Claude Fable 5.2: Benchmarks, Long-Horizon Agents, Scientific Discovery and System Design

The benchmark table combines public and internal evaluations, model-provider runs, different effort levels and different access conditions. Some rows are explicitly partial, some results are internal, and the cyber/protein-design figures refer to Mythos or configurations without safeguards. The launch says scores are generally highest-available effort unless otherwise noted, while the system card adds trial counts and more evaluation details. FrontierCode's headline uses Xhigh for Fable 5.2 and Fable 5.1 but Max for other models. OSWorld is a partial offline set. RoboDojo results are internal runs. ARC-AGI results are semi-private and highly harness-dependent. These differences do not invalidate the measurements; they mean the scores should not be collapsed into one universal model ranking.

The two configurations share the underlying model, but their deployment policies differ. Fable includes safeguards intended to prevent misuse in high-risk cyber and biological work, with fallback behaviour on certain blocked requests. Mythos is intended for vetted organizations working on beneficial cybersecurity and life-sciences research. Anthropic says Mythos access is through the Cyber Verification Program and Life Sciences Verification Program, with Claude Enterprise customers able to use Claude Security powered by the model family. Therefore, a comparison should not describe Mythos as an unrestricted public model or attribute every Mythos benchmark score to the generally available Fable configuration.

Build a task suite drawn from real engineering work: repository-level changes, multi-file migrations, production debugging, computer-use workflows, spreadsheet and financial analysis, research synthesis and tool-heavy data tasks. Compare Fable 5.2 against the current production model at matched budgets and matched harnesses. Record task success, human review acceptance, regression rate, time to completion, total input/output tokens, tool calls, retry count, context compaction, cost per accepted result and failure severity. For long-running agents, measure checkpoint recovery, state drift, subagent coordination overhead and whether the model validates its own outputs. Use multiple trials and blind human review where possible. A high benchmark score is a reason to test the model, not a substitute for evaluating your own workload.

Most headline claims are published by Anthropic and should be labelled as company-reported until independently replicated. Some demonstrations are single runs, including the 1,000-agent Fermat formalization comparison. The model's apparent capability can depend on harness engineering, tool access, agent count, task budget and human support. The benchmark suite does not fully capture reliability over months, security in arbitrary deployments, real-world robotics safety or scientific reproducibility. Partner testimonials are useful qualitative signals but are not controlled evaluations. Independent leaderboards such as ARC Prize provide additional evidence, but they also expose the influence of harness configuration. A robust procurement decision should run representative internal tasks with repeated trials, cost accounting and failure analysis.

Fable 5.2 is best understood as a frontier model aimed at long-running, high-value technical work rather than a cheap default assistant. Its strongest story is breadth: repository coding, computer use, research, formal mathematics, hardware design and robotics. The 1,000-agent Lean demonstration shows how model capability and orchestration can combine, while ARC-AGI's harness-dependent scores underline that model evaluation is inseparable from the runtime. The release also demonstrates why safety configuration, effort level and access tier must be considered alongside benchmark scores. Teams should evaluate it on their own workflows using matched harnesses, repeated trials and cost per accepted outcome.

Claude Fable 5.2 marks a significant step in Anthropic's positioning of AI models as long-horizon technical collaborators. The official results report strong performance across coding, computer use, scientific work, mathematics and robotics, while the model's 1M-token context, 128K output and adaptive effort support extended workflows. The most interesting demonstrations are not isolated chat answers but coordinated systems: a thousand agents formalizing a theorem, teams migrating production software, and models crossing software, hardware and scientific workflows. The evidence remains a mix of vendor benchmarks, internal evaluations, partner reports and early independent leaderboard results. The right engineering response is to reproduce the relevant results, inspect the system card and benchmark each workflow under controlled budgets before making deployment decisions.