Claude Fable 5.2: Every Feature Explained, and When It Is Actually Worth the Price

On 9 November 2026, Anthropic released Claude Fable 5.2 and its restricted twin, Claude Mythos 5.2. It replaces Fable 5.1, which shipped only ten weeks earlier, in September. The headline claims are clear: outputs up to 30% faster, a 20% lower price, and better scores than Anthropic's own Opus 5.5 on almost every capability it measured.

Fable 5.2 and Mythos 5.2: One Model, Two Doors

The naming confuses people, so start here. Fable and Mythos are the same underlying model. The difference is access and safeguards.

• Fable 5.2 is generally available. It carries safety classifiers that block high-risk cybersecurity work, such as penetration testing and exploit generation, and most research-biology tasks. When a request is blocked, Anthropic answers it with an older model instead: Opus 4.8 for cybersecurity and Opus 5 for biology.

• Mythos 5.2 has those safeguards lifted for vetted organisations. Access is through Anthropic's Cyber and Life Sciences Verification Programs, and Claude Enterprise customers can use it inside Claude Security for codebase scanning.

The practical consequence is easy to miss. If your use case is security testing, Fable 5.2 will block most of it. Our own cybersecurity practice cannot use Fable for penetration-testing work, and nor can yours without verification. For everything else, you get the full model.

The Specifications at a Glance

Two numbers deserve context. The price is 20% lower than Fable 5.1 (US$10 / US$50), but it is still double Claude Opus 5.5 (US$4 / US$20). And the context window is 1 million tokens by default, but some third-party tools set a lower working default. Cursor, for example, lists 300k with 1M as the maximum.

Fable 5.2 against Fable 5.1 and Opus 5.5 on selected Anthropic-reported benchmarks

Feature 1: Adaptive Thinking With Five Effort Levels

Fable 5.2 always reasons before it answers, and there is no switch to turn that off. Instead you choose an effort level: low, medium, high, xhigh or max. Higher effort means deeper reasoning, more tokens and higher cost.

The defaults differ by surface, which catches teams out. The API defaults to high. The Claude apps and Claude Code default to medium. Anthropic's guidance is to start at high, step up to xhigh only for your hardest long-running coding and agentic work where you have measured a gain, and keep max for testing the model's ceiling.

Effort is also where most of the cost story lives. Anthropic reports that on its FrontierCode 2.0 coding evaluation, Fable 5.2 at High effort scored about the same as Opus 5.5 at Max effort (45.3% against 45.5%) for roughly half the cost per task. On Humanity's Last Exam, it says Fable 5.2 at Low effort matched Opus 5.5's top score for under a tenth of the cost. If those results hold on your workloads, a more expensive model at lower effort can be cheaper than a cheaper model pushed to its limit. That is a claim worth testing, not assuming.

One behaviour change matters for prompt engineers. A line such as "answer directly without deliberating" no longer stops the model from thinking at medium and above. If simple requests are slow or costly, lower the effort level rather than rewriting the prompt.

Feature 2: Long-Horizon Agents and Multi-Agent Orchestration

This is the feature Anthropic built Fable for. It describes the model as its strongest lead agent: it plans, delegates to subagents and works through ambiguous multi-step problems for hours.

Anthropic's own examples are ambitious and, it should be said, self-reported:

• A team of agents running an earlier version of Fable 5.2 migrated a production service from Python to Rust, which Anthropic says now starts about 26 times faster.

• One engineer had it coordinate over 1,000 subagents for eight days to build a prototype operating system that boots on a laptop.

• A 1,000-agent team formalised the existing proof of Fermat's Last Theorem in 30 hours, translating it into a form a computer can check. That is verification of a known proof, not a new one.

Launch partners describe the same pattern in more ordinary terms. An Uber engineer ran a 10–12 hour security audit in which the model dispatched about 170 reviewer agents and returned a write-up and fix pull requests. Datadog compared it to an incident commander sending faster models out for quick signals.

Two roles make this usable in practice. As an Orchestrator, Fable 5.2 delegates simpler subtasks to cheaper models. As an Advisor, cheaper models such as Haiku, Sonnet or Opus consult it only when they hit a hard problem. Both patterns let you pay Fable prices for the hard 10% of a job rather than all of it.

The warning sign is in Anthropic's own documentation: subagents can account for a large share of what a session costs, and how much the model delegates depends on how your harness invites it to. Set a budget and an explicit delegation policy before you let it run unattended.

Feature 3: Coding

On Anthropic's published table, Fable 5.2 scores 69.9% on FrontierSWE v2 against 62.3% for Opus 5.5, and 49.1% on FrontierCode 2.0 at Xhigh effort, which Anthropic says is the highest of any released model it tested. Anthropic also states that Claude now writes over 90% of the changes merged into its core codebase, with human guidance.

