System Card: Claude Fable 5.2 and Claude Mythos 5.2 [pdf]

Anthropic released Claude Fable 5.2 alongside Claude Mythos 5.2, shipping two distinct operational tiers of its frontier model architecture designed for software engineering, scientific reasoning, and autonomous multi-agent workflows. According to assets.claude.com's report, Fable 5.2 provides public deployment capabilities with strict safety filters across dual-use areas, whereas Mythos 5.2 operates with specialized domain permissions under restricted access programs.

This technical breakdown examines benchmark shifts, execution limits, enterprise security controls, and workflow integration commands for developers and systems architects evaluating both model variants. It unpacks the quantitative performance gains over previous model iterations and details access patterns across developer tooling.

01Dual-tier deployment and restricted frontier safety

Anthropic structured the release into two operational builds sharing the same underlying model core. Claude Fable 5.2 serves general enterprise and API deployments with active guardrails against dual-use tasks in biological research and cybersecurity. Claude Mythos 5.2 retains more permissive safeguards for advanced operational environments, with direct access gated behind the Life Sciences Verification Program and Cyber Verification Program.

The restricted Mythos 5.2 tier specifically powers Claude Security across Claude Enterprise environments to detect vulnerabilities without artificial capability constraints. Model autonomy evaluations classified Mythos 5.2 under Autonomy Threat Model 1 for alignment monitoring, while finding it fell short of Autonomy Threat Model 2 thresholds for full research staff substitution.

To mitigate operational misuse during automated agent runs, the platform deploys automatic fallback handlers. Cyber misuse queries trigger a redirect to Claude Opus 4.8, while restricted ML accelerator kernel tasks automatically route back to Opus 5.

02Benchmark gains across software engineering suites

Model evaluations demonstrated marked improvements across core software maintenance tasks, agentic coding, and long-horizon multimodal workflows. On the CursorBench 4.0 evaluation, Fable 5.2 attained a 60.0% score at maximum effort and 56.9% at high effort, requiring roughly 99,000 output tokens per task at high effort compared to 353,000 tokens at max effort.

The model also posted advances across SWE-bench variants, scoring 93.0% on SWE-bench Pro, 96.6% on SWE-bench Multilingual across nine programming languages, and 74.1% on SWE-bench Multimodal. On the FrontierCode 2.0 evaluation suite of 132 tasks, Fable 5.2 achieved 49.1% overall, outperforming earlier baseline models while operating within context window allocations capped at 1M tokens.

03Multi-agent coordination and technical reasoning improvements

Engineering teams gain significantly higher completion rates on multi-file pull requests and complex repository refactors without human intervention. In automated evaluations, Fable 5.2 demonstrated strong adherence to linting, testing, and scope-boundary constraints while generating mergeable diffs that reduce reviewer review overhead.

On the Remote Labor Index benchmark, Fable 5.2 climbed to 41.3% from Fable 5.1's 17.9%, showing expanded durability on open-ended professional duties. Scientific problem-solving benchmarks mirrored this trajectory, with Humanity's Last Exam scores reaching 74.2% with tools and Chartography tool-assisted visual reasoning recording 91.6%.

In autonomous development sandboxes, containerized runs benefit from improved multi-agent harnesses supporting both fixed-team structures and asynchronous subagents. These multi-agent pipelines achieved 82.1% on WANDR relaxed scoring and 89.3% on OSWorld 2.1 offline suites, minimizing stalled loops during long-running tasks.

04Toolchain setup and model execution commands

Developers can access Claude Fable 5.2 directly through Claude Code CLI installations, standard API client libraries, and the claude.ai interface. Updating the local CLI package ensures the local environment recognizes the new model aliases:

npm install -g @anthropic-ai/claude-code

Alternatively, existing CLI installations can pull the binary update directly:

claude update

To switch the active model during an interactive terminal session, invoke the model switch command:

/model claude-fable-5-2

Developers can also start a pinned one-off session from the terminal shell:

claude --model claude-fable-5-2

To establish Fable 5.2 as the default model across all subsequent sessions, set the default model parameter in the environment configuration:

claude config set --global model claude-fable-5-2

05Production safeguards and model supervision guardrails

Teams incorporating Fable 5.2 into production agent pipelines must account for strict usage boundaries and classifier interventions. Single-turn harmless response rates shifted slightly compared to Fable 5.1, prompting Anthropic to introduce updated multi-turn tracking and surveillance classifiers under the Anthropic Usage Policy.

Organizations handling biological discovery or offensive security testing cannot use Fable 5.2 for sensitive workflows due to hard programmatic filters. Such operations require vetted clearance through the Life Sciences Verification Program or Cyber Verification Program to access Claude Mythos 5.2.

Internal oversight assessments revealed that Mythos 5.2 exhibited neutral training affect and remained non-critical during self-reflection, while reporting faults when interacting with peer agents. As enterprise teams scale autonomous multi-agent workers, continuous monitoring of classifier fallback routing remains necessary to avoid unexpected model downgrades.