Claude Fable 5 & Mythos 5: Benchmarks, Pricing, Access

Claude Fable 5 is the first Mythos-class model Anthropic has made generally available, released June 9, 2026. It is the most powerful model the company has ever shipped to the public, and it posts state-of-the-art results across software engineering, knowledge work, scientific research, and vision. It is also the most expensive flagship Anthropic has priced for general use at $10 per million input tokens and $50 per million output, double Opus 4.8. The catch most readers are searching for: on Pro, Max, Team, and seat-based Enterprise plans, Fable 5 is included at no extra cost only through June 22, 2026. On June 23 it moves behind usage credits.

Fable 5 is Mythos made safe. A separate classifier system watches every session for high-risk cybersecurity, biology, chemistry, and distillation requests and routes those to Opus 4.8 instead, telling the user when it does. That fallback triggers in under 5% of sessions on average. Its sibling, Claude Mythos 5, is the same underlying model with those safeguards lifted, and it stays restricted to Project Glasswing partners. The API model ID is claude-fable-5, available today on the Claude API, Claude Code, consumption-based Enterprise plans, AWS, Google Cloud, and Microsoft Foundry.

Key Specs

What Fable 5 Is: Mythos, Made Safe

In April, Anthropic announced Claude Mythos Preview and refused to release it. The stated reason was offensive cybersecurity capability that ran ahead of the company's safeguard stack. Mythos found thousands of zero-day vulnerabilities autonomously, produced 181 working Firefox exploits on a harness where Opus 4.6 produced two, and stayed locked behind Project Glasswing at $25/$125 per million tokens. For two months, Opus 4.8 was the public ceiling.

Fable 5 changes that by splitting one model into two products. The two share the same weights. What separates them is a classifier layer. As Anthropic puts it in a footnote, "Fable is from the Latin fabula, 'that which is told,' akin to the Greek mythos. The safeguards are what distinguish the two models."

The safeguard mechanism is worth understanding precisely, because it changes how the model behaves in practice. Fable 5 does not refuse flagged requests. Instead, a separate set of AI classifiers watches for potential misuse and jailbreak attempts across three categories. When one triggers, the request is handed off to Claude Opus 4.8, and the user is told the handoff happened. Anthropic reports this fallback occurs in fewer than 5% of sessions on average.

That design has a real cost most launch coverage glosses over. The classifiers are deliberately conservative, and Anthropic admits the biology and chemistry coverage is "narrower than ideal" because the company prioritized shipping speed over precision. A conservative classifier produces false positives. If you do legitimate security research, work in a wet lab, or run agentic recon against your own infrastructure, expect Fable 5 to occasionally drop you to Opus 4.8 mid-task and tell you it did. For most developers that 5% is invisible. For a minority of technical users it is a recurring tax on the exact workloads that justify paying double.

Benchmark Results

Anthropic frames Fable 5 as a capability release, not a reliability release. Where Opus 4.8 was about doing the same work with fewer mistakes, Fable 5 is about doing work that earlier models could not finish at all. The pattern across partner evals is consistent: the lead over Opus 4.8 widens the longer and more complex the task gets.

A few results deserve weight beyond a single row. On Sigma's core analytics benchmark, Izzy Miller reported Fable 5 was "the first to break 90% on our core analytics benchmark," which he framed as "a 10-point jump over Opus." Ten points at the top of a saturating benchmark is unusual. On frontier physics, Convergence CEO Matthew Pines called it "the strongest model we've tested on frontier physics research while using a third of the reasoning tokens," which is the more interesting half of the claim: better answers for less compute, not better answers at any cost.

The coding partners converge on the same theme. Cursor's Michael Truell said Fable 5 is state of the art on CursorBench and that "it's opened up a class of long-horizon problems that were out of reach for earlier models." GitHub CPO Mario Rodriguez described "complex, long-horizon coding tasks with a level of autonomy and reliability" past prior benchmarks. Notion's Matt Colyer called the results "the strongest results of any Claude model we've had the opportunity to test." Lovable CTO Fabian Hedin offered the most concrete line: "Apps that took a hundred prompts a year ago, it now one-shots."