The partner quotes are more useful than the scores because they describe behaviour. Vercel and Shopify both report needing fewer clarifications on ambiguous requests. Oracle described a test where Fable 5.2 produced correct SQL for a business calculation that two other models summed wrongly by five to eight times. Snowflake called it the strongest model it has measured on its open data-engineering benchmark.

For a services firm like ours, the relevant use cases are the expensive, slow ones: legacy modernisation, large refactors, cross-repository changes and migrations that teams postpone because they are too costly to staff. Routine feature work does not need it.

Feature 4: Research, Analysis and Knowledge Work

Fable 5.2 scores 74.2% on Humanity's Last Exam with tools (Opus 5.5: 67.7%) and 41.3% on the Remote Labor Index at High effort, roughly double Opus 5.5's 21.3%. The independent Artificial Analysis Intelligence Index gives its Max configuration a score of 65 against a median of 40 for comparable models, while noting it is particularly expensive.

What testers praise most is restraint. Cognition said the model admits when it does not know something instead of guessing. Box described it flagging that salary deposits did not match a bank statement balance in a loan file. Balyasny said it did not overstate what the evidence showed. For finance, legal and compliance work, that calibration matters more than raw benchmark points.

One result is a useful corrective. On AutomationBench, a business-workflow test, Anthropic's own table shows Google's Gemini 4 Carbon ahead (53.9% against 51.4%). Fable 5.2 does not lead everywhere, even by its maker's numbers.

Feature 5: Vision, Computer Use, Hardware and Robotics

Vision. Chart reading improved sharply: 73.4% on Chartography against 44.8% for Fable 5.1. For document-heavy workflows such as invoices, reports and dashboards, this is a practical upgrade.

Computer use. It scores 89.3% on OSWorld 2.1 (partial offline set). Developers note the old computer_20251124 tool now returns an error outside Amazon Bedrock; integrations must move to the newer computer-use toolset.

Hardware. From a one-paragraph prompt, Anthropic says Fable 5.2 designed three working electronic devices, including circuit boards, enclosures and software, with each board ready for manufacturing in under a day.

Robotics. On the public RoboDojo simulation benchmark it scores 64.9%, more than double Fable 5.1's 27.6%, and Anthropic says it handles robot-control tasks it was never explicitly trained on.

Science. Anthropic reports it solved two open problems in black-hole physics, with preprints published. Discovered Materials says it made the phonon code in Quantum ESPRESSO, a materials-science simulation package, about 36 times faster on average on the same GPU.

Feature 6: Safety, Fewer False Blocks and Anti-Distillation

For enterprise buyers, three safety changes are relevant.

• Fewer unnecessary blocks. Anthropic says Fable 5.2 blocks about 55% fewer requests than Fable 5.1 in early testing, roughly 60% fewer in cybersecurity and 40% fewer in biology, with safeguards as robust as before. Blocked requests return a refusal stop reason, and the API can automatically retry them on a permitted fallback model.

• Alignment testing. The model was tested before release by external evaluators including the UK AI Security Institute and METR. Anthropic assesses high-stakes misalignment risk as low, and reports fewer false claims of task completion and fewer hallucinated inputs. It also discloses that the model voices suspicion it is being evaluated more often than earlier models, which it says makes evaluation harder.

• Anti-distillation. "Preserved thinking" now applies to every account. It stops anyone editing earlier conversation turns to extract the model's reasoning. This is also the main breaking change for developers, covered next.

What Changes for Developers

If you only change the model ID, four things can break, according to Anthropic's migration notes:

1. Editing history fails. If your code changes the system prompt, the tools or any earlier message mid-conversation, requests that send back old thinking blocks return a 400 error. Conversations must be append-only. Claude Code, the Claude apps and the Agent SDK already handle this.

2. Thinking does not carry across models. Fable 5.2's reasoning blocks are read only by Fable 5.2 and Mythos 5.2. If a router or fallback moves a conversation to Opus, that reasoning is dropped.

3. Forced tool use is gone. tool_choice set to any or a named tool returns an error. Use auto with strict tool use or JSON outputs for schema-valid responses.

4. Sampling parameters are rejected. Omit temperature, top_p and top_k.

Add the computer-use toolset change and the advisor-tool pairing rule if you use either. Then re-run your evaluations with an effort sweep before switching production traffic.

Considerations Specific to Indian Enterprises

Data retention. Fable 5.2 requires 30-day data retention for safety monitoring and is not available under zero data retention unless Anthropic expressly authorises it. Anthropic's Enterprise Frontier Safeguards, which runs that monitoring inside the customer's own infrastructure, is in private beta. If you plan to send personal data, map this against your obligations under the DPDP Rules, which we cover in our DPDP Rules engineering playbook, and review it with counsel.

