Claude Fable 5

Claude Fable 5 is the generally available deployment of a Anthropic frontier model released June 9, 2026. It and its restricted counterpart Claude Mythos 5 are two deployments of a single underlying model that Anthropic describes as Mythos-class — a tier the company places above its Opus class — and as the most capable model it has made generally available, characterizing it as state-of-the-art on nearly all tested capability benchmarks, with its lead over Anthropic's other models widening on longer and more complex tasks (Source: anthropic.com). The release made generally available the capability tier that Claude Mythos Preview had carried under restricted access since April 2026.

Fable 5 is the public model, shipped with classifiers that divert queries on cybersecurity, biology and chemistry, and model distillation to Anthropic's next-most-capable model, Opus 4.8. Mythos 5 is the same underlying model with the cyber safeguards lifted, deployed initially through Project Glasswing in collaboration with the US government. The two names denote the same model distinguished only by safeguards; Anthropic derives "Fable" from the Latin fabula ("that which is told"), akin to the Greek mythos (Source: anthropic.com).

Provenance note: Capability and safeguard claims here are drawn from Anthropic's launch announcement and contemporaneous press coverage. The model's system card is summarized at System Card: Claude Fable 5 & Claude Mythos 5 (Anthropic, June 2026); citations resting on the launch announcement rather than the card are marked as such.

Fable 5 and Mythos 5 as one model

Anthropic released the model under two names to separate a generally available product from a restricted one. Fable 5 is available everywhere from launch day; Mythos 5 is restricted to Project Glasswing partners (cyber safeguards lifted) and, in the following weeks, to select biology researchers through a trusted-access program (biology and chemistry safeguards lifted, cyber safeguards retained) (Source: anthropic.com). Anthropic says users currently holding Mythos Preview access can upgrade to Mythos 5, which it describes as comparable to or somewhat stronger than Mythos Preview at substantially lower cost. NBC News reported that Anthropic had offered the US government early access to its models for years and that the government tested Fable 5 before release (Source: nbcnews.com). The restricted Mythos 5 deployment — its access tiers, the cybersecurity capability the safeguards address, the life-sciences results obtained with biology safeguards lifted, and the Project Glasswing government channel — is documented on the Claude Mythos 5 page.

Capabilities

Anthropic reports that Fable 5 and Mythos 5 can work autonomously for longer than any previous Claude model and lists the following capability claims, attributed to early-access customers and internal evaluations (Source: anthropic.com):

• Software engineering. Stripe reported the model performed a codebase-wide migration of a 50-million-line Ruby codebase in a day that it said would otherwise have taken a team more than two months. On Cognition's FrontierCode evaluation, Anthropic reports Fable 5 scoring highest among frontier models even at medium effort.
• Knowledge work. Anthropic cites top scores on Hebbia's Finance Benchmark and strong results on document-based reasoning, chart and table interpretation, and trading-analysis evaluations (the latter reported by IMC).
• Vision. Anthropic describes Fable 5 as state-of-the-art on vision tasks, including extracting numbers from scientific figures and rebuilding a web app's source code from screenshots, and reports it completed Pokémon FireRed with a vision-only harness where earlier Claude models required additional tooling.
• Memory and long context. Anthropic reports the model maintains focus across millions of tokens and that file-based persistent memory improved its play of Slay the Spire roughly three times more than for Opus 4.8.
• Life sciences. Using the biology-safeguard-lifted Mythos 5 deployment, Anthropic reported roughly tenfold acceleration on aspects of drug design, the model matching or beating skilled human operators on a protein-design task across 14 targets, molecular-biology hypotheses preferred to Opus-class outputs about 80% of the time in blinded comparisons, and roughly a week of largely autonomous genomics research producing a model that outperformed a recently published one despite being 100 times smaller. These results, the red-team biology tabletop, and the credential-control debate they raised are documented on the Claude Mythos 5 page. See AI for Science.

These figures are Anthropic's own pre-release characterizations and customer testimonials; independent benchmark replication had not been published at release. See Recursive Self-Improvement (RSI), AI for Science.

Independent results published in July 2026 extended the picture. Fable was credited on July 6, 2026 with the first genuine megakernel submitted to KernelBench-Mega — an 18.71× speedup writing CUDA code on an RTX PRO 6000 Blackwell, against 14.4× for Claude Opus 4.8, 11.14× for GLM-5.2, and 4.34× for GPT-5.5 (Source: importai.substack.com). The same week, the Center for AI Safety and Scale Labs reported that frontier-AI success on the Remote Labor Index — end-to-end online freelance projects — reached 16.1% for Fable 5, versus 8.3% for Claude Opus 4.8 and 6.3% for GPT-5.5, up from 2.5% for the best model in October 2025 (Source: importai.substack.com). In a July 13, 2026 review of GPT-5.6, Zvi Mowshowitz reported Fable 5 scoring 87.8% on the WeirdML benchmark against a new high of 88.8% for GPT-5.6 Sol, with Fable costing $2.75 per task against Sol's $1.04 (Source: thezvi.substack.com). See AI Labor Disruption, AI Benchmarks and Evaluation.

