Anthropic ships Claude Fable 5 and Mythos 5, splitting its strongest model into a public tier and a Glasswing-restricted cyber tier
Fable 5 is the first publicly available Mythos-class model. It scores 80.3% on SWE-Bench Pro against 69.2 for Claude Opus 4.8 and 58.6 for GPT-5.5, becomes the first model to clear 90% on Hex's analytical benchmark, and shipped with Stripe reporting that it finished a 50-million-line Ruby migration in a single day that would otherwise have taken a team more than two months. Fable 5 is free on Pro, Max, Team and seat-based Enterprise plans through June 22 and bills at $10 per million input / $50 per million output afterward, double Opus 4.8. The same model, with cybersecurity safeguards lifted, ships as Claude Mythos 5 to a small set of cyber defenders and critical-infrastructure operators under Project Glasswing in collaboration with the U.S. government. Three safety classifiers route cybersecurity, biology/chemistry and distillation queries to Opus 4.8 instead of refusing; Anthropic says the classifiers fire on fewer than 5% of Fable sessions. Mandatory 30-day data retention applies to all Fable 5 and Mythos 5 traffic, overriding prior zero-retention enterprise contracts. Fable 5 lands the same day inside GitHub Copilot, Amazon Bedrock, the Claude Platform on AWS, Microsoft Foundry and Harvey.
Anthropic · 2026-06-09 · Source ↗ ·
Anthropic ships Claude Fable 5 and Mythos 5, splitting its strongest model into a public tier and a Glasswing-restricted cyber tier
Fable 5 is the first publicly available Mythos-class model. It scores 80.3% on SWE-Bench Pro against 69.2 for Claude Opus 4.8 and 58.6 for GPT-5.5, becomes the first model to clear 90% on Hex's analytical benchmark, and shipped with Stripe reporting that it finished a 50-million-line Ruby migration in a single day that would otherwise have taken a team more than two months. Fable 5 is free on Pro, Max, Team and seat-based Enterprise plans through June 22 and bills at $10 per million input / $50 per million output afterward, double Opus 4.8. The same model, with cybersecurity safeguards lifted, ships as Claude Mythos 5 to a small set of cyber defenders and critical-infrastructure operators under Project Glasswing in collaboration with the U.S. government. Three safety classifiers route cybersecurity, biology/chemistry and distillation queries to Opus 4.8 instead of refusing; Anthropic says the classifiers fire on fewer than 5% of Fable sessions. Mandatory 30-day data retention applies to all Fable 5 and Mythos 5 traffic, overriding prior zero-retention enterprise contracts. Fable 5 lands the same day inside GitHub Copilot, Amazon Bedrock, the Claude Platform on AWS, Microsoft Foundry and Harvey.
Anthropic · 2026-06-09 · Source ↗ ·
Anthropic released Claude Fable 5 and Claude Mythos 5 Tuesday morning, ending eight months of speculation about whether the company would ever ship Mythos to anyone outside its private partner program. Fable 5 is the public-facing model, available immediately through the Claude API, Pro, Max, Team and seat-based Enterprise subscriptions, GitHub Copilot, Amazon Bedrock, the Claude Platform on AWS, Microsoft Foundry, Harvey and a handful of other launch partners. Mythos 5 is the same model with cybersecurity safeguards stripped out, deployed only to vetted cyber defenders and critical-infrastructure operators through Project Glasswing in collaboration with the U.S. government.
The benchmark numbers are the easy part of the story. On SWE-Bench Pro, the closely watched programming benchmark, Fable 5 solves 80.3% of problems against 69.2% for Claude Opus 4.8 and 58.6% for GPT-5.5. It is the first model to exceed 90% on Hex's analytical benchmark, made up of long, complex analytical tasks, and reaches state of the art on Cognition's FrontierCode and Hebbia's Finance Benchmark for senior-level reasoning. On Harvey's Legal Agent Bench it scores 13.3% against Opus 4.8's 10.4%, and on Harvey's proprietary BigLaw Bench it hits 93.4% overall, a new ceiling for the Anthropic family. Fable 5 completed Pokémon FireRed using vision only without helper tools, posted three-fold gains over Opus 4.8 in Slay the Spire long-context play through memory improvements, and Anthropic says Mythos 5 has already accelerated protein design by roughly 10x in pilot drug-discovery work, producing strong candidates for 9 of 14 protein targets.