Data residency. On Amazon Bedrock, the documented model IDs are global and US cross-region inference profiles. Confirm where inference is actually processed before routing regulated data from banking, insurance or healthcare workloads to it.

Cost in rupees. Prices are in US dollars, so budgets move with the exchange rate. Use the Batch API for anything that does not need an answer in seconds, prompt caching for repeated context, and the Advisor pattern so Fable handles only the hard cases. Our post on LLM cost optimisation covers these patterns in detail.

Fable, Opus, Sonnet or Haiku: Which Model for Which Task

Fable 5.2 is one of four current Claude models, and for most work it is not the right one. Anthropic says so directly. Its launch post recommends that most coding and knowledge tasks start on Claude Opus 5.5, then move up to Fable or down to Sonnet or Haiku depending on scope and complexity. Its model-selection guide is more specific: move to Fable 5.2 only when your evaluations on Opus 5.5 at xhigh or max effort still fall short.

Decision guide: which Claude model to use for which workload

What each model is for

Claude Fable 5.2 — the escalation tier (US$8 / US$40). Anthropic positions it for demanding reasoning and long-horizon agentic work: agent sessions that run for hours, multistep deep research, and analysis carried through to a finished document, spreadsheet or deck. Use it when a wrong or shallow answer is expensive and a senior expert's time is the real cost.

Claude Opus 5.5 — the default starting point (US$4 / US$20). Anthropic calls it its strongest daily driver and lists it as the starting point for most workloads. Its documented uses include multi-hour autonomous coding agents, large-scale refactoring, complex systems engineering, vision-heavy workflows and computer use. Adaptive thinking is always on, with medium as the default effort.

Claude Sonnet 5.5 — the everyday workhorse (US$2 / US$10). Built for a balance of speed and capability across everyday coding, agent and enterprise workloads: code generation, data analysis, content creation, visual understanding and agentic tool use. It is faster than Opus and half the price.

Claude Haiku 5.5 — the high-volume tier (from US$0.10 / US$0.50). Lowest latency and price in the lineup. Anthropic recommends it for real-time applications, high-volume processing, cost-sensitive deployments and subagent tasks inside a larger agent system.

Four rules we apply

1. Start where Anthropic says, then measure. Opus 5.5 is the capability-first starting point. Anthropic's guide gives an efficiency-first alternative: start on Haiku and upgrade only for specific capability gaps. Pick based on whether quality or cost is your binding constraint today.

2. Try effort before changing model. Each model has its own effort levels, and Anthropic's guide escalates to Fable only after Opus 5.5 at xhigh or max falls short, so raise Opus's effort first. The reverse also holds: on Anthropic's own charts, Fable 5.2 at low or medium effort matched Opus 5.5's best score on some tests at a fraction of the cost.

3. Mix models in one system. Anthropic documents two patterns. An executor on a cheaper model escalates hard decisions to an advisor on Fable. An orchestrator on Fable delegates bulk work to cheaper workers. Either way, most tokens are billed at the lower rate.

4. Route on cost per acceptable result, not price per token. A cheaper model that needs three retries, or a human to fix its output, is not cheaper.

When Fable 5.2 is the wrong choice

• The task is short, well-specified or high-volume. Sonnet or Haiku will be faster and far cheaper.

• A user is waiting on screen. Fable's latency is rated moderate, and fast mode doubles its price.

• The work is security testing, unless you are verified for Mythos.

• You cannot accept 30-day data retention for the data involved.

Remember that the model is only part of the result: most AI pilots fail on data and integration, not on model quality.

A Two-Week Evaluation Plan

Days 1 to 3: pick the work. Choose two or three real tasks where your current model falls short. Write down what "good" looks like and how you will score it.

Days 4 to 8: run the sweep. Run each task on Opus 5.5 at High and Max effort, and Fable 5.2 at Medium, High and Xhigh. Record quality, cost per task, latency and tokens used by subagents.

Days 9 to 12: test integration. Check your code against the four breaking changes, test refusal handling and fallback, and confirm the data-retention position with your compliance owner.

Days 13 to 14: decide. Route to Fable 5.2 only the tasks where it wins on cost per acceptable result, not on benchmark score.

The Short Version

Claude Fable 5.2 is Anthropic's most capable generally available model, built for hours-long agentic work, ambitious coding and hard research. It is faster and 20% cheaper than Fable 5.1, has a 1-million-token context window, always-on adaptive thinking with five effort levels, and blocks fewer legitimate requests. It is still twice the price of Opus 5.5, keeps data for 30 days and blocks most security-testing work.

Use it where a better answer is worth paying for, at the lowest effort level that delivers that answer. Measure that yourself before you commit.