What is Claude Fable 5.1?
Claude Fable 5.1 is Anthropic's September 2026 model for long-running agentic coding, multistep research, and document, spreadsheet, and slide work. Released September 1, 2026, it replaces Claude Fable 5 at the top of the Claude line. On Unifically it runs as anthropic/claude-fable-5-1. It shares its model with Claude Mythos 5.1, which stays limited to approved organizations; Fable 5.1 is the generally available version with safeguards for cybersecurity and biology. It takes text and image input, outputs text, and carries a 1M token context window with up to 128k output tokens. Adaptive thinking is always on.
What's new in Claude Fable 5.1
Agentic coding: 55.8% on Terminal-Bench 4.0
Claude Fable 5.1 scores 55.8% on Terminal-Bench 4.0 against 52.3% for Claude Opus 5, 42.0% for Claude Fable 5, and 37.3% for GPT-5.6 Sol. On Terminal-Bench-Science 0.1, agentic scientific research in a terminal, it more than doubles Fable 5 at 52.6% against 24.7%. CursorBench 3.2 lands at 73.4% at max effort.
#1 on the Artificial Analysis Intelligence Index
The independent Artificial Analysis index scores Claude Fable 5.1 at 66 at max effort and 65 at xhigh, ahead of Claude Opus 5 at 63 and Kimi K3 at 60, the top of 184 models as of September 2, 2026. Vals AI agrees: its Vals Index puts Fable 5.1 first of 51 models at 67.87%, with #1 on LiveCodeBench (90.52%), MMLU Pro (92.38%), and MMMU Pro (90.64%).
90.0% on ARC-AGI-2
ARC Prize verifies 90.0% on ARC-AGI-2 Semi-Private and 97.5% on ARC-AGI-1 Semi-Private at max effort. The effort ladder is the useful part: ARC-AGI-2 reaches 78.3% at low, 86.3% at medium, and 88.8% at high, so most of the reasoning gain is available well below max.
Knowledge work: 1853 on GDPval-AA v2
On GDPval-AA v2, an evaluation of real occupational tasks, Claude Fable 5.1 scores 1853 against 1824 for Claude Opus 5 and 1723 for Claude Fable 5. Humanity's Last Exam reaches 65.0% with tools and 60.9% without, and AutomationBench, which measures business workflows, nearly doubles Fable 5 at 31.4% against 17.1%.
Effort that pays off, and effort you can change mid-run
Anthropic reports Fable 5.1 matches or beats Fable 5 at low and medium effort and pulls far ahead at xhigh and max. New in the API: a per-message effort change that raises depth for a hard step and lowers it for routine ones without resetting the prompt cache, turn-scoped system messages for one-turn reminders, and a display mode that returns short progress updates between tool calls.
Best for
Multi-hour coding sessions
Multi-file features, large refactors and migrations, and code review that runs across sessions.
Root-cause debugging
Traces rare failures through vendor libraries and core dumps instead of patching symptoms.
Documents, spreadsheets, slides
Takes an analysis from first question to a finished deck or live-formula spreadsheet.
Deep research
Multistep web research that follows up on what it finds and grounds answers in primary sources.
Long-context reasoning
Connects details across the full 1M token window: filings, repos, or a whole agent transcript.
Computer use
Operates a browser and desktop apps more reliably and recovers from failed steps.
Use cases
Build an orchestrator agent that plans a change across several services, runs unattended for hours with verification loops, and reports back with evidence of what it built. Run it as the reviewer of last resort on production incidents: give it logs, a core dump, and the vendor library, and let it produce a root-cause analysis. Put it behind a finance or legal research product that reads long filings and contracts with charts and tables inline, and returns a structured brief with exact figures from the primary source. It also fits scientific research pipelines: Anthropic used the model family to design protein binders with a hit rate near 50% across 12 targets, build a new elevation map of a third of Venus from 30-year-old radar data, and rewrite GPU kernels that sped up seven open-source biology models by up to 2.5 times.
