Claude Fable 5 is Anthropic's first publicly available Mythos-class model, released June 9, 2026 (source: Anthropic announcement). The current shipping version is Fable 5.1 (model ID claude-fable-5-1), priced at $10 per million input tokens and $50 per million output tokens, with a 1M context window and 128K max output (models overview, checked 2026-10-20).
The baseline has moved. At launch the comparison was Opus 4.8; the current workhorse is Opus 5.5, which Anthropic says "performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5" at $4/$20 (Opus 5.5 announcement). So the question is no longer whether to upgrade from Opus 4.8, but how to split work between Fable 5.1 and Opus 5.5.
It also ships with safety classifiers that intercept requests related to cybersecurity, biology/chemistry, and model distillation (less than 5% of sessions on average, per the launch announcement). At launch, flagged requests were downgraded to Opus 4.8.
Should you switch? Pick Fable 5.1 when you need the strongest reasoning and long-horizon agentic work and can absorb both the topic restrictions and the higher unit price. For everyday coding and knowledge work, start with Opus 5.5: Anthropic positions it as Fable 5.1-level on most work, and its API price is 40% of Fable 5.1's ($4/$20 versus $10/$50). If your work is cybersecurity or biology/chemistry research itself, go through the trusted-access programs rather than trying to prompt around the classifiers.
Model Tier
Mythos is a new model tier above Opus. The first one, Claude Mythos Preview, was released in April 2026 through Project Glasswing exclusively for cybersecurity defenders. Fable 5 is the first publicly accessible Mythos-class model—same base model as Mythos 5, but with safety classifiers added.
Mythos 5 has partial safety restrictions lifted (Glasswing partners bypass the cybersecurity classifier, biology researchers bypass the biology/chemistry classifier), but is currently limited to Project Glasswing partners and select biology researchers.
Published Evaluations and Cases
Everything below comes from the customer cases and benchmarks listed in Anthropic's 2026-06-09 announcement plus the third-party reports that announcement cites. This is not AIGCDev hands-on testing, and we did not independently reproduce any of it.
The shared limitation: every figure is self-reported by the vendor or its partners. No methodology, sample size, eval script or failure case is public, and the Stripe, Cognition, Hebbia and IMC results describe those companies' own workloads — they do not transfer to your tasks. "Source" below is who made the claim in the announcement; "Evidence type" is what kind of claim it is.
| Claim | Source | Evidence type |
|---|---|---|
| Full migration of a 50M-line Ruby codebase that would have taken a team over two months, done in a day | Stripe (Anthropic customer) | Customer case, no before/after migration data |
| Top score among frontier models at medium effort on FrontierCode | Cognition (Anthropic partner) | Partner benchmark, no full leaderboard or scoring method published |
| Highest score on a finance benchmark, leading on document reasoning, chart reading and problem solving | Hebbia (Anthropic customer) | Customer case, no sample or scoring detail published |
| Near-perfect across a trading analysis evaluation | IMC (Anthropic customer) | Customer case, eval set not public |
| Persistent memory tripled both the improvement and the final-act completion rate versus Opus 4.8 | Anthropic internal test | Vendor self-test on Slay the Spire, a game environment that does not represent production tasks |
| "Substantially outperforms almost all other public models I've used, consistently" | Ethan Mollick (University of Pennsylvania professor) | Third-party personal report, not a controlled experiment |
Software Engineering
Stripe used it to migrate a 50-million-line Ruby codebase in a single day—work that would have taken a team over two months by hand. On Cognition's FrontierCode benchmark, Fable 5 scored highest among frontier models at medium effort.
Vision
It extracts precise numbers from scientific figures and rebuilds web app source code from screenshots alone. Anthropic demonstrated it beating Pokémon FireRed using only raw game screenshots—no maps, navigation aids, or extra game-state info.
Knowledge Work
Fable 5 scored highest on Hebbia's finance benchmark, with substantial leads in document reasoning, chart interpretation, and problem solving. IMC reported near-perfect scores across their trading analysis evaluations—factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis, all close to perfect.