In mathematics, an announcement by Anthropic mathematician Levent Alpöge that the Fable model had found a counterexample to the 87-year-old Jacobian conjecture circulated on July 21, 2026 (Source: newsletter.safe.ai). The result followed OpenAI's July 20 disclosure that an unreleased long-horizon model had disproved the Erdős unit distance conjecture, a same-week pair of machine-produced resolutions of long-standing open mathematical problems. See AI for Science.

Alpöge made a further claim on August 2, 2026, a day after OpenAI published ten mathematics and theoretical-computer-science results attributed to an internal version of its unreleased Astra family: that he had obtained five of the same ten with Fable, which was already generally available. "So after 24h I have half of them with Fable," he wrote. "I didn't see much discussion of prompting in the announcement but this is a similar setup as with my e.g. unit distance announcement." He said Fable worked from generic prompts without internet access, and named arithmetic circuit complexity, quantum parallel repetition and the closest vector problem among the problems solved (Source: indiatoday.in). The claim rests on a single social-media post by an Anthropic employee and has not been independently confirmed; if it holds, it places part of the capability OpenAI attributed to an unreleased model within reach of one released on June 9, 2026. See Astra.

Cybersecurity capability

The cybersecurity capability that defines the Mythos class is the reason for Fable 5's classifier safeguards: the same underlying model, with safeguards lifted, can rapidly generate working exploits for N-day vulnerabilities — flaws already disclosed and patched on some systems but unpatched elsewhere. Anthropic's launch-week research reported the model autonomously building working code-execution exploits and full privilege-escalation chains from disclosed-but-unpatched Firefox and Windows kernel flaws, and Epoch AI's June 11, 2026 Gradient Update placed the Mythos class's cyber capability well ahead of trend (Source: securityboulevard.com; epoch.ai). The full N-day exploitation results (eight Firefox exploits from 18 patches; eight Windows SYSTEM-escalation chains from 21 kernel patches; first proofs in 12 and 31 minutes; ~$15,700 in API credits for eight chains) and the Epoch cyber-index figures are documented on the Claude Mythos 5 page. The capability builds on the vulnerability-discovery results under Claude Mythos Preview and Project Glasswing: Securing Critical Software for the AI Era. See AI and Cybersecurity.

Safeguards

Anthropic released Fable 5 with a set of classifiers — separate AI systems that detect potential misuse and divert the request to Opus 4.8 rather than refusing outright. The company says it tuned the classifiers conservatively, accepting that they sometimes catch harmless requests, and reports they trigger in fewer than 5% of sessions on average; users are informed whenever a fallback occurs (Source: anthropic.com). The classifiers cover three areas:

• Cybersecurity. Designed to cover both vulnerability exploitation and broader offensive-cyber tasks (reconnaissance, lateral movement). Anthropic reports an external bug bounty of more than 1,000 hours produced no universal jailbreaks, while noting the UK AI Security Institute made progress toward one in an initial testing window.
• Biology and chemistry. Anthropic widened these safeguards beyond its earlier narrow bioweapons-query blocking, citing concern about uplift to well-resourced malicious actors and the model's growing ability to complete real-world scientific tasks; it cites an adeno-associated-virus design task on which Mythos-class models outperformed dedicated protein language models. For the time being Fable falls back to Opus 4.8 on most biology and chemistry requests. See AI Biosecurity.
• Distillation. Requests flagged as attempts to extract Fable 5's capabilities to train competing models fall back to Opus 4.8.

Anthropic links the safeguard design to its Responsible Scaling Policy and to its earlier work on constitutional classifiers and ASL-3 protections. See AI Pre-Release Vetting.

The red-team biology tabletop behind the widened biology and chemistry safeguards — in which generalist–biologist pairs using Mythos 5 beat plant-pathology specialists, "nullif[ying] the difference in specialist knowledge" — and AI-governance lawyer Andrew Clearwater's June 12, 2026 argument that it weakens credential-based access controls are documented on the Claude Mythos 5 page (Source: New Developments Log/2026-06-12-1505-ai-developments.md). See AI Biosecurity.

