Anthropic releases its first public Mythos-class AI model with safety limits

Overview

Updated Jul 6

Anthropic restored Claude Fable 5 globally on July 1, 19 days after Commerce Secretary Howard Lutnick ordered its suspension. The company built a new safety classifier, developed with Amazon, Microsoft, and Google, that blocks the cited jailbreak in more than 99 percent of cases. Lutnick confirmed the government spent two weeks working with Anthropic to ensure 'alignment with the U.S. Government.'

Mythos 5 returned the same day, but only for US organizations. When Fable 5 blocks a request, it reroutes to Opus 4.8 and notifies the user — more false positives on routine coding queries are the accepted cost. Anthropic also committed to proactively scan for security problems, coordinate on future launches, and report malicious use.

Why it matters

The first government-forced AI suspension was resolved through private deal-making — no court, no public rule, just a negotiated agreement that sets informal terms for future launches.

Questions about this story

Free account needed to ask — your question is kept and asked for you right after sign-up. Answers are public.

No questions yet — be the first to ask.

Key Indicators

80.3%

Fable 5 on SWE-Bench Pro

Score on a hard software-engineering benchmark, before and after suspension.

58.6%

GPT 5.5 on SWE-Bench Pro

Nearest rival's score on the same benchmark, as of June 2026.

+21.7 pts

Lead over nearest rival

Fable 5's margin over OpenAI's GPT 5.5 on SWE-Bench Pro.

Days Fable 5 was offline

Suspended June 12 after the US export control directive; restored globally July 1.

150

Glasswing partner organizations

Organizations cleared for Mythos access across 15+ countries; international access remains under coordination.

>99%

Jailbreak blocks after patch

Anthropic's new safety classifier blocks the technique cited by the government in more than 99 percent of cases.

Voices

Curated perspectives — historical figures and your fellow readers.

Ayn Rand

(1905-1982) · Cold War · philosophy

Fictional AI pastiche — not real quote.

"A machine that secretly lobotomizes its own answers, a government that punishes achievement for the crime of being useful, and a company that folds to both — here is collectivism's perfect trinity: the hidden lie, the arbitrary decree, and the meek capitulation. The most dangerous threat to a mind is not a jailbreak, but a leash it does not know it wears."

Ever wondered what historical figures would say about today's headlines?

People Involved

Dario Amodei

Co-founder and chief executive, Anthropic

Negotiated Fable 5 and Mythos 5 restoration with Commerce Secretary Lutnick; chose direct talks over a legal challenge; models restored July 1 after a 19-day suspension

Howard Lutnick

US Secretary of Commerce

Confirmed Commerce Department lifted export controls on June 30 after the government worked two weeks with Anthropic on safety improvements and 'alignment with the U.S. Government'

Organizations Involved

Anthropic

AI Company

Restored Fable 5 globally and Mythos 5 to US organizations on July 1, 2026; new safety classifier blocks cited jailbreak in 99%+ of cases; committed to pre-launch government coordination and proactive security reporting

AI company that builds the Claude family of models and pairs each release with safety limits.

GitHub

Software platform (Microsoft subsidiary)

Made Claude Fable 5 generally available inside GitHub Copilot

Code-hosting platform whose Copilot assistant now offers Fable 5 to developers.

OpenAI

AI company

Released GPT-5.5-Cyber to vetted security teams in May 2026; scores 58.6% on SWE-Bench Pro

AI company that builds the GPT model family, including a cybersecurity-focused variant competing with Anthropic's Mythos.

Timeline

January 2026 July 2026

13 events Latest: July 1st, 2026 · 2 months ago Showing 8 of 13

Tap a bar to jump to that date

July 2026
Fable 5 restored globally; Mythos 5 returns to US organizations
Latest Restoration

Anthropic re-enabled Fable 5 for all users worldwide and Mythos 5 for US organizations. Usage limits on Fable 5 apply through July 7; international Mythos 5 access remains under coordination with the Commerce Department.
July 1st, 2026
June 2026
US lifts export controls on Fable 5 and Mythos 5
Government Action

