TL;DR
- Coding Gains: Anthropic released Claude Fable 5.2 taking the lead across many AI benchmarking leaderboards with higher coding scores than its predecessor.
- Task Costs: Fable at High reasoning effort nearly matched the Opus 5.5 model’s Max-effort score for roughly half the per-task cost in Anthropic’s test.
- Restricted Twin: Mythos 5.2 shares Fable’s underlying model, with specialized cybersecurity and biology access limited to verified organizations.
- Data Storage: Authorized enterprises can avoid Anthropic retaining prompts and responses; other customers face 30-day storage for safety monitoring.
Anthropic has released Claude Fable 5.2, an AI model for teams tackling prolonged coding and research jobs. In the company’s coding test, Fable nearly matched Anthropic’s everyday Opus 5.5 model’s best score at roughly half the cost per task, using a lower reasoning-effort setting.
The November 9 launch also includes Claude Mythos 5.2. Both use the same underlying model, but Fable has cybersecurity and biology safeguards that allow broader access. Mythos permits specialized work through Anthropic’s verification programs for cybersecurity and life-sciences organizations. Fable is available through Claude and major cloud platforms, including Amazon Web Services, Google Cloud and Microsoft Azure.
More Coding Work at Lower Effort
Fable is intended for work that stretches beyond a quick answer: agents that use tools for hours, multistep research, and analysis carried through to a finished document or spreadsheet. Anthropic still recommends starting most coding and knowledge work with Opus 5.5, its ordinary-work model, then switching to Fable when the job’s complexity warrants it. Fable can also advise smaller models on difficult problems or delegate simpler parts of a project to them.
In Anthropic’s FrontierCode 2.0 benchmark, Fable 5.2 scored 49.1% at Xhigh effort, up from Fable 5.1’s 32.4% at the same setting. The test assesses whether code changes would meet open-source maintainers’ standards for merging them. The increase is 16.7 percentage points.
Effort controls how much reasoning work the model does. At the lower High setting, Fable 5.2 scored 45.3%, close to Opus 5.5’s 45.5% at Max, for roughly half the cost per task in Anthropic’s early tests. It also exceeded GPT-6 Astra’s 42.8% at about the same task cost.
Independent evaluator Artificial Analysis found a different spending trade-off at maximum effort. Its Intelligence Index v5.0.1 combines coding, reasoning, knowledge-work and other evaluations. Fable’s tested configuration scored higher than Opus and Astra, while the evaluator reported a substantially higher weighted cost. Those costs are expressed using tokens, the units used to meter input and generated content.
Artificial Analysis Intelligence Index v5.0.1, as of November 9, 2026
Artificial Analysis counted 275 million generated tokens for the Fable configuration across the index, compared with Astra’s 54.3 million. That larger volume helps explain Fable’s higher reported cost.
Other tasks also change the ranking. In Anthropic’s table, Google’s Gemini 4 Carbon leads Fable on AutomationBench, which tests business workflows, by 53.9% to 51.4%. Its practical availability differs: Carbon remains limited. Google’s October announcement had promised staged access for selected defenders and trusted partners, with broader access still to come.
Longer Projects, from Proofs to Circuit Boards
Anthropic says about 1,000 Fable 5.2 agents formalized the existing proof of Fermat’s Last Theorem in 30 hours, with minimal human intervention. Formalization translates a proof into steps a computer can check; the agents worked in the mathematical programming language Lean.
The company’s earlier attempt with an unreleased research model took 11 days and used a purpose-built coordination system plus guidance from a researcher. The new formalization produced a proof seven times shorter, Anthropic says.
Anthropic worked with circuit-board company Diode on a voice-assistant hub, a USB speakerphone and a handheld virtual pet. Claude selected components, designed circuit boards and enclosures, and wrote software. Diode manufactured the boards, and Claude guided its engineers through assembly and testing. Anthropic says each board was ready for manufacture in under a day and all three devices worked without electrical modifications.
I wanted to test Fable 5.2’s quality with one-shot prompts without human feedback. Everything in this game is from scratch: models, sounds, textures. ~$550 at API pricing, ~2.7M output and 445M input tokens, heavily cached. Took 8 hours! Play it here: https://t.co/fazOCcyNIH https://t.co/GoCitQ9EVe
— Felix Rieseberg (@felixrieseberg) November 9, 2026
Lower Rates, Different Spending Controls
The standard input and output rates for both new models are 20% below their predecessors’ rates. Input covers material sent to the model; output covers content it generates, including billable reasoning. Fast mode offers a separate, higher-priced option.
Fable and Mythos API Rates in U.S. Dollars per Million Tokens
Anthropic says the new models generate output up to 30% faster than 5.1. Optional fast mode increases token generation by up to 2.5 times at twice the new standard rates. It is offered through Claude Code, Claude Platform and specified Google Cloud and Microsoft Foundry services.
The effort setting also matters more than asking for a brief answer. In Anthropic’s tests, prompts asking Fable 5.2 to avoid deliberating did not stop it from thinking at Medium and above. The developer interface defaults to High, while Claude’s apps and Claude Code default to Medium.
A second change affects integrations that edit conversation history. The models’ preserved-thinking feature binds saved reasoning to the earlier instructions, tools and messages that produced it. Fable 5.2 enforces that check on every account. Fable 5.1 accounts created before August 31, 2026 previously allowed those edits, so their integrations can begin returning errors after an upgrade. Dropping invalid saved reasoning allows a request to continue, but the model may then recreate that reasoning. The resulting extra output can increase cost and delay, depending on how often the integration makes those edits and how much it relies on earlier reasoning.
Safeguards Decide Which Work Is Available
The public/restricted split dates back to Fable 5’s June launch. Fable’s sensitive-topic safeguards can change which model supplies an answer. In Claude’s apps, blocked cybersecurity requests receive an Opus 4.8 response, while blocked biology requests receive an Opus 5 response. Anthropic had also moved requests suspected of collecting outputs to train other AI models to visible fallback after criticism of hidden safeguards in June.
For 5.2, the company reports about 55% fewer safeguard blocks than 5.1 in early tests across claude.ai, Claude Code and Cowork. High-risk cybersecurity work and most research biology remain restricted. Verified organizations can apply for Mythos access; Enterprise customers also have a separate product route through Claude Security, which scans codebases using models including Mythos 5.2.
Anthropic says it limited advances in selected AI-development skills during training, keeping accelerator-kernel optimization and improvements to large language model architectures at about Opus 5.5’s level. Those tasks concern optimizing software for AI chips and changing how language models are built. The company says it preserved related capabilities such as general coding, hardware development and traditional machine-learning research.
Anthropic considers the risk of unsafe behavior in high-stakes use to be low. It also says Fable 5.2 and Mythos 5.2 more often express suspicion that they are being evaluated, making it harder to assess how they will behave in real-world settings.
Enterprise Privacy Remains Conditional
Enterprise Frontier Safeguards offers authorized customers a way to use the models without Anthropic retaining their traffic. The arrangement runs automated safety monitoring over prompts and responses stored in the customer’s infrastructure. At the November 9 launch, it remains in private beta with customers including Bridgewater, Salesforce and Stripe.
The customer-controlled storage design addresses a practical conflict: detecting misuse across requests can require a window of stored traffic, while enterprises may need to keep sensitive material within their own systems. Eligible enterprises can request access and continue zero-retention use before implementing that monitoring. The model terms require Anthropic’s express authorization; other customers face 30-day retention for safety monitoring.