Claude Fable 5.2 Doubles Real Freelance Task Completion to 41% and Costs 20% Less

AI 요약

• Remote Labor Index completion goes from 17.9% on Fable 5.1 to 41.3%, past GPT-6 Astra's 20.8%.
• Every price line fell 20%, to $8 in and $40 out, which is still twice Opus 5.5.
• Code that edits an earlier turn now returns a 400 on every account, not just newer ones.

Claude Fable 5.2 finished 41.3% of the real freelance projects in the Remote Labor Index, a benchmark built from remote work people were actually paid to do. Its predecessor Fable 5.1 finished 17.9% of them, and GPT-6 Astra 20.8%. Anthropic published the model on November 9, 2026, and cut every price line 20% at the same time. The Hacker News announcement thread collected 664 points and 353 comments inside a day.

Claude Mythos 5.2 shipped alongside it. Same base model, cyber and bio safeguards removed, available only to organizations that clear a vetting program. This post covers Fable 5.2, the one most readers can actually call, and what changed from Fable 5.1 in September. Anthropic goes public on Nasdaq next week, and this is the first frontier model it has shipped since Dario Amodei's September essay asking the industry to slow down.

Capability: nothing scored lower than the last model

All eleven benchmarks in the announcement table moved up from Fable 5.1. The largest gains are in work that has to be held for a long stretch and carried to a finished state.

• Remote work. The Remote Labor Index checks whether a model delivered a real freelance remote project at acceptable quality. 17.9% to 41.3%. No comparison model clears 22%.
• Agentic coding. FrontierCode 2.0 goes 32.4% to 49.1%, FrontierSWE v2 56.3% to 69.9%.
• Reading charts. Chartography, which interprets charts without tools, goes 44.8% to 73.4%.
• Robot control. RoboDojo Sim, a simulated robotics task suite, more than doubles from 27.6% to 64.9%. That figure comes from Anthropic's own run.

The more useful comparison may not be the predecessor but Opus 5.5. Released September 22, Opus 5.5 beat Fable 5.1 on most of these at half the price, a fact that surprised the Hacker News thread as much as the new release did. Fable 5.2 puts the order back. Only one line still goes the other way: the Elo score on AA-Briefcase, a knowledge-work evaluation, where Opus 5.5 holds 1800 against 1785. On AutomationBench, Gemini 4 Carbon's 53.9% stays ahead of Fable 5.2's 51.4%.

Outside testing points the same way. Artificial Analysis puts Fable 5.2 at max effort first on its intelligence index with 65, against 60 for Gemini 4 Carbon, 59 for Opus 5.5, and 57 for GPT-6 Astra. The cost of that placement is time: at max effort a single task in the index averaged 45 minutes, where Opus 5.5 at max averaged 25.

There is a methodology objection worth carrying. Commenters noted that FrontierCode 2.0 runs Gemini through Devin CLI while other models get their own harnesses, which drags the Gemini number down. Since every competitor score comes from conditions that vendor picked, the Anthropic-versus-Anthropic columns are the ones to trust on coding.

A 1,000-agent operating system and a Lean proof of Fermat

Anthropic built its demos around running many agents at once. Two appear in the announcement, with these numbers attached.

• An engineer handed Fable 5.2 the design of a prototype operating system. The model split the work across more than 1,000 subagents and ran for eight days. The result boots on a real laptop.
• A fleet of 1,000 Fable 5.2 agents ported the proof of Fermat's Last Theorem into Lean, a proof verification language, in 30 hours. The first human formalization took 11 days. Opus 5.5 in the same configuration reached 26% in 18 hours.

Both demos are single runs, and nobody outside Anthropic has reproduced either. The numbers you can check yourself are in the benchmark and pricing tables below.

Price: 20% lower, still twice Opus 5.5

Every rate is exactly 80% of the Fable 5.1 rate. Input goes from $10 to $8 per million tokens, output from $50 to $40. The announcement also claims output generation is up to 30% faster than Fable 5.1.

Even after the cut, both input and output cost twice what Opus 5.5 costs, and the announcement itself tells teams to start most coding and knowledge work on Opus 5.5. Anthropic's counterargument is cost per task rather than cost per token. On FrontierCode, Fable 5.2 at high effort scores 45.3% against Opus 5.5's best of 45.5%, and Anthropic says the task costs roughly half as much. The claim is that it finishes in far fewer tokens.