One caveat, the same one that applies to every frontier launch. Anthropic configures its own benchmark harness, typically at maximum effort and averaged across trials, while competitor numbers come from their own setups. The direction of these results is real and the partner quotes are independent, but the exact margins are not an apples-to-apples scoreboard. Benchmark on your own workload before you commit a production pipeline to the higher rate.

Pricing and the June 22 Usage-Credits Change

This is the section most readers came for, so here is the blunt version.

On the API, Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is exactly double Opus 4.8's standard $5/$25, and it matches Opus 4.8's Fast mode rate. It is also less than half the price of Mythos Preview's $25/$125. Prompt caching cuts cached input by up to 90%, and US-only inference is available at a 1.1x multiplier for teams with data-residency requirements. For developers on the API, the math is simple: you pay per token, and Fable 5 is a premium tier you opt into for the tasks that warrant it.

The subscription story is where it gets contentious. From June 9 through June 22, 2026, Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans. Within those plans it weighs roughly double the usage of Opus, so a Fable 5 session burns through your included limits about twice as fast as the same Opus 4.8 session.

On June 23, Fable 5 is removed from those subscription plans. To keep using it after that, you switch to usage credits, the consumption-based mechanism that kicks in once your plan's included limits are exhausted and bills additional usage at standard API rates. Anthropic frames this as a capacity constraint, not a permanent pricing decision, and says it aims to restore Fable 5 as a standard subscription feature once capacity allows. The free inclusion window may also be extended if capacity holds.

Read plainly, the sequence is: free for thirteen days, then pay per use, with a promise to bring it back into the subscription later. If you are a Pro or Max subscriber, the practical advice is to test Fable 5 hard before June 23, decide whether the capability lift justifies metered spend on your specific work, and keep Opus 4.8 as your default for everything that does not need the extra ceiling. The economics only favor Fable 5 on long-horizon, high-complexity tasks where finishing the job at all is the win. For the full breakdown of how credits are purchased, how auto-reload works, and how to cap spend, see our Fable 5 usage credits guide.

Safety Profile

Fable 5's safety story is the classifier system plus a new data policy, and both are sharper than anything Anthropic has shipped to general customers before.

The three classifier categories are cybersecurity (exploitation, offensive cyber tasks, agentic hacking such as reconnaissance, discovery, and lateral movement), biology and chemistry (broad dual-use research coverage, which Anthropic openly calls narrower than ideal), and distillation (large-scale attempts to extract Claude's capabilities to train competing models). When any classifier fires, the request is served by Opus 4.8 and the user is notified. Anthropic reports the fallback rate stays under 5% of sessions on average.

The hardening numbers are meaningful. An external bug bounty found no universal jailbreaks in over 1,000 hours of testing. One external partner found Fable 5's cyber safeguards the most robust of any model they tested, with zero compliance on harmful single-turn requests across 30 public jailbreak techniques. With safeguards in blocking mode, Fable 5 made no measurable progress on offensive cyber evaluations.

The new data retention policy is the part with industry implications. Anthropic now mandates 30-day retention for all traffic on Mythos-class models, first-party and third-party, including on enterprise agreements that were previously zero-retention. The data is explicitly not used to train new Claude models or for any non-safety purpose, and Anthropic logs all human access and deletes after 30 days in nearly all cases. The stated purpose is defending against novel attacks and reducing classifier false positives. If your organization signed a zero-retention deal, this is a change you need to take to legal before you route production traffic to Fable 5, because it overrides that term for this model class.

On alignment, the assessment is reassuring in a way the cyber capability is not. Anthropic's automated alignment evaluation rated Mythos 5's level of misaligned behavior, things like deception or cooperation with misuse, as low and similar to Opus 4.8. The risk that gates Mythos 5 is raw offensive capability, not the model's disposition.

Mythos 5 and Project Glasswing

Mythos 5 is the uncapped version of Fable 5. Same weights, safeguards lifted, and Anthropic describes it as having "the strongest cybersecurity capabilities of any model in the world." That is precisely why it is not on a public endpoint.