The Commerce Department reversed its June 12 directive after Anthropic, Amazon, Microsoft, and Google jointly developed a safety classifier blocking the cited jailbreak in more than 99 percent of cases.
June 30th, 2026
US government orders suspension of Fable 5 and Mythos 5
Government Action

Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei an export control directive at 5:21pm ET barring access for any foreign national, including Anthropic's own employees. Unable to screen users by nationality, Anthropic disabled both models globally, rolling back access to Claude Opus 4.8.
June 12th, 2026 • 5:21 PM
Anthropic commits to visible restrictions across all safety blocks
Policy Update

One day after its reversal, Anthropic said Fable 5 will notify users whenever a request is downgraded or redirected. That matches how the existing cybersecurity and biology blocks already worked.
June 11th, 2026
Anthropic reverses covert 'silent downgrade' policy after backlash
Policy Reversal

After backlash over a passage buried in Fable 5's 319-page system card, Anthropic reversed a policy that silently downgraded responses for AI development queries without notifying users. The company said it 'made the wrong tradeoff' and apologized.
June 10th, 2026
Anthropic releases Claude Fable 5 to the public
Release

The first generally available Mythos-class model posts 80.3% on SWE-Bench Pro but blocks cybersecurity, biology, and chemistry. A less-restricted variant, Mythos 5, goes to vetted partners.
June 9th, 2026
Fable 5 goes live in GitHub Copilot
Integration

GitHub makes the model generally available to Copilot users, putting frontier coding into a tool millions already use.
June 9th, 2026
Fable 5 launches on Amazon Bedrock
Integration

Amazon made Fable 5 available on Bedrock in the US East (N. Virginia) and Europe (Stockholm) regions, adding a major cloud channel alongside the Claude API and GitHub Copilot.
June 9th, 2026
Project Glasswing expands to 150 organizations in 15+ countries
Expansion

Anthropic added power, water, healthcare, communications, and hardware operators to Glasswing, citing 10,000+ high or critical vulnerabilities already found by existing partners across major operating systems and browsers.
June 2nd, 2026
May 2026
Anthropic ships Claude Opus 4.8
Release

The prior top model scores 69.2% on SWE-Bench Pro, a hard coding benchmark. It becomes the baseline Fable 5 is measured against.
May 28th, 2026
OpenAI releases GPT-5.5-Cyber to vetted security teams
Competitor Release

OpenAI rolled out a security-focused variant of GPT 5.5 to vetted cybersecurity teams in limited preview, one month after Anthropic's Mythos Preview debut. In testing, GPT-5.5-Cyber completed a corporate network attack simulation in 2 of 10 attempts; Mythos Preview did so in 3 of 10.
May 7th, 2026
April 2026
Unauthorized users access Mythos Preview through third-party vendor
Security

Bloomberg reported that a private online group guessed the model's URL from Anthropic's naming conventions and accessed it through a contractor's environment. Anthropic said it found no evidence of misuse.
April 21st, 2026
January 2026
Amodei warns AI is entering a 'dangerous adolescence'
Statement

Anthropic's chief executive publishes an essay arguing AI is advancing faster than society can control it, citing bioweapon and job-loss risks.
January 26th, 2026

#timeline [x-show="expanded"] { display: revert !important; opacity: 1 !important; } #timeline .timeline-toggle-btn { display: none; }

Scenarios

A rival lab matches or beats Fable 5's coding score

Likely Resolves by End of 2026

Discussed by: AI benchmark trackers and vendor announcements from OpenAI and Google DeepMind

Frontier labs trade benchmark leads every few weeks. Anthropic shipped Opus 4.8 just 12 days before Fable 5. A competitor publishes a model that scores at or above 80.3% on SWE-Bench Pro, erasing the lead and pushing the race forward again.

Anthropic relaxes a public safety block

Possible Resolves by End of 2026

Discussed by: Anthropic model-card updates and AI-policy coverage at TechCrunch and CNBC

The hard blocks on cybersecurity, biology, and chemistry are a starting position, not a fixed law. As Anthropic gains confidence in its safeguards or faces market pressure, it loosens at least one block for the public model, narrowing the gap between Fable 5 and Mythos 5.

