Claude Fable 5.2: What Changed, What It Costs and Who Gets It

What did Anthropic ship on 9 November?

Two models arrived together: Claude Fable 5.2 (claude-fable-5-2) and Claude Mythos 5.2 (claude-mythos-5-2). Anthropic says they are the same underlying model with different safeguards. The figures below are Anthropic's own, and later sections give the caveats.

• •Adaptive thinking is always on. You steer depth with five effort levels: low, medium, high, xhigh and max.
• •Medium is the default in the Claude apps and Claude Code. High is the default on the API, which matters when you move code across.
• •Reliable knowledge runs to June 2026, and Anthropic commits to not retiring the model before 9 November 2027.
• •Anthropic's docs say when to move: when your evals on an Opus model at higher effort still do not meet your requirements.

What is the Fable version history, and is there a Fable 4.5?

There is no model called Claude Fable 4.5. The Fable line began with Fable 5 in June 2026, so a page reviewing Fable 4.5 is wrong; the name is probably a mix-up with older 4.x Claude models. The line has had a turbulent five months, which is worth knowing before you build a product on it.

What is the difference between Fable 5.2 and Mythos 5.2?

The capability is shared and the safeguards differ. Anthropic's system card calls them two configurations of one model: Fable 5.2 is for general use with extra safeguards in dual-use areas such as biology and cybersecurity, and Mythos 5.2 has more permissive safeguards there. Both cost the same. Unless you do defensive security or life sciences research, Mythos is nothing to chase.

• •Mythos 5.2 access runs through Anthropic's Cyber Verification Program and Life Sciences Verification Program, for vetted organisations.
• •Claude Security, available to Enterprise customers, includes Mythos 5.2.

How much does Claude Fable 5.2 cost in dollars and rupees?

Prices are per million tokens from Anthropic's pricing page. Rupee figures use the RBI reference rate of Rs 93.8579 per US dollar on 9 November 2026, rounded, and exclude tax and bank charges. The Opus 5.5 column is there because most readers are really choosing against it.

• •US-only inference (inference_geo set to us) multiplies every token category by 1.1.
• •Fable 5.2's tokenizer produces roughly 30% more tokens for the same text than Sonnet 4.6 and earlier, so re-measure if you are coming from an older model.

Will a 20% price cut make your bill 20% smaller?

Only if the model uses the same tokens, and often it does not. The list-price arithmetic below takes a job of 2 million input and 400,000 output tokens, no caching, with identical token use. That last assumption is the weak point. In Anthropic's own runs on 463 SWE-bench Pro problems, Fable 5.2 at high effort solved 96.3% at $1.39 per solved task (about Rs 130), while Fable 5.1 at its default solved 92.4% at $1.17 (about Rs 110). Opus 5.5 at high solved 95.5% for $0.30 (about Rs 28). On a research benchmark Fable 5.2 cost 20% to 61% less than Fable 5.1 per task and scored 4.7 to 7.0 points lower. Measure cost per finished job on your own work; /blog/claude-fable-5-2-vs-opus-5-5-cost-per-task-rupees does that comparison in full.

• •Fable 5.1: (2 x $10) + (0.4 x $50) = $40, about Rs 3,754.
• •Fable 5.2: (2 x $8) + (0.4 x $40) = $32, about Rs 3,003.
• •Fable 5.2 fast mode: (2 x $16) + (0.4 x $80) = $64, about Rs 6,007.
• •Opus 5.5: (2 x $4) + (0.4 x $20) = $16, about Rs 1,502.
• •Independent check: Artificial Analysis lists Fable 5.2 at max effort as the most expensive of 73 entries, at $26.54 per Intelligence Index task (about Rs 2,491).

How much faster is Fable 5.2, and is fast mode worth double?

These are two separate claims. Standard Fable 5.2 generates output up to 30% faster than Fable 5.1. Fast mode raises token generation by up to 2.5 times at twice the price. Both figures are Anthropic's, and we found no independent before-and-after measurement yet. Faster tokens do not guarantee a shorter job, because the model may write more of them: Artificial Analysis clocks Fable 5.2 at max effort at 103 tokens per second and 44.4 minutes per index task. Pay for fast mode only when a person is waiting on the answer.

• •Fast mode is set with speed: "fast" on the request. It is on the Claude API, Claude Models on Google Cloud, and Anthropic-hosted Microsoft Foundry deployments. Amazon Bedrock is not on Anthropic's list.
• •It cannot be combined with the Batch API, and batch at half price is usually the better trade for overnight work.