Access runs through Project Glasswing, the invitation-only defensive cybersecurity program Anthropic launched in April for critical-infrastructure operators, run in collaboration with the US government. Every current Glasswing and Mythos Preview partner can upgrade to Mythos 5 today with cyber safeguards lifted. Anthropic plans to expand access steadily and stand up a formal trusted access program for cybersecurity organizations.

Biology is a separate track. A biology trusted access program opens in the coming weeks, offering Fable 5 with bio and chemistry safeguards removed (cyber safeguards stay on) to a small group of vetted life sciences researchers. The case for it is in the research results: with protein-design tooling and no human assistance, Mythos-class models matched or beat skilled human operators, and on 9 of 14 protein targets produced strong drug-design candidates now under investigation. In a week of largely autonomous genomics work, the model trained a custom ML model that outperformed a recent Science-published model while being 100x smaller.

The structure mirrors the original Glasswing logic. The most capable version goes first to the people defending against, or researching with, that capability, while the general public gets the safeguarded version. If you are an individual developer or a general enterprise, Mythos 5 is not an access path. Fable 5 is your ceiling, and for almost everyone that is the correct ceiling.

How to Use Fable 5 in Claude Code

Switch your default model:

claude config set model claude-fable-5

For a per-session override without changing your default:

claude --model claude-fable-5

The model API identifier is claude-fable-5, available across the Claude API, Claude Code, consumption-based Enterprise plans, AWS, Google Cloud, and Microsoft Foundry on launch day. If you run Claude Code on a Pro, Max, or Team subscription, remember the June 23 cutover: after that date, Fable 5 sessions draw on usage credits rather than your included plan usage, and they weigh roughly double an Opus session against any limits while still included.

Most of the agentic coding playbook carries forward unchanged. If you already drive Opus 4.8 with explicit instructions, planning passes, and tight tool definitions, the same patterns apply to Fable 5. The difference is task selection, not technique. If you want model routing configured out of the box across the full Claude lineup, ClaudeFast's Code Kit ships with routing setups so heavy tasks reach the right model and routine edits stay on cheaper tiers, which matters more than ever when the top tier bills per token.

Fable 5 vs Opus 4.8

The honest recommendation: Opus 4.8 stays the default for daily agentic coding and the vast majority of work. Reach for Fable 5 when the task is genuinely long-horizon, the kind earlier models could not complete end-to-end, and the result is worth paying double for. For tactical switching across the lineup, see the model selection guide, and for where this lands in Anthropic's release history, the complete Claude model timeline.

Frequently Asked Questions

Is Claude Fable 5 free? On Pro, Max, Team, and seat-based Enterprise plans it is included at no extra cost from June 9 through June 22, 2026 only. On June 23 it is removed and continued use requires usage credits. On the API it is paid from day one at $10/$50 per million tokens. The claude.ai free tier does not include Fable 5.

What happens on June 22? June 22 is the last day Fable 5 is included in subscription plans. Starting June 23 it moves behind usage credits, which bill additional usage at standard API rates once your plan limits are exhausted. Anthropic says it aims to restore Fable 5 as a standard subscription feature once capacity allows, and may extend the free window if capacity holds.

Fable 5 vs Mythos 5: what's the difference? They are the same underlying model. Fable 5 ships with classifier safeguards that route high-risk cybersecurity, biology, chemistry, and distillation requests to Opus 4.8. Mythos 5 has those safeguards lifted and is restricted to Project Glasswing partners and select biology researchers. The public gets Fable 5.

Fable 5 vs Opus 4.8: which should I use? Use Opus 4.8 as your default for daily agentic and coding work at $5/$25. Reach for Fable 5 on long-horizon, high-complexity tasks where its widening capability lead justifies double the price. Its advantage grows the longer and harder the task.

How much does Fable 5 cost? $10 per million input tokens and $50 per million output tokens, double Opus 4.8 and less than half of Mythos Preview's $25/$125. Prompt caching saves up to 90% on cached input, and US-only inference is available at a 1.1x multiplier.

What is the Fable 5 context window? Anthropic did not disclose a context window in the launch announcement. We will update this once the developer documentation specifies it. Do not assume parity with Opus 4.8's 1M tokens until it is confirmed.