The Stripe migration is the demo
The customer headline is Stripe. Anthropic disclosed that Stripe ran Fable 5 across a 50-million-line Ruby codebase to execute a codebase-wide migration. The model finished the work in a single day. A team of human engineers had estimated more than two months for the same effort. The result has become the launch's signature demonstration of what Anthropic is now selling: not better completion of individual coding tickets, but autonomous completion of larger units of work — codebase migrations, multi-step refactors, app prototyping, end-to-end test generation, debugging across unfamiliar tools and multi-day internal projects.
"Claude Fable 5 compressed months of engineering into days." — Stripe, in Anthropic's launch materials
Cursor CEO Michael Truell shipped a launch blurb the same morning: "Claude Fable 5 is the state of the art model on CursorBench. It's opened up a class of long-horizon problems that were out of reach for earlier models." GitHub's CPO Mario Rodriguez framed it for the developer audience: "What excites us most is the direction it points: a future where developers can hand increasingly ambitious work to agents and trust the results." Rakuten, which had publicly tested Mythos Preview in March, said the extra thinking time during long agent runs "pays for itself" when sustained autonomy is the goal.
Three classifiers, one fallback model
The safety architecture is the part of the launch that Anthropic spent most of the briefing on. Fable 5 ships with three safety classifiers: one for cybersecurity, one for biology and chemistry, and one for distillation attempts targeting unauthorized model replication. When a classifier fires, the request does not get refused outright. Instead it falls back to Claude Opus 4.8, which still answers but with the older model's narrower autonomy. Anthropic says the classifiers trigger on fewer than 5% of Fable sessions on average, with more than 95% of requests running entirely on Fable's own responses. An external bug-bounty program ran more than 1,000 hours of red-team testing before launch and produced "zero harmful single-turn requests" that succeeded across jailbreak techniques, according to the launch post.
The single biggest contractual change is data retention. Mythos-class models, including the public Fable 5, require 30-day data retention on all traffic, including prompts and outputs. The retained data is used to operate safety classifiers and detect jailbreak attempts in flight. It is not used to train future models or for any non-safety purpose. The policy overrides prior zero-retention agreements with enterprise customers — GitHub's changelog spells this out explicitly: "Anthropic retains prompts and outputs for up to 30 days to operate safety classifiers that detect harmful or abusive use," while every other Claude model in Copilot, including Opus 4.8, Sonnet 4.5 and Haiku 4.5, continues to operate under Zero Data Retention. Enterprise and Business administrators on GitHub must explicitly enable the Fable 5 policy; it is disabled by default.
Mythos 5 and Project Glasswing
Mythos 5 is the same model, same weights, with the cybersecurity classifier lifted. Anthropic calls it the strongest cybersecurity model in the world. Initial deployment is limited to vetted partners through Project Glasswing — the cyber-defender access program that, as of the prior week, spans roughly 200 critical-infrastructure organizations across more than 15 countries. Cisco, BeyondTrust and a number of US-government partners are among the publicly named participants. Anthropic plans to extend Mythos access to a separate biology research program with similar guardrails on who gets in.
The Project Glasswing track has had a complicated week. Bruce Schneier published a public critique Monday that of roughly 23,000 vulnerabilities Mythos flagged across 1,000 open-source projects during the Preview phase, almost none have been patched, and described as "just not true" the Anthropic framing that Mythos's results are uniquely strong relative to other frontier models. Cisco's Foundry Security Spec team, which scanned 1.8 billion lines of Cisco code with six frontier models, came to a broadly model-agnostic result on raw discovery: any of the top US-lab frontier models surfaces a similar vulnerability surface. Tuesday's Mythos 5 launch is Anthropic's argument that the uniqueness lives downstream in autonomous remediation, not in initial discovery.
Pricing, distribution and the IPO clock
Pricing for both Fable 5 and Mythos 5 is identical: $10 per million input tokens and $50 per million output tokens. That is double Opus 4.8 on output and roughly half what Mythos Preview cost on the private program. Fable 5 is included at no additional cost on Pro, Max, Team and seat-based Enterprise plans from June 9 through June 22; after June 23 it bills against usage credits on those plans. The launch lands inside Anthropic's IPO clock — the company filed a confidential S-1 with the SEC on June 1 at a reported $965 billion valuation after raising $65 billion in its Series H. Fable 5 is the largest single product Anthropic has shipped between filing and pricing, and it is the artifact most analysts will use to value the company against OpenAI heading into the second half.