Fable 5.2 matches Opus 5.5 on FrontierCode at half the cost

Claude Fable 5.2 and Claude Mythos 5.2 arrived on November 9. Fable’s price drops to $8 per million input tokens and $40 per million output tokens, down from $10/$50, and in Anthropic’s early testing Fable 5.2 at High effort scored about level with Opus 5.5 at Max effort on FrontierCode 2.0, for roughly half the cost per task. That saving was measured on one benchmark at High effort. Per token, Fable still costs twice Opus 5.5’s $4/$20, and Anthropic still recommends starting most coding and knowledge work on Opus 5.5.

Fable and Mythos are the same underlying model under different safeguards. Fable carries strong cybersecurity and biology safeguards, and that is what lets Anthropic offer it to everyone: it is out today on all paid Claude plans, on the Claude Platform and on AWS, Google Cloud and Microsoft Azure. Mythos 5.2 stays limited to organizations in Anthropic’s Cyber and Life Sciences Verification Programs. Anthropic’s platform docs say it has the same specifications and pricing as Fable, including a 1M-token context window and 128K max output. Both models generate output up to 30% faster than the 5.1 versions.

“All paid plans” has come with conditions before. At the end of October, Anthropic’s pricing page listed Fable on Max and Ultra up to half the weekly limit, and Pro users paid for it with usage credits.

Which effort, which model

Medium is now the default effort level in Claude apps and Claude Code, and Anthropic says Fable 5.2 at Medium handles challenging tasks with fewer tokens than Fable 5.1. With Fable 5.1, Anthropic’s launch post listed High as the default in Claude Code. So Claude Code sessions on Fable now start at Medium rather than High, while the result that matched Opus 5.5 at Max was measured at High. Of the two savings in the announcement, only the price cut is on the rate card. The half-cost figure comes from early testing on a single benchmark.

At Xhigh effort, Fable 5.2 scored 49.1% on FrontierCode 2.0, which Anthropic calls the highest of any released model it tested. Opus 5.5 scored 45.5%, GPT-6 Astra 42.8%, Gemini 4 Carbon 41.1% at Medium effort, and Fable 5.1 32.4% at Xhigh. FrontierSWE v2 shows a similar order: Fable 5.2 at 69.9%, Astra at 65.5% and Opus 5.5 at 62.3%. Unless the table notes otherwise, each score is the model’s highest available effort with safeguards on.

Anthropic’s own ladder starts with Opus 5.5, which it calls “our strongest daily driver”. From there it suggests moving up to Fable or Mythos for the most ambitious, open-ended technical work, or down to Sonnet 5.5 or Haiku 5.5. Fable 5.2 can also act as an Advisor that Haiku, Sonnet or Opus models consult on harder problems, or as an Orchestrator that hands simpler tasks to smaller models.

A new fast mode for both models speeds up token generation by up to 2.5 times and costs $16/$80, double the standard rates. The New Stack lists it in Claude Code, on the Claude Platform, through Claude Models on Google Cloud and through Anthropic-hosted Microsoft Foundry. The outlet also puts GPT-6 Astra at $10/$50 at standard rates for prompts of up to 272,000 input tokens. On those rates, Fable’s standard pricing is 20% below OpenAI’s flagship.

Where it doesn’t lead

Gemini 4 Carbon beats Fable 5.2 on AutomationBench, a business-workflow test, 53.9% to 51.4%. On the knowledge-work benchmark AA-Briefcase v1.1, Fable leads Opus 5.5 on the rubric score, 66.1% to 61.2%, but Opus is ahead on Elo, 1800 to 1785. On Terminal-Bench-Science 0.1, Fable 5.2 scores 74.6% and Mythos 5.2 scores 75.3%. Two of the rows need context. Anthropic ran RoboDojo Sim internally, where Fable 5.2 scored 64.9% against 39.4% for Opus 5.5. And on CyScenarioBench, each lab’s scores come from its own unique tasks.

The New Stack reports that Anthropic deliberately limited some AI-development capabilities during training. In accelerator kernel optimization and work on large language model architectures, the models perform at about Opus 5.5’s level. The company says those limits don’t touch agentic coding, hardware development or traditional machine learning research.

Fewer blocks, same lines

Anthropic says Fable 5.2’s cyber and biology safeguards are as robust as Fable 5.1’s but produce fewer unnecessary blocks. According to The New Stack, in early testing across Claude.ai, Claude Code and Cowork, Anthropic saw about 55% fewer blocks overall: about 60% fewer in cybersecurity and 40% fewer in biology. The rules themselves haven’t moved. Fable still blocks penetration testing, exploit generation and most research biology tasks, and Anthropic warns that its safeguards will block most cybersecurity-related tasks. A blocked cybersecurity query gets its answer from Opus 4.8 instead, and a blocked biology query from Opus 5.

Security teams outside the verification program have another route. Claude Enterprise customers can use Claude Security, which scans codebases for vulnerabilities and now runs Mythos 5.2, even if they aren’t members of the Cyber Verification Program. A recent update lets it watch repositories as code changes, track findings until they’re fixed and connect to a team’s tools through an API and MCP.

On Anthropic’s internal CVE Rediscovery evaluation, Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1. One security company found that one in six confirmed bugs from Mythos 5.2 were missed by every other frontier model it tested. Another tester found that earlier models had overstated the severity of about 1 in 20 findings, and Mythos 5.2 overstated none.

Developers on the API get two other changes, The New Stack reports. Preserved thinking, which limits edits to Claude’s earlier context that could be used to pull out its reasoning for distillation, is expanding to all accounts. Enterprise Frontier Safeguards is entering private beta, with Bridgewater, Salesforce and Stripe among the early customers. It runs automated safety monitoring on prompts and responses stored in the customer’s own infrastructure, so it works with zero data retention. Eligible customers can keep using Fable 5.2 and Mythos 5.2 with zero data retention before they set it up. For everyone else, Anthropic keeps data for 30 days for safety monitoring. Before release, the model was also tested by external evaluators, including the AI Security Institute and METR.

What the long runs produced

Most of Anthropic’s showcase involves agents running for hours or days. A team of agents on an earlier version of Fable 5.2 moved a production service from Python to Rust, and the service now starts about 26 times faster. In another test, starting from an empty folder, Fable 5.2 coordinated more than 1,000 subagents for eight days and built a prototype operating system that boots on a laptop, with drivers, permanent file storage, windows, a login screen and an installer. A team of 1,000 Fable 5.2 agents formalized Fermat’s Last Theorem in Lean in 30 hours, turning the proof into a form a computer can check step by step, with minimal human intervention. With the same setup, Opus 5.5 finished 26% of the work in 18 hours and made little progress after that. The New Stack notes that the comparison was a single run per model.

Anthropic’s other examples come from outside software. It says Fable 5.2 solved two open problems in black-hole physics, and that Mythos 5.2 found a way to predict the structure of a whole gene-therapy capsid, modelling 28 of the 34 it attempted. Working with circuit-board company Diode, Fable 5.2 also designed a voice-assistant hub, a USB speakerphone and a handheld virtual pet: it picked the components, designed the boards and enclosures and wrote the software. According to Anthropic, all three devices worked without electrical modifications.