Does Fable 5.2 really refuse less often?

For most users, probably yes. Anthropic reports about 55% fewer blocks than Fable 5.1 in early testing across claude.ai, Claude Code and Cowork, including about 60% fewer in cybersecurity and 40% fewer in biology. It gave no absolute block rate. The classifiers still stop penetration testing, exploit generation, binary-based vulnerability work and most research biology tasks. Blocked queries are answered by Opus 4.8 for cyber and Opus 5 for biology, and that fallback has a cost: in the system card's Gray Swan prompt injection test, 53% of attempts against Fable 5.2 fell back to Opus 4.8, against 23% for Fable 5.1, and attack success at 15 attempts was 4.09% against 1.02%. If your agent reads untrusted pages or email, treat fallback as part of your threat model. See /blog/claude-fable-5-2-system-card-plain-english-findings.

• •On the API, a declined request returns HTTP 200 with stop_reason set to refusal, so a status-code check alone will miss it.

Who gets Fable 5.2 on each Claude plan?

The launch post's line about Pro has no end date; the help article supplies it. Pro and standard Team seats get Fable as a one-week promotion, from 9 November to 16 November 2026 at 11:59 pm Pacific, which is 1:29 pm IST on 17 November. After that, Fable returns to pay-as-you-go usage credits on those plans. Your default model does not change, so pick Fable in the model picker.

• •Tenure sets resets: under six months earns one, six to 12 months two, more than 12 months three, counted from your first paid bill.
• •A reset gives back only weekly usage spent on Fable, plus a full five-hour window where you have one. It does not raise your limits.
• •On Enterprise, Fable 5.2 stays off until an admin enables it. Seven days after launch it joins every role that already has a Fable model enabled, unless an admin opts out.
• •On cloud, it is on the Claude API, Amazon Bedrock, Claude Platform on AWS, Claude Models on Google Cloud and Microsoft Foundry.
• •Both models require 30-day data retention and are not available under zero data retention unless Anthropic expressly authorises it.

Did Fable 5.2 really prove Fermat's Last Theorem?

It formalised a proof humans had already found, and the fine print matters. The announcement says "a team of 1,000 Fable 5.2 agents fully formalized the theorem in just 30 hours, with minimal human intervention", using one-third as many action tokens as Anthropic's first formalisation and producing a proof seven times shorter. The system card is narrower: the run used "an earlier snapshot of Fable 5.2", and the teams "started with definitions and supporting lemmas from our earlier formalization". We would rely on the system card's wording. Treat it as evidence about multi-agent persistence, and do not read it as a forecast for your codebase. Neither passage gives a cost for the run.

• •It was one run per model at 1,000 agents, so repeatability is unknown.
• •At 18 hours the snapshot's team had reached 0.86 of the proof against 0.26 for Opus 5.5. Each team then resumed for 12 more hours, and the snapshot's team finished.
• •With smaller teams over 18 hours the gap narrows: 0.35 against 0.21 at 100 agents, and 0.12 against 0.09 for one agent.

How good is Fable 5.2, and who has checked?

It ranks first on the two broad independent boards we could read, and it does not lead everywhere. Anthropic's system card says it scored highest on 14 of 16 results in its own table, trailing Opus 5.5 on the GDPval-AA and AA-Briefcase Elo ratings. The full comparison is at /blog/fable-5-2-vs-gpt-6-astra-vs-gemini-4-carbon-benchmarks-independent.

• •Where it does not lead: Vals AI ranks it 20th of 37 on its Legal Index and 6th of 39 on Finance.
• •Still missing as of 11 November: a METR time-horizon figure, a listing on Scale's public SWE-bench Pro board and, as far as we could read, a full Arena text ranking.

Should you care about Fable 5.2? Our verdict by reader

Yes, if you run long, hard, open-ended agent work. For most other jobs, Anthropic's own advice stands: "For most coding and knowledge tasks, we recommend that users start with Claude Opus 5.5". Here is what we would do in each seat.

What mistakes should you avoid when switching?

Most early problems will come from assumptions carried over from earlier models.

Which dates should go in your calendar?

Two of the launch terms fall within days of this article.

Frequently asked questions about Claude Fable 5.2

Short answers to what people are searching for this week.

Where Tech Arion fits

Tech Arion is a Hyderabad software and AI consultancy that builds with Claude Code and advises teams on model choice. If you want a second opinion on which workloads justify Fable 5.2, see /services/ai-consulting.