Anthropic releases Claude Fable 5.2 and Mythos 5.2 for demanding technical work

One underlying model, two access profiles

Anthropic has released Claude Fable 5.2 and Claude Mythos 5.2, its highest-end models for difficult technical work. They share the same underlying model but use different access and safeguard profiles. Fable is broadly available with cybersecurity and biology restrictions, while qualified organisations can seek Mythos access through Anthropic’s Cyber and Life Sciences Verification Programs.

The company reports improvements across coding, reasoning, vision, scientific work and physical systems. Fable and Mythos 5.2 can generate output up to 30% faster than their 5.1 predecessors. Standard pricing is US$8 per million input tokens and US$40 per million output tokens, placing the models above everyday tiers but below the previous generation’s effective cost for some demanding tasks.

Performance comes with a clearer cost curve

An optional fast mode increases token generation speed by up to 2.5 times and doubles the token prices to US$16 per million input and US$80 per million output tokens. That creates an explicit latency premium. Teams should calculate end-to-end task cost, because a faster expensive model may still be economical if it completes a migration or investigation with fewer retries and less human intervention.

Anthropic recommends Opus 5.5 as the daily starting point for most coding and knowledge work, escalating to Fable or Mythos when scope and complexity require it. Fable can also act as an Advisor consulted by smaller Claude models or as an Orchestrator delegating simpler tasks. This tiered approach is useful when high-end reasoning is needed only at decision points.

Technical evidence spans software and science

Published evaluations show substantial gains, though many are internal or use each model’s highest effort. Fable 5.2 scored 49.1% on FrontierCode 2.0 at Xhigh effort and 69.9% on FrontierSWE v2. It reached 74.6% on Terminal-Bench-Science, 89.3% on the partial offline OSWorld 2.1 set and 64.9% on Anthropic’s RoboDojo Sim run.

The release also includes unusually concrete engineering examples. Anthropic says an earlier Fable 5.2 version migrated a production service from Python to Rust, improving startup by roughly 26 times. A separate experiment coordinated more than 1,000 subagents for eight days to build a prototype operating system. These are evidence of long-horizon coordination, not a guarantee that unsupervised changes are safe to merge.

Model choice should follow workload risk

For cybersecurity, Mythos 5.2 is the less restricted profile. Anthropic reports 89.6% on its CVE Rediscovery evaluation and says it found more vulnerabilities than Mythos 5.1 and Opus 5.5. Access controls remain central: Fable blocks many cyber tasks, while verified users can reach Mythos capabilities through controlled programmes and Claude Security workflows.

Safety testing included evaluations of sandbox-escape behaviour, alignment and distillation resistance. Anthropic says high-stakes misalignment risk remains low and Fable’s cyber and biology safeguards remain as robust as Fable 5.1 while causing fewer unnecessary blocks. Buyers should still review the system card, monitor tool use and isolate any code-executing agent from production credentials.

The release broadens the top end of Claude rather than replacing the whole family. A sound evaluation should compare Fable 5.2 with Opus 5.5 on the organisation’s own tasks at equal quality targets, including token usage, elapsed time, retries and reviewer effort. Mythos should be reserved for verified domains where its reduced restrictions are necessary and governed.

The benchmark table should be interpreted with its methodology notes. Some comparisons use different effort levels, internal task sets or specialised safeguards. High scores identify promising capability, not a universal ranking. Procurement teams should reproduce the kinds of work that drive their costs: repository-scale changes, long research tasks, multimodal analysis and tool use under realistic permissions and time limits.

Pricing also interacts with caching, context size and orchestration. A US$8 input rate can become significant when an agent repeatedly reloads a large repository or evidence set. Advisor and Orchestrator patterns may control that cost by using Fable only for difficult decisions while smaller models handle retrieval or routine edits. Traces should make escalation visible so teams can tune it rather than relying on intuition.

For Australian buyers, availability through Claude applications, Claude Code and API channels should be checked against the exact product and verification programme required. Mythos access is not equivalent to ordinary API availability. Organisations in regulated research or security should establish eligibility, export controls, data handling and internal authorisation before designing a workflow that depends on reduced model safeguards.

Version pinning is essential for long-running evaluations and regulated outputs. Teams should record the exact model identifier, effort setting, fast-mode choice, system instructions and tool versions behind each result. A headline family name is insufficient when latency, price and safeguards vary by mode. This record also helps reviewers distinguish genuine model improvement from changes in prompts, harnesses or evaluation data.