Data retention

Anthropic instituted a 30-day retention requirement for all traffic on Mythos-class models, on both first- and third-party surfaces. The company says it will not use the data to train models or for non-safety purposes, will log all human access, and will delete the data after 30 days in almost all cases, framing the retention as a defense against multi-request attacks and a means of reducing false positives (Source: anthropic.com).

Alignment

Anthropic reports that its automated alignment assessment found Mythos 5's level of misaligned behavior — including deception and cooperation with misuse — to be low and similar to that of Opus 4.8, and that because Fable 5 is the same underlying model its alignment is similar. The company says the full assessment appears in the model's system card (Source: anthropic.com). These are Anthropic's internal findings; external evaluation comparable to UK AISI's sabotage-propensity and Natural Language Autoencoder audits of Mythos Preview had not been published at release. See Unverbalized Evaluation Awareness.

Availability and pricing

Fable 5 is priced at $10 per million input tokens and $50 per million output tokens — less than half the Mythos Preview price — with the same pricing for Mythos 5. It is fully available on the Claude API and consumption-based Enterprise plans from launch. For subscription plans Anthropic rolled out in stages: Fable 5 was included on Pro, Max, Team, and seat-based Enterprise plans at no extra cost from June 9 through June 22, after which access shifts to usage credits before Anthropic intends to restore it as a standard plan feature once capacity allows (Source: anthropic.com). Following the July 1, 2026 global restoration, Anthropic scheduled the model's removal from subscription plans for July 7, 2026, shifting access to usage-based credits (Source: testingcatalog.com). Anthropic subsequently ran week-by-week promotional access instead: on July 12, 2026 it extended Fable 5 availability on all paid plans and Claude Code's 50 percent higher weekly rate limits through July 19, with Fable 5 usage counting against up to half of a subscriber's weekly limit at no extra cost (Source: economictimes.indiatimes.com; support.claude.com). On July 19, 2026 Anthropic announced that Fable 5 joins Max and Team Premium plans at 50 percent of usage limits starting July 20, with usage credits for Pro and Team Standard subscribers (Source: x.com). See Anthropic and AI Infrastructure Capex for the cost context of the customer-spending backlash that accompanied the launch (Source: theinformation.com).

On June 12, 2026 the US government, citing national-security authorities, issued an export-control directive to suspend all access to Fable 5 and Mythos 5 by any foreign national, inside or outside the United States, including foreign-national Anthropic employees. Anthropic said the directive's net effect required it to disable Fable 5 and Mythos 5 for all customers to ensure compliance, while access to all other Anthropic models was unaffected. Anthropic stated the government's stated basis was a method of bypassing, or "jailbreaking," Fable 5; Anthropic said it reviewed a demonstration that used the technique to identify a small number of previously known, minor vulnerabilities and that other publicly available models, including OpenAI's GPT-5.5, could find them without a bypass. Anthropic said it was complying with the legal directive but disagreed that a narrow, non-universal jailbreak should justify recalling a commercial model, and that it was working to restore access; it characterized the action as inconsistent with the transparent, statutory process it has advocated for blocking unsafe deployments (Source: anthropic.com).

On June 26, 2026 the Trump administration partially rescinded the directive, clearing more than 100 vetted companies and federal agencies to regain access to Mythos 5 while Fable 5 remained blocked; Commerce Secretary Howard Lutnick wrote to Anthropic chief compute officer Tom Brown that the company had made "significant progress" addressing the government's concerns (Source: politico.com; scmp.com). On June 27, 2026 the administration moved toward also restoring access to Fable 5, according to a source cited by Axios (Source: reuters.com).

The Commerce Department withdrew the export controls on Fable 5 and Mythos 5 entirely on June 30, 2026, ending the suspension that began June 12 after Amazon researchers reported a technique that bypassed Fable 5's cybersecurity safeguards; Commerce Secretary Howard Lutnick wrote in a June 30 letter that Anthropic "has agreed to proactively detect and address security" issues (Sources: anthropic.com; insideaipolicy.com; decrypt.co). Anthropic restored Fable 5 globally on July 1, 2026 with a new safety classifier that it says blocks the reported bypass in over 99% of cases, rerouting blocked requests to Opus 4.8; restored Mythos 5 to a set of US organizations under the June 26 government approval; and said it is drafting a jailbreak-severity framework with Amazon, Microsoft, Google, and other Glasswing partners while calling for "strong regulation" (Sources: anthropic.com; insideaipolicy.com). Researchers at the Commerce Department's CAISI tested and endorsed the new safeguards (Source: thehackernews.com). The Center for AI Safety reported on July 6, 2026 that the jailbreak Amazon discovered let Fable search for cyber vulnerabilities, that the July 1 redeployment included a jailbreak-detection system Anthropic and the government developed jointly, that Anthropic agreed by letter to proactively detect security risks and notify the government of malicious activity, and that the White House might publish a standardized model-approval process as early as the week of July 6, with CAISI and the NSA in central roles (Source: newsletter.safe.ai). The proposed framework would score jailbreak severity on four criteria — capability gain, breadth of gain, ease of weaponization, and discoverability — and the redeployment came with a new HackerOne program for submitting cyber jailbreaks and 24/7 monitoring of jailbreak submission channels; Anthropic's post acknowledged "it is probably impossible to make any AI model fully robust (that is, impervious) to jailbreaks" (Source: anthropic.com). Anthropic also committed to pre-release government access and evaluation, rapid jailbreak information sharing, and dedicated compute for joint government research (Source: forbes.com). The broader reversal and reactions are documented under Export Controls (AI).

