ANTHROPIC / MODEL RELEASE

Claude Fable 5.2: Turning Long Tasks into Reviewable Workflows

The upgrade is not simply “longer answers.” It is a model designed to keep moving through difficult coding, research, and design work while leaving checkpoints behind.

This article uses Anthropic’s release page, Claude Platform documentation, migration guide, and public reporting. Benchmarks, prices, availability, and safety claims are vendor information; the workflow recommendations are Evolve’s editorial judgment.

Claude Fable 5.2 launched on November 9, 2026. Anthropic positions it for the hardest knowledge and coding work: long-running agents, cross-repository migrations, research, and analysis that ends in a finished report. Fable 5.2 and Mythos 5.2 share the same underlying model; Fable is the broadly available configuration with stronger safeguards for cybersecurity and biology.

Claude Fable 5.2 long-horizon agent loop from ambiguous brief to verified delivery — Figure 1: An original Evolve redraw. Longer runs still need evidence and a human approval gate.

1Mcontext window
tokens

128Kmaximum output
tokens

$8 / $40input / output
per MTok

2.5×maximum fast-mode
speedup

The numbers point to long chains, not every prompt

Anthropic’s table puts Fable 5.2 above Opus 5.5 on most listed capabilities. The clearest signal is in tasks that require decisions and tools to work together: FrontierCode 2.0 for agentic coding, OSWorld 2.1 for computer use, and Terminal-Bench-Science 0.1 for scientific agents.

Selected Claude Fable 5.2 benchmark scores compared with Fable 5.1 and Opus 5.5 — Figure 2: Four selected rows from Anthropic’s table. These are vendor-reported results, with different effort settings and test conditions.

Benchmark	Fable 5.2	Fable 5.1	Opus 5.5	What it tests
FrontierCode 2.0	49.1%	32.4%	45.5%	Agentic coding
FrontierSWE v2	69.9%	56.3%	62.3%	Software engineering
OSWorld 2.1	89.3%	83.5%	87.2%	Computer use
Terminal-Bench-Science 0.1	74.6%	52.6%	58.7%	Scientific agents

Fable 5.2 also reports 51.4% on AutomationBench, 82.1% on WANDR’s relaxed score, 74.2% on Humanity’s Last Exam with tools, and 73.4% on Chartography. None of these numbers is your business success rate: tools, stopping conditions, test sets, and cost accounting all matter.

Read benchmarks this way: ask whether the model makes a task easier to finish, then ask what one successful delivery costs. Fable 5.2’s advantage is sustained judgment and self-checking, not automatically cheaper short answers.

Price fell, but effort is the billing switch

Standard pricing is $8 per million input tokens and $40 per million output tokens, down 20% from Fable 5.1’s $10/$50. Cache reads cost $0.20 per MTok, and Batch API input and output receive a 50% discount. Fast mode can generate up to 2.5 times faster at $16/$80 per MTok.

The documentation treats effort as a resource dial: it changes thinking depth, tool calls, latency, and token use. Start at high, the API default; use low for clear high-volume tasks, medium for interactive work, and xhigh or max only when your evals show a gain.

Claude Fable 5.2 effort and cost map — Figure 3: Effort and fast mode are separate cost and latency choices. Re-run an effort sweep on your own tasks.

Anthropic says early FrontierCode testing put Fable 5.2 at high effort around Opus 5.5 at max effort for roughly half the task cost. That is a vendor test claim; your tool calls, retries, and input lengths decide the actual bill.

Migration: remove the 400 errors before traffic moves

Changing the model ID is the smallest part of a migration. Adaptive thinking is always on, so thinking: {type: "disabled"} is invalid. Fable 5.2 also defaults to high, while Opus 5.5 defaults to medium.

Tool choice: tool_choice values any and tool return 400. Keep auto, state the desired action, and validate the tool schema.
Thinking blocks: return them unchanged in every tool loop, including empty blocks; select content by type rather than position.
Computer use: outside Bedrock, replace computer_20251124 with computer_toolset_20260801.
Data retention: 30-day retention is required by default. Zero data retention needs an eligible Enterprise Frontier Safeguards path.

The migration guide also calls out separate rate limits and lower token ceilings, and says a Fable 5.2 executor can use only a Fable 5.2 or Mythos 5.2 advisor. Treat these as launch checks, not post-incident documentation.

Where it earns a real trial

Cross-repository migrations

Give the model acceptance criteria, compatibility constraints, and test commands. Let it inspect dependencies, make staged changes, run tests, and explain failures. The value is the trail of diffs and evidence, not just more code per turn.

Research to delivery

For work that reads papers, runs experiments, compares results, and produces a report, a large context can help. Require sources, assumptions, unresolved questions, and reproducible commands.

Hardware and visual design

Ask it to list objects, dimensions, and uncertainties before letting it generate CAD, code, or simulations. That gives a human a chance to catch a visual mistake before it becomes an action.

Our production gate: recognition (state the input and gaps) → action (pause before payment, deletion, or sending) → delivery (keep sources, logs, tests, and failure reasons).

Conclusion: treat Fable 5.2 as a long-task lead

Claude Fable 5.2 moves the model from one-shot answers toward sustained work: split the goal, use tools, check the result, and revise until there is a document, code change, or research artifact. The lower price and larger context come with always-on thinking, stricter retention requirements, and API compatibility changes.

Choose one familiar process with measurable acceptance criteria. Run it at medium, high, and fast mode; record success rate, tokens, tool calls, human takeover, and rework time. That cost sheet is the meaningful answer for your team.

Sources and limits: Anthropic’s release, model documentation, migration guide, prompting guide, product page, and TechRadar’s report. The illustrations are original Evolve redraws.