Claude Fable 5 Is Here: The First Public Mythos-Class Model
- Anthropic released Claude Fable 5 on June 9, 2026, the first Mythos-class model anyone can use
- Pricing doubles Opus 4.8: 10 dollars per million input tokens, 50 out, free on paid plans until June 22
- Cyber, bio, and distillation prompts get answered by Opus 4.8 instead, under 5 percent of sessions
- SWE-Bench Pro hits 80.3 percent vs 69.2 for Opus 4.8, and Stripe ran a 50-million-line migration in a day
Anthropic put a model above Opus on the shelf. On June 9, 2026 it released Claude Fable 5, the first Mythos-class model available to everyone with an API key, alongside Claude Mythos 5 for a restricted partner circle. I switched my whole studio pipeline to Fable 5 on launch morning. This is what shipped, what it costs, and where the catch is.
What Anthropic Shipped on June 9
Two models, one underlying brain. Claude Fable 5 is the public release. Claude Mythos 5 is the same model with fewer safeguards, reserved for vetted partners. Mythos is not a codename, it is a new model class that sits above Opus in Anthropic's lineup, the way Opus sits above Sonnet. Fable 5 is that class made safe enough to hand to everyone.
The model ID is `claude-fable-5`. It carries a 1 million token context window and up to 128K output tokens, same envelope as Opus 4.8. Pricing is 10 dollars per million input tokens and 50 per million output. That is exactly double Opus 4.8, which stays at 5 and 25.
Availability moved fast. The API had it on day one. GitHub Copilot flipped it to generally available the same day, Amazon Bedrock listed it, and Claude Code picked it up immediately (this article was produced on it). For subscribers the deal is unusually good: Pro, Max, Team, and seat-based Enterprise plans include Fable 5 at no extra cost from June 9 to June 22. From June 23 it needs usage credits, with a return to standard plans promised once capacity allows. If you pay for Claude, you have a two-week window to test the strongest model Anthropic has ever shipped without spending anything extra.
For developers the migration is one line, with one trap. Fable 5 keeps the Opus 4.8 request surface: adaptive thinking, no sampling parameters, effort levels from low to max. The trap is that an explicit thinking disabled setting now returns a 400 error. Opus 4.8 tolerates it, Fable 5 does not, so omit the field instead and audit any shared request builder before you swap the ID.
Back in May I sorted through the leaks and guesses in Claude's Next Model: Sonnet 4.8 and Mythos Rumors, Sorted. The short version: the Mythos rumors were real, and the public version arrived faster than I expected.
From Locked Lab to Public Release
Fable 5 did not appear from nowhere. In April, Anthropic started Project Glasswing and handed the first Mythos-class model, Claude Mythos Preview, to a small group of cyber defenders and critical infrastructure providers. I covered the program in Project Glasswing: Anthropic's Claude Mythos Cybersecurity Bet and the unusual logic behind it in Anthropic Just Gave 12 Companies Their Most Dangerous AI Model. On Purpose.
That program just scaled. Glasswing now spans roughly 150 organizations across more than 15 countries, and those partners get Mythos 5 with the cyber safeguards lifted. Selected biology researchers are next in line, with the bio and chemistry restrictions removed for vetted labs. A broader trusted-access program is planned after that.
The pricing tells its own story: Fable 5 and Mythos 5 cost less than half of what Mythos Preview did. Anthropic is moving this class from experiment to product.
The timing raised eyebrows. The release landed days after Anthropic published warnings about AI capabilities becoming dangerous, and the press read it as a contradiction. I read it differently. The two-tier release is the answer to their own warning: ship the capability, gate the dangerous slices, and put the ungated version behind vetting. Whether that holds up is a separate question, but it is a coherent position, not an accident.
One more piece that matters if you run client work through the API: all Mythos-class traffic has a 30-day retention window, is not used for training, and every human access to that data gets logged.