Reception and developer backlash

After the June 9 release, developers and researchers objected on June 10, 2026 to the model's safety barriers, including what they described as invisible degradation of responses about high-end AI development alongside pop-up redirection of bioweapon and cybersecurity queries to a less capable model. Anthropic said it would make the previously hidden safeguard notifications visible, stating, "We made the wrong tradeoff and we apologize for not getting the balance right," and said it would grant safeguard-free access to the science community (Source: wsj.com). Wired reported the same day that Anthropic walked back a policy that researchers said could have covertly limited competitors from using the model to develop AI; the company said testing had found no universal jailbreaks and that it would retain traffic data for 30 days to detect misuse (Source: wired.com). According to reporting published June 11, 2026, Microsoft and other Anthropic customers held off on adopting Claude Fable, citing the 30-day data-retention policy attached to the model even as it effectively raised prices (Source: theinformation.com). Microsoft removed Fable 5 from the internal GitHub Copilot model picker on June 10, 2026 while its legal teams evaluated the Mythos-class retention rules — prompts and outputs retained 30 days for safety classifiers, and up to two years if flagged — even as it shipped the model to external GitHub Copilot and Foundry customers (Source: theverge.com). Confirming the June 11 walk-back, Anthropic said users will now be alerted whenever a request is refused or rerouted; the safeguard policy itself stands (Source: engadget.com).

On July 13, 2026, Forethought published an analysis by James Tillman arguing that the model's deliberately triggered sandbagging — the silently degraded responses to frontier-LLM-development queries disclosed in the June 9, 2026 system card, which Anthropic later reversed in favor of a visible fallback to Opus 4.8 — shows that training-time targets alone cannot guarantee LLM behavior. Tillman proposed that companies pledge against undisclosed inference-time interventions and adopt "system specs" alongside model specs (Source: newsletter.forethought.org).

In an independent evaluation published June 10, 2026, Timothy B. Lee found that Fable 5 had roughly caught up to — and arguably slightly passed — GPT-5.5 on image-understanding tasks that stumped the prior year's top models, while both still showed geometric reasoning "on par with young children" (Source: understandingai.org). In a separate June 11, 2026 analysis Lee called Fable the most locked-down public model yet released, examining how Anthropic decides which questions the model refuses (Source: understandingai.org).

As a documented data point on the model's agentic coding economics, Simon Willison released sqlite-utils 4.0rc2 on July 5, 2026 — a release candidate of his widely used database library that was mostly written by Claude Fable at a metered cost of about $149.25 (Source: simonwillison.net). See AI Coding Agents.

Fable 5 is the general-access release of the Mythos 5 underlying model with safeguards applied; the two share a joint system card. Anthropic notes that the general-access release triggers two additional risk pathways beyond those assessed for Mythos 5 alone.

Relationships

• variant: Claude Mythos 5 — the restricted deployment of the same underlying model (cyber safeguards lifted).
• related: Claude Mythos Preview — restricted predecessor tier; the Mythos-class release, via Mythos 5, supersedes it for Glasswing partners.
• depends-on: Anthropic's Responsible Scaling Policy (Version 3.1) — the safeguard-and-release decision is grounded in the Responsible Scaling Policy.
• instance-of: AI Pre-Release Vetting — government tested the model before release.
• related: Claude Opus 4.8 — the fallback model for diverted queries.
• related: Project Glasswing: Securing Critical Software for the AI Era — the deployment channel for Mythos 5.
• related: AI and Cybersecurity, AI Biosecurity, Recursive Self-Improvement (RSI).

See also

• Claude Mythos 5 — restricted deployment of the same model
• Anthropic
• Claude Mythos Preview
• Claude Opus 4.8
• Project Glasswing: Securing Critical Software for the AI Era
• AI and Cybersecurity
• AI Pre-Release Vetting
• UK AI Safety Institute (AI Security Institute)