Limitations
Claude Fable 5.1 carries safety classifiers for cybersecurity and biology. It can now be used to find software vulnerabilities but not to write exploits, and some benign requests near those areas still get declined. Anthropic reports about 60% fewer cyber interventions per session than Fable 5 and 85% fewer benign biology refusals. Handle the refusal stop reason, and consider the built-in fallback that reruns a declined request on Claude Opus 5 or Claude Opus 4.8.
Three API changes are breaking if you come from Claude Fable 5. Forced tool use returns an error, so keep tool_choice on auto and name the tool in the prompt. Thinking blocks are tied to the model that produced them, and editing anything before a thinking block invalidates every block after it, so conversation history must stay append-only. Parallel tool calling is also more variable: in long agent loops it may issue one tool call per turn where Fable 5 batched several, which costs round trips but not answer quality.
It is the slowest model in the current Claude lineup. Artificial Analysis measures 66.4 output tokens per second, and hard tasks at high effort can run for many minutes, so plan for streaming and progress updates. It requires 30-day data retention and is not available under zero data retention unless Anthropic authorizes it. Output is text only.
Claude Fable 5.1 vs Claude Fable 5
Same context window, same output limit, same tokenizer, same always-on thinking. The difference is how far the model gets on hard, long tasks. Fable 5.1 wins Terminal-Bench 4.0 at 55.8% against 42.0%, Terminal-Bench-Science 0.1 at 52.6% against 24.7%, AutomationBench at 31.4% against 17.1%, GDPval-AA v2 at 1853 against 1723, and CursorBench 3.2 at 73.4% against 70.5%. Humanity's Last Exam moves from 63.8% to 65.0% with tools. On OSWorld 2.0 computer use, strict scoring rises from 36.1% to 41.7%. At low and medium effort the two models score about the same, so the upgrade is about ceiling, not floor. Early testers also report the writing is more readable over long multi-step runs. On Arena.ai there is no Fable 5.1 row yet as of September 2, 2026; Claude Fable 5 still holds #1 on the Text board at 1508, so re-check that board once votes land.
When to use Claude Fable 5.1
Use Claude Fable 5.1 when a task sits at the edge of what models can do and Claude Opus 5 at high effort still falls short: multi-hour agent runs, research that has to find something new, or debugging nobody else has cracked. Start it at the default high effort, drop to low or medium for routine turns mid-conversation, and reserve max for the steps where correctness matters more than time. For most day-to-day coding and document work, Claude Opus 5 remains the sensible default, and Claude Sonnet 5 covers high-volume, latency-sensitive traffic.
API examples
Call Claude Fable 5.1 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Parameter docs live at docs.unifically.com/models/llm/overview.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "anthropic/claude-fable-5-1",
"messages": [
{ "role": "user", "content": "This service crashes about once in a million runs. Here are the logs and the core dump. Find the root cause." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
anthropic/claude-fable-5-1, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
A 1M token context window, which is both the default and the maximum, with up to 128k output tokens per response. Reasoning across the full window is one of the six areas Anthropic lists as improved over Claude Fable 5.
Yes. It takes text and image input in one request and reads dense charts, filings, and tables nested in PDFs. Output is text.
Yes, and the gap grows with effort. It scores 55.8% on Terminal-Bench 4.0 against 42.0% for Claude Fable 5, 52.6% on Terminal-Bench-Science 0.1 against 24.7%, and 73.4% on CursorBench 3.2 against 70.5%. At low or medium effort it matches Fable 5 while using fewer tokens.
90.0% on ARC-AGI-2 Semi-Private and 97.5% on ARC-AGI-1 Semi-Private at max effort, verified by ARC Prize as of September 2026. ARC-AGI-2 already reaches 88.8% at high effort and 78.3% at low.
Adaptive thinking is always on and cannot be turned off. The effort setting from low to max decides how deep it goes, the default is high, and effort can change mid-conversation without resetting the prompt cache. Raw chain of thought is never returned.
Three things. Forced tool use returns an error, so use auto with an instruction naming the tool. Thinking blocks are tied to the model and break if you edit earlier turns, so keep conversation history append-only. And safety classifiers can decline cyber or biology requests, with an optional fallback that reruns the request on Claude Opus 5 or Claude Opus 4.8.
Guides and comparisons
Review pricing, limits, and tested outputs before running this model.