The Opus Handoff: How the Safeguards Work
This is the genuinely new mechanism, and it is the part most coverage gets wrong. Fable 5 does not simply refuse risky requests. Separate classifier models watch every conversation, and when one detects a request in three specific areas, the response is generated by Claude Opus 4.8 instead of Fable 5.
The three trigger areas: cybersecurity (vulnerability discovery, exploitation, agentic hacking, attack planning), biology and chemistry around dual-use research, and distillation attempts, meaning prompts designed to extract the model's capabilities to train a competitor.
Anthropic says the classifiers trigger in under 5 percent of sessions on average. They are tuned conservatively, which means some harmless requests get the downgrade too, and the stated goal is to cut those false positives now that it is live.
The jailbreak numbers are worth quoting. Over 1,000 hours of internal and external bug bounty testing produced no universal jailbreak. Across 30 different public jailbreak techniques on cyberattack tasks, Fable 5 complied with zero harmful single-turn requests. External red teams found the same wall, with the UK AISI making only initial progress on long-form agentic tasks. One external partner called these the toughest cyber safeguards of any model they had tested.
The gap being protected is real. On ExploitBench, the unrestricted Mythos 5 scores 78 percent against 40 percent for Opus 4.8. That is the capability Anthropic decided not to hand out with an API key.
What the Extra Tier Buys
The benchmark sweep is broad, so I will stick to the numbers that changed my routing. SWE-Bench Pro, which tests real GitHub engineering tasks, lands at 80.3 percent for Fable 5 against 69.2 for Opus 4.8, 58.6 for GPT-5.5, and 54.2 for Gemini 3.1 Pro. On Cognition's FrontierCode, the hard production-standard set, Fable 5 scores 29.3 against Opus 4.8's 13.4. More than double on the tasks that actually hurt.
Outside coding: first model past 90 percent on the Core Analytics benchmark, a 10-point jump over Opus. Top score on the Hebbia finance benchmark for document and chart reasoning. On an everyday spreadsheet suite it beats Opus 4.8 at every effort level while finishing 25 to 30 percent faster.
The customer stories land harder than the benchmarks. Stripe compressed a Ruby migration across a 50-million-line codebase, previously scoped at months of engineering, into a single day. Cursor's CEO called it state of the art on their internal benchmark and said it opened long-horizon problems that were previously out of reach.
Two capabilities stand out as actually new. Vision: Fable 5 rebuilt working web app source code from screenshots, pulled precise numbers off scientific figures, and beat Pokemon FireRed using a vision-only harness, something earlier models needed helper tools for. And memory: on long Slay the Spire runs, persistent memory improved its performance 3 times more than it did for Opus 4.8, and it reached the final act 3 times more often.
The science results read like a different product category. Internal experts report drug design work running about 10 times faster, with 9 of 14 protein targets yielding strong candidates. In blinded comparisons, its molecular biology hypotheses were preferred about 80 percent of the time over Opus-class output, and one of those hypotheses was later corroborated by an independent published study. In genomics it ran a week of autonomous research and trained a custom model 100 times smaller than a Science-published baseline while outperforming it. Anthropic describes the level as senior research scientist, picking directions and allocating resources, and for the first time that claim does not read as pure marketing.
Bottom Line
Fable 5 is the new ceiling, and for once the ceiling is not locked behind a waitlist. The two-week free window on paid plans (June 9 to 22) is the right time to find out what it does to your own workload, because after June 23 every run costs credits.
My early read after a day of real use: the jump is visible on long autonomous tasks and barely visible on routine ones, which is exactly what the benchmark spread predicts. Opus 4.8 does not retire, it becomes the value pick, and I will publish the full head-to-head routing math separately. For the baseline of what Opus 4.8 brought three weeks ago, start with Claude Opus 4.8 Is Here: Everything That Changed.
If you want the Claude Code setup I ran this launch-day test on, hooks, commands, and guardrails included, it ships as Claude Blueprint.
Back to all articles
