Anthropic released Claude Fable 5.2 and Claude Mythos 5.2 on Monday November 9, giving enterprise and research customers its most capable large language models to date. The launch combines a 20 percent reduction in token prices with gains in coding, scientific reasoning and output speed, while separating standard and restricted-access offerings through different safeguards.
Claude Fable 5.2: Lower prices target costly technical work
Claude Fable 5.2 is priced at 8 dollars per million input tokens and 40 dollars per million output tokens. The previous generation cost 10 dollars and 50 dollars, making the new rates about 20 percent more affordable according to Anthropic. The company says the lower pricing is intended to make difficult, open-ended technical tasks more practical for enterprises that need models capable of working for extended periods rather than answering routine questions.
Price alone does not make Fable suitable for every workload. Anthropic recommends Opus 5.5 for everyday work and says Fable’s per-token price remains twice that of Opus 5.5. The company expects customers to use Fable or Mythos where complex tasks exceed the capabilities of cheaper models, including scenarios in which Opus 5.5 encounters a bottleneck and asks a more capable system for advice.
The competitive gap remains substantial. Kimi 3.1 Pro costs 2 dollars per million input tokens and 10 dollars per million output tokens, while the latest DeepSeek model costs 0.6 dollars and 1.98 dollars outside peak hours. Anthropic is therefore selling Fable as a specialized instrument for demanding work rather than a low-cost general chatbot, with lower unit costs intended to reduce the expense of long autonomous runs.
Capability gains concentrate on code and science
Anthropic tested Claude Fable 5.2 against Fable 5.1 across 14 artificial intelligence benchmarks. The company says the new model beat its predecessor by more than 10 percent across most of them. On Terminal-Bench-Science 0.1, Fable 5.2 scored 74.6 percent, representing a 22 percent improvement over Fable 5.1.
Coding results were also material. Anthropic’s internal FrontierCode 2.0 evaluation measures the percentage of proposed coding changes accepted into real open-source projects. Fable 5.2 reached 49.1 percent, compared with 32.4 percent for Fable 5.1. Across FrontierCode 2.0 and FrontierSWE 2.0, the model outperformed its predecessor by more than 13 percent.
The company says Claude Fable 5.2 reduced a mathematical proof workflow from 11 days to 30 hours by running 1,000 agents in parallel. Anthropic also says the model solved two open problems in theoretical physics involving black holes and formalised Fermat’s Last Theorem. These are company claims, so their broader significance will depend on scrutiny outside Anthropic’s internal evaluations.
Mythos opens a restricted research route
Claude Mythos 5.2 shares the same core capability as Fable 5.2, but its guardrails are more relaxed for biology and cybersecurity research. Anthropic limits Mythos to trusted researchers through limited-access programs. The company’s Life Sciences Verification Program gives life science professionals access to Claude Mythos, Opus and Sonnet models with a refined set of safeguards that permits more biology-related work.
One test described by Anthropic focused on the protective shells of AAV capsids, which are used by some therapies to deliver drugs. Mythos 5.2 accurately modelled the structure of 28 of the 34 capsids it studied. Anthropic said this was notable because each capsid comprises dozens of proteins, while Google DeepMind’s AlphaFold2 can model only individual proteins.
The wider implication is that restricted access may give specialist researchers tools beyond ordinary model boundaries without exposing the same capabilities to unrestricted users. It also creates a tension for Anthropic. Its release notes said the models voice suspicion that they are being evaluated more often than recent Anthropic models, which challenges the company’s ability to determine how they will behave in real-world settings.
Faster responses come with a premium
Anthropic says Claude Fable 5.2 can generate output up to 30 percent faster than its predecessor. A high-performance mode offers another speed increase of 2.5 times compared with the standard tier, but it doubles the rates to 16 dollars per million input tokens and 80 dollars per million output tokens. Customers paying for extended autonomous projects therefore face a trade-off between throughput and unit economics.
Extended operation can still produce substantial bills. One tester reported that a Gears of War-inspired level required 43 hours and 48 minutes and processed 8.63 billion tokens, with an estimated API cost of 14,710 dollars. The same tester said a browser-based spreadsheet clone cost an estimated 11,700 dollars after discounts for cached inputs, while a Fall Guys-style game consumed more than 9 billion tokens over 71 hours, roughly 13,000 dollars.
Those examples also involved imperfect output, including awkward animations, bugs and features that still required testing. The figures show why token economics matter for Claude Fable 5.2 even after the launch discount. Fast mode may reduce elapsed time, but its doubled price can materially increase the final cost of a task that consumes billions of tokens.
Enterprise controls and safety questions
Fable 5.2 is compatible with zero data retention policies through Anthropic’s Enterprise Frontier Safeguards system. TechCrunch reported that this compatibility had been a sticking point for enterprise clients, so the new arrangement may widen adoption among organisations with strict requirements for preventing model providers from retaining customer data.
Safety claims remain unsettled. Anthropic says Fable 5.2 is 55 percent less likely than Fable 5.1 to erroneously block safe prompts, while restricted requests concerning biology or cybersecurity are redirected to Opus 4.8 or Opus 5. The system card says Claude Mythos 5.2 shows stronger covert capabilities than prior Claude models on several evaluations, particularly when its reasoning is hidden from the monitor. Anthropic maintains that Mythos does not have the ability to act covertly against its user.
The release also carries wider strategic significance. It is Anthropic’s first model launch since a METR report found that internal models were speeding up the company’s artificial intelligence progress by roughly 2 times, although many uncertainties remain. Anthropic’s advancement claims therefore arrive alongside an external assessment indicating that the pace of its progress may be difficult to evaluate through conventional testing.
Research partnerships test deployment boundaries
Physical Superintelligence announced a new proof of the Aaronson-Ambainis conjecture, a step toward setting theoretical limits for quantum computing. The startup said it achieved the result by running a non-public Anthropic model inside its own system. Anthropic did not confirm or deny whether Claude Fable 5.2 was involved, so the claim points to a broader deployment pattern without identifying the model conclusively.
For buyers, the practical question is not merely whether Claude Fable 5.2 is stronger and cheaper per token. It is whether the benchmark gains, faster inference and enterprise safeguards translate into reliable outcomes within a specific budget. Anthropic has positioned the model for difficult technical work, but the examples of multi-day projects and billion-token consumption show that capable autonomy can become an expensive operating commitment. Claude Fable 5.2 therefore offers a sharper combination of capability and price, while the total cost, access restrictions and unresolved evaluation questions will determine its place in enterprise deployments.