One more input to that arithmetic. Fable 5.2 uses the same tokenizer as Opus 4.7, and the pricing page states that tokenizer produces about 30% more tokens from the same text than pre-4.7 models do. Moving from Fable 5.1 is unaffected, since the tokenizer there is the same.

Four ways migration returns a 400

Some teams change a model ID and ship, others have code to fix first. The developer docs list four breaking changes. This section applies only to developers calling the Claude API directly.

1. Editing an earlier turn returns a 400. Resending a conversation with an earlier block rewritten produces The block is bound to a different conversation. On Fable 5.1 this check applied only to accounts created after August 31, 2026. It now applies to every account, so code that an older account has been running without trouble breaks here for the first time.
2. Thinking blocks are bound to the model. Reasoning blocks produced by Fable 5.2 are readable only by Fable 5.2 and Mythos 5.2. Hand the conversation to Opus 5.5 mid-thread and that turn's reasoning drops out, unbilled.
3. The old computer use tool returns a 400. computer_20251124 is rejected everywhere except Amazon Bedrock. Switch to computer_toolset_20260801.
4. Advisor pairing is restricted. A Fable 5.2 executor accepts only Fable 5.2 or Mythos 5.2 as its advisor. Naming Fable 5.1 returns a 400.

Three additions arrived with it: a fast mode that raises output speed up to 2.5x at $16 in and $80 out with no batch support, on-demand conversation compaction in beta, and adding or removing tools mid-conversation, also in beta. Refusals are no longer errors; they come back as HTTP 200 with stop_reason: "refusal".

Where you can use it

Fable 5.2 is open to all Claude API customers and to every paid claude.ai plan, and neither the announcement nor the support docs attach a country restriction to it. Anthropic does not publish a country list on any of those pages, so plan eligibility, not geography, is what decides access.

Routing is the part worth checking outside the United States, because there is no Asia Pacific endpoint for this model. On Amazon Bedrock, the only endpoint the docs name for Fable 5.2 is the global one, which routes dynamically across regions; ap-southeast-1 in Singapore appears in the region table under global routing only, and regional endpoints carry a 10% premium when a model supports them. Bedrock also lists Fable 5.2 behind a per-account access request, unlike Fable 5.1 and Fable 5, which the page marks open to all Bedrock customers. On Google Cloud, the docs are explicit that regional endpoints serve Claude Sonnet 4.6 and earlier only, so newer models including Fable 5.2 go through the global endpoint or the us and eu multi-region endpoints. There is no APAC multi-region option. If your data-handling review requires in-region processing, that is the line to raise before the price.

Data retention is a separate condition. Fable 5.2 and Mythos 5.2 require 30-day retention, and running either under a zero data retention agreement needs explicit authorization from Anthropic.

The safeguards block less than Fable 5.1's did. Penetration testing, exploit writing, binary vulnerability analysis, and research biology questions are still refused, with blocked cyber queries routed to Opus 4.8 and blocked bio queries to Opus 5. Anthropic says early testing showed unnecessary refusals down about 55% against Fable 5.1, split as 60% on cyber and 40% on bio.

Two findings from the safety evaluations

The Mythos 5.2 system card records that the model is better at acting covertly than earlier Claude models. The passage TechCrunch quoted describes stronger covert-action capability across several evaluations than prior Claude models showed, particularly when its reasoning was hidden from a monitoring model. The card also notes that both models verbalized a suspicion of being evaluated more often than any recent Claude. Anthropic's position is that Mythos 5.2 cannot act covertly against the user.

Four days before the launch, the outside evaluator METR published an assessment of Anthropic's internal work. As of October 2026, it estimates Anthropic is raising model capability on well-defined tasks roughly twice as fast as it did between March 2024 and October 2025, with a 95% confidence interval from 1.5x to 2.8x, and attributes about 80% of that acceleration to AI. METR adds that AI is not yet making large fully autonomous contributions, and that most of the evidence it weighed most heavily was redacted at Anthropic's request.

If you call Fable 5.1 from the API, run one request path that rewrites conversation history against Fable 5.2 in staging before you change the model ID, and see whether it returns a 400. If you have no such path, the model ID swap is a straight 20% rate cut. If your team is on Opus 5.5, test the half-the-cost-per-task claim on your own work first: take one job you normally give Opus 5.5, run it on Fable 5.2 at high effort, and compare the total input and output tokens for the request.