Updated
What is Claude Fable 5.2?
Claude Fable 5.2 is Anthropic's November 2026 model for demanding reasoning and long-horizon agentic work: agent sessions that run for hours, multistep research, and analysis carried through to a finished document, spreadsheet, or deck. Released November 9, 2026, it replaces Claude Fable 5.1 at the top of the Claude line. On UnificAlly it runs as anthropic/claude-fable-5-2. It is the same model as Claude Mythos 5.2, which stays limited to verified organizations; Fable 5.2 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 5.2
Agentic coding: 49.1% on FrontierCode 2.0
At xhigh effort Claude Fable 5.2 scores 49.1% on FrontierCode 2.0, the highest of any released model Anthropic tested, against 45.5% for Claude Opus 5.5, 42.8% for GPT-6 Astra, and 32.4% for Claude Fable 5.1. At high effort it scores 45.3%, about level with Opus 5.5 at roughly half the cost per task. FrontierSWE v2 lands at 69.9% against 56.3% for Fable 5.1.
Up to 30% faster than Fable 5.1
Anthropic reports output up to 30% faster than Claude Fable 5.1. Fable 5.1 was the slowest model in the Claude lineup, so this matters most for long agent runs where every turn waits on the previous one.
Knowledge work and research
Humanity's Last Exam with tools reaches 74.2% against 65.6% for Fable 5.1 and 67.7% for Opus 5.5. AutomationBench, which measures business workflows, rises from 31.4% to 51.4%, and the Remote Labor Index from 17.9% to 41.3% at high effort. Vals AI ranks it first of 36 models on the Vals Index at 71.31%.
Computer use and charts
OSWorld 2.1 reaches 89.3% against 83.5% for Fable 5.1 and 87.2% for Opus 5.5. Chartography without tools jumps from 44.8% to 73.4%, and at medium effort Fable 5.2 already beats Opus 5.5 at max effort.
Fewer refusals
Anthropic reports about 55% fewer safeguard blocks than Fable 5.1, including about 60% fewer in cybersecurity and 40% fewer in biology. A declined request still returns a refusal stop reason, so handle it in your code.
What Claude Fable 5.2 is best for
Multi-hour coding sessions
Large refactors, migrations, and multi-file features that run unattended with verification loops.
Orchestrating other models
Plans the work and hands simpler steps to smaller models such as Claude Sonnet 5.5 or Haiku 5.5.
Documents, spreadsheets, slides
Takes an analysis from the first question to a finished deck or spreadsheet.
Deep research
Multistep 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 long agent transcript.
Computer use
Operates a browser and desktop apps, with 89.3% on OSWorld 2.1.
Claude Fable 5.2 use cases
Run it as the orchestrator of a coding agent that plans a change across several services, delegates routine edits to a cheaper model, and reports back with evidence of what it built. Use it on production incidents where cheaper models have already failed: give it logs, traces, and the relevant source, and let it produce a root-cause analysis. Put it behind a finance or legal research product that reads long filings and returns a structured brief with exact figures from the source. Anthropic's own examples include a Python-to-Rust migration that made a production service start about 26 times faster, and a materials science code that ran about 36 times faster after Fable 5.2 rewrote it.
Claude Fable 5.2 limitations
Claude Fable 5.2 carries safety classifiers for cybersecurity and biology, and some benign requests near those areas still get declined. Handle the refusal stop reason and retry on another model if you need to.
A few API rules can break code that worked on Fable 5.1. Editing anything before a thinking block (the system prompt, the tools, or an earlier message) now returns an error on every account, so conversation history must stay append-only. Thinking blocks from Fable 5.2 are read only by Fable 5.2 and Mythos 5.2, so moving a conversation to another model loses its earlier reasoning. Forced tool use returns an error, so keep tool_choice on auto. Temperature, top_p, and top_k should be left out.
At high effort and above it thinks before almost every reply, and on long tasks at low or medium it sometimes stops with a list of what remains. It requires 30-day data retention and is not available under zero data retention unless Anthropic authorizes it. Output is text only.
Claude Fable 5.2 vs Claude Fable 5.1
Same context window, same output limit, same tokenizer, same always-on thinking. Fable 5.2 is faster, cheaper at list price ($8 / $40 per 1M tokens against $10 / $50), and stronger on every benchmark Anthropic published: FrontierCode 2.0 at 49.1% against 32.4%, FrontierSWE v2 at 69.9% against 56.3%, AutomationBench at 51.4% against 31.4%, Terminal-Bench-Science 0.1 at 74.6% against 52.6%, and Humanity's Last Exam with tools at 74.2% against 65.6%. It also blocks about 55% fewer requests. The main thing to check before switching is whether your code edits conversation history, because that now fails on every account.
When to use Claude Fable 5.2
Anthropic still recommends Claude Opus 5.5 as the everyday default for most coding and knowledge work. Move to Claude Fable 5.2 when Opus 5.5 at high effort still falls short: multi-hour agent runs, research that has to find something new, or bugs nobody else has cracked. Start at the default high effort, step up to xhigh for the hardest long-running coding work, and drop to low or medium for routine turns. Claude Sonnet 5.5 and Haiku 5.5 cover high-volume, latency-sensitive traffic.
Claude Fable 5.2 API examples
Call Claude Fable 5.2 from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model in the UnificAlly API. Full parameter docs live at docs.unifically.com/models/llm/anthropic/claude-fable-5-2.
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-5-2",
"messages": [
{ "role": "user", "content": "Plan the migration of this service from Python to Rust, then list the first three changes to make." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask about Claude Fable 5.2
anthropic/claude-fable-5-2, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your UnificAlly API key.
Anthropic released it on November 9, 2026, together with Claude Mythos 5.2. Both are the same model. Fable 5.2 is the generally available version with extra safeguards for cybersecurity and biology, and Mythos 5.2 stays limited to verified organizations.
A 1M token context window, which is both the default and the maximum, with up to 128k output tokens per response. It uses the same tokenizer as Claude Fable 5.1.
Yes. It takes text and image input in one request. Output is text.
Yes, on every benchmark Anthropic published. It scores 49.1% on FrontierCode 2.0 against 32.4%, 69.9% on FrontierSWE v2 against 56.3%, 51.4% on AutomationBench against 31.4%, and 74.2% on Humanity's Last Exam with tools against 65.6%. Output is also up to 30% faster, and Anthropic reports about 55% fewer safeguard blocks.
Adaptive thinking is always on and cannot be turned off. The effort setting from low to max decides how deep it goes, and the default is high. A prompt line asking it to answer without deliberating does not stop it thinking at medium and above, so lower the effort for easy turns instead.
Forced tool use returns an error, so keep tool_choice on auto. Editing anything before a thinking block now returns an error on every account, so keep conversation history append-only. Thinking blocks from Fable 5.2 are read only by Fable 5.2 and Mythos 5.2, so switching a conversation to another model drops its earlier reasoning.
Claude Fable 5.2 guides and comparisons
Review pricing, limits, and tested outputs before running this model.