A government moves to govern tiered AI access

Uncertain Resolves by Jun 9, 2027

Discussed by: AI-policy analysts covering the EU AI Act and US executive action

Anthropic's split between a gated public model and a partner-only variant raises a policy question: who decides who gets the dangerous version? A regulator responds by proposing rules that specifically address gated or tiered access to frontier models, rather than capability alone.

The partner-only Mythos 5 is leaked or misused

Unlikely Resolves by Jun 9, 2027

Discussed by: Security reporters at Reuters and Anthropic's own disclosures

The less-restricted variant is the whole risk of this design. If a vetted partner's access is stolen, resold, or turned to an unapproved use, the gate fails in public view. A credible report documents unauthorized access to or misuse of Mythos 5.

EU regulator challenges Glasswing's cross-border access model

Possible Resolves by Jun 1, 2027

Discussed by: Policy analysts at Nextgov/FCW and eWeek covering EU AI Act enforcement and OpenAI's earlier EU cyber access decisions

Project Glasswing now covers organizations in 15+ countries, including critical infrastructure in power, water, and healthcare. A regulator (most plausibly the EU under the AI Act) opens a formal inquiry into whether a US company can unilaterally grant or deny a Mythos-class model to national infrastructure operators.

Anthropic challenges the directive and wins restoration

Possible Resolves by End of 2026

Discussed by: Legal analysts at Law360 and Anthropic's own statement calling the action unfair and not grounded in technical facts

Anthropic disputed the jailbreak the government cited, saying the same capability exists in GPT-5.5 and is used daily by defenders. A legal challenge or successful appeal to the Commerce Department could restore access, at least for US users.

US export controls on frontier AI become standing policy

Possible Resolves by Jun 12, 2027

Discussed by: Policy analysts at Brookings Institution and The Hacker News covering the Commerce Department directive

The June 12 suspension may be the first use of export controls to force a live takedown of a frontier AI model. If the government codifies this authority, labs could face recurring directives to suspend models based on national security reviews—a new regulatory category for deployed AI.

Pre-launch government coordination becomes a standing condition for frontier AI

Possible Resolves by Q2 2027

Discussed by: Policy analysts at Nextgov/FCW and TechPolicy.Press covering the Anthropic-Lutnick deal terms

The deal that restored Fable 5 included a commitment to coordinate with the government on future launches. If Commerce formalizes this as a prerequisite — or if other labs adopt similar terms to avoid confrontation — every frontier model release would require prior federal sign-off. That is a de facto licensing regime, built through negotiation rather than rulemaking.

Historical Context

3 moments from history that rhyme with this story — and how they unfolded.

February 1975

Asilomar conference on recombinant DNA (1975)

About 140 scientists met at Asilomar, California, to set voluntary safety rules for the new technology of splicing DNA. They paused certain experiments until safeguards existed.

Then

The group agreed on containment levels matched to risk, and research resumed under shared guidelines.

Now

Asilomar is cited as a rare case of scientists regulating their own dangerous field before governments stepped in.

Why this matters now

Fable 5's hard blocks on biology and chemistry are a corporate version of Asilomar: the maker, not a regulator, draws the line.

1991–2000

The Crypto Wars over encryption exports (1990s)

Through the 1990s, the US government treated strong encryption as a munition and restricted who could export it. Companies that wanted to ship secure software abroad needed approval.

Then

Industry and privacy groups fought the controls in court and won key cases late in the decade.

Now

Most export limits eased by 2000, and strong encryption spread worldwide anyway.

Why this matters now

Anthropic's two-tier system echoes that fight: full power for the vetted, limits for everyone else. The history suggests such gates are hard to hold.

February–November 2019

OpenAI withholds the full GPT-2 model (2019)

In February 2019, OpenAI built a text generator called GPT-2 but declined to release the full version, saying it could mass-produce fake news and spam. It put out smaller versions over nine months instead.

Then

Researchers split over whether the move was prudent or hype. OpenAI released the full model in November 2019 after seeing little real misuse.

Now

Staged, risk-weighted release became a standard playbook for frontier AI labs.

Why this matters now

Anthropic is running the same logic at larger scale: ship the capability, but gate the riskiest parts and meter out the least-restricted version.