking.block_binding.prefix_mismatch_behavior accepts reject (400) or drop (silent), and the input_transformations response field tells you what was dropped.claude-fable-5-1) succeeds Fable 5 for agentic coding, knowledge work, and research — same pricing ($10/$50 per MTok input/output), same context window (1M tokens), same max output (128k tokens), but with always-on adaptive thinking and a dramatically different cache read price ($0.25/MTok = 0.025x, vs. $1/MTok = 0.1x on all other Claude models). Claude Mythos 5.1 (claude-mythos-5-1) ships simultaneously for Project Glasswing participants only — not generally available.thinking blocks even if you didn't opt in. (3) Fable 5.1 requires 30-day data retention — if you're on a ZDR contract, you cannot use Fable 5.1 until EFS is available to you (fall 2026) or you receive explicit authorization. Available now on Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Azure AI Foundry. New beta features available with Fable 5.1: per-message effort control (beta header mid-conversation-output-config-2026-07-01), turn-scoped system messages (mid-conversation-system-clear-at-2026-08-21), and thinking display updates (thinking-display-updates-2026-08-18).tool_choice: any/tool (see Breaking Changes above); you parse Claude API responses and don't handle thinking blocks.EffortSignificant — multiple breaking changes require code review before switching; data retention constraints may require organizational coordination.Nothing cleared the quality gate this period. arXiv RSS feeds (cs.AI, cs.CL, cs.LG, cs.CV) were egress-blocked from this environment. Papers surfaced via search (benchmark rewriting for LLM math evals, trajectory-judge for agent evaluation, LLM evaluation validity) had no associated code repos from recognized labs or were EMNLP/NeurIPS workshop submissions without confirmed September 1 arXiv submission dates. No papers with benchmark numbers, code repos, and confirmed September 1 submission date were verifiable.
Quick hit only — see Quick Hits section.
Claude Mythos 5 has moved to #1 on the LMArena text leaderboard with ~1531 ELO, overtaking Claude Fable 5 (~1525) which held the top position as of the August 31 digest. 56 models are currently tracked. The cluster just below: Claude Opus 4.8 (~1510 ELO), GPT-5.5 Pro (~1510 ELO), Gemini 3.1 Pro Preview — still tightly contested.
Context: Mythos 5 was released in June 2026 for Project Glasswing participants; it has been accumulating arena votes since then. The leaderboard re-ranks weekly as votes accumulate, so this reflects sustained preference, not a sudden surge.
Note: Claude Fable 5.1 and Mythos 5.1 (both launched September 1) are not yet on the leaderboard — arena ranking requires public availability and vote accumulation. Fable 5.1 is now broadly available; expect it to enter rankings in coming weeks.
tool_choice: any/tool pattern — used extensively in agentic frameworks to guarantee structured outputs — becomes permanently unsupported. The Anthropic-recommended replacement (structured outputs + auto tool choice) is more declarative and arguably more robust, but requires rewriting a large surface area of LangChain, LlamaIndex, and custom agent code that relies on forced tool selection. Monitor whether major agentic frameworks update their Claude adapters in the next few weeks; if they do, this migration will be invisible to most developers. If they don't, it becomes a hard blocker for Fable 5.1 adoption.Nothing cleared the quality bar this period. HN threads from September 1 were not individually verifiable at score >200 with concrete technical data within the scan window.
maybe_load_text_tensor_range for slab-cached reads). Pre-release, binaries for macOS/Linux/Windows/Android. linkFiltered from 30+ primary sources against a published quality rubric. No press releases, no fluff — only what changes what you build.