Memory and Long Context
It stays focused across millions of tokens in long-running tasks and improves outputs using persistent file-based notes. In Slay the Spire, Fable 5 with persistent memory improved three times more than Opus 4.8 and reached the final act three times more often.
Autonomous Task Execution
University of Pennsylvania professor Ethan Mollick tested it independently and said Fable 5 "consistently outperformed basically every other public model I have used by a considerable margin," working autonomously on multi-page specs for hours. He generated multiple playable games from a single prompt in Claude Code—Snake, Strata, and a Rilke-based Duino.
Safety Restrictions
Fable 5 is the first publicly available Mythos-class model with content classifiers. When the system detects any of the following three request types, the launch behaviour was to downgrade the response to Opus 4.8:
- Cybersecurity: vulnerability discovery, exploit development, defense evasion, automated hacking. Anthropic's internal evaluations show the cybersecurity classifier prevents the model from making any progress on these tasks. External red teams also found no universal jailbreaks on long-form agentic tasks.
- Biology and chemistry: including dual-use capabilities like viral vector design for gene therapy. Anthropic found that Mythos-class models outperformed dedicated protein language models on this task.
- Model distillation: large-scale extraction of model capabilities to train competing models.
Over 95% of sessions don't trigger any classifier. When triggered, users are notified and the response comes from the next-strongest model instead—much better than an outright refusal. Anthropic ran over 1,000 hours of external bug bounties with no universal jailbreak found, though the UK AISI made some progress within a short testing window.
Fable 5 also adds a 30-day data retention policy to defend against novel attacks and reduce classifier false positives. Anthropic promises not to use this data for model training or non-safety purposes.
Pricing
API pricing is $10 per million input tokens and $50 per million output tokens—roughly half the price of Mythos Preview.
The subscription rollout ran in three stages, and the June timeline is already behind us:
| Period | Access |
|---|---|
| June 9 – June 22 | Free for Pro/Max/Team/Enterprise users (ended) |
| Starting June 23 | Removed from subscription plans, requires usage credits (in effect) |
| Now (checked 2026-10-20) | Pro still needs usage credits; Max 5x / Max 20x have access at 50% of weekly limits |
Current availability comes from the "Models and usage" table on claude.com/pricing: the Fable row reads No for Free, "Usage credits" for Pro, and "50% of weekly limits" for both Max tiers. Heavy subscribers can already use it directly; Pro users still need to buy credits.
Developers reach it through the Claude API using the model name claude-fable-5-1; the older claude-fable-5 alias is still Active and not retiring before 2027-06-09 (model deprecations).
FAQ
What's the difference between Fable 5 and Mythos 5?
Same underlying model. Fable 5 has safety classifiers (cybersecurity, biology/chemistry, distillation). Mythos 5 lifts the corresponding restrictions (Glasswing partners bypass cybersecurity; biology researchers bypass biology/chemistry). Mythos 5 is currently limited to Project Glasswing partners and trusted biology researchers.
What if the classifier blocks legitimate requests?
The classifiers are currently tuned to be conservative and may block harmless requests. Anthropic says it will keep reducing false positives after launch.
Can it handle coding and game development?
Yes. Stripe used it for a large-scale code migration. Ethan Mollick generated multiple playable games from a single prompt in Claude Code. Fable 5 also tops CursorBench, FrontierCode, and other coding benchmarks.
How does it compare to GPT-5.5?
Anthropic partner Matthew Pines (CEO) independently tested it and said Fable 5 "got nearly to where GPT-5.5 landed after four days in 36 hours using a third of the reasoning tokens." This is a third-party user report, not an official comparison.
What if my work gets blocked by the classifiers?
For everyday coding and knowledge work, switch to Opus 5.5 at $4/$20, which Anthropic positions as Fable 5.1-level on most work. Cybersecurity professionals can apply for the Cyber Verification Program for reduced blocking; biology and chemistry researchers go through the trusted biology access program.
How is Fable 5 data handled?
All Fable 5 and Mythos 5 traffic is retained for 30 days, used only for security and defense purposes. Data won't be used to train new models. All human access is logged, and data is deleted after 30 days.