Meet Claude Fable 5, Anthropic’s New Mythos-Class Model

On June 9, 2026, Anthropic released Claude Fable 5, the first model in a brand new top tier the company calls the Mythos class. Seven weeks later it still sits at number one on Arena’s text leaderboard, the largest human-preference board in AI, with 16,056 votes behind a 1507.6 rating.

What changed since launch is the competition inside Anthropic’s own lineup. Claude Opus 5 arrived on July 24 at half the token price, and on most agentic and coding work it now matches or beats Fable 5. This article covers what Fable 5 is, how the Mythos class works, the full benchmark picture, the safeguard system, pricing, and the 18-day government shutdown that pulled it offline three days after launch.

It launched alongside a sibling model, Claude Mythos 5. That model is the same underlying system with the safety guardrails lifted, available only to a small set of vetted users. Fable 5 even became a punchline in the viral Le Chaton Fat meme, where users joked you would not need Fable if Mistral shipped a fat kitten model.

The new tier is the real headline. For two years Anthropic ran a three step ladder of Haiku, Sonnet, and Opus. Mythos is a fourth step above all of them, and Anthropic still describes Fable 5 as its most capable widely released model.

The Key Takeaways

• Fable 5 is number one on Arena’s text leaderboard at 1507.6, and also tops the vision and image-to-code boards.
• On Artificial Analysis it scores 60, third of 264 rows, just behind Claude Opus 5 at 61, which costs half as much.
• Pricing is $10 / $50 per million tokens, exactly twice Opus 5 ($5 / $25). It is not on the Free plan.
• Classifiers cover three domains, and Anthropic says 95%+ of sessions never trigger one. Blocked biology requests now route to Opus 5, not Opus 4.8.
• Mythos 5 is the same model with guardrails off, still restricted to vetted cyber defenders and approved biology researchers.

What Claude Fable 5 Actually Is

Claude Fable 5 is a frontier reasoning model and the first generally available member of Anthropic’s Mythos class. Anthropic’s own model documentation calls it the company’s most capable widely released model. Here are the basics at a glance:

• API identifier: claude-fable-5, with anthropic.claude-fable-5 on AWS Bedrock and Google Cloud
• 1 million token context window and 128,000 tokens of maximum output, both published by Anthropic
• Reliable knowledge cutoff of January 2026, four months behind Opus 5’s May 2026
• Adaptive thinking is always on, which is why Anthropic rates its latency as the slowest of the four current models
• Same model as Mythos 5, separated only by a safety layer added at inference time

The most important fact is that Fable 5 and Mythos 5 are the same model. The difference is not the weights or the training. It is the safety layer wrapped around Fable 5 that intercepts high risk requests. So every benchmark number Fable 5 posts is the Mythos class capability operating with a safety net underneath it.

Anthropic positions the model for autonomous work that runs for hours or even days inside an agent harness. In that setting the model plans a multi step job, calls tools, reads the results, validates its own output, and corrects course without a human in the loop.

The trait partners keep citing is self verification. Rakuten told Anthropic the model reflects on and validates its own work, which is what makes the autonomous operation practical rather than risky.

The Mythos Class Explained

Anthropic’s lineup has always been a ladder:

• Haiku is the small, fast, cheap tier
• Sonnet is the balanced workhorse
• Opus is the frontier reasoning tier
• Mythos is the new step above all three

Mythos represents a capability level Anthropic had been holding back from general release, because of the risks a model that strong could pose in the wrong hands.

Fable 5 is what you get when Anthropic takes a Mythos class model and adds the guardrails needed to make it safe for a broad audience. The pitch is that the safeguards are not a tax on the model. They are the thing that makes shipping it possible at all.

That framing mattered more at launch than it does now. In June the only comparison was Opus 4.8, a full tier below. Since July 24 the honest question is narrower, because Opus 5 delivers most of the same capability for half the price.

A Brief History of Mythos and Project Glasswing

Mythos did not appear out of nowhere on launch day. The first Mythos model, Claude Mythos Preview, shipped quietly in April 2026 through a limited program named Project Glasswing. An early Mythos build later made headlines when it bypassed Apple’s M5 memory protections in five days.

Project Glasswing was not a public product. It was a controlled access program aimed at a narrow set of users, mainly cyber defenders and critical infrastructure providers. These are organizations that could use a frontier model’s offensive security knowledge to strengthen their own defenses.

Anthropic used the program to learn how a Mythos class model behaves in the field before deciding whether to release it more widely. The June 9 launch is the graduation of that experiment into two official products:

• Claude Mythos 5, the direct upgrade for Glasswing partners and select biology researchers, available only through trusted access
• Claude Fable 5, the safeguarded version built for everyone else

That program has since grown well beyond its original scope. Glasswing started with 12 launch partners and reached roughly 50 organizations with Mythos Preview access by late May. On June 2, 2026 Anthropic extended it to approximately 150 more organizations in over 15 countries, covering power, water, healthcare, communications, and hardware.

Benchmark Results from the June 9 Launch

Anthropic published a full benchmark table with the launch, putting the Mythos class against Opus 4.8, GPT-5.5, and Gemini 3.1 Pro, which were the current flagships that day. Launch coverage from VentureBeat and Tom’s Hardware echoed the topline that it is state of the art on nearly all of them.

Two things matter for reading the table. The Anthropic column combines Mythos 5 and Fable 5, showing the higher of the two, which are normally within 1 to 3 points of each other. The starred rows are the exception, because that is where Fable 5’s safeguards hand the request to a fallback model, so the public Fable 5 score lands lower than the starred figure.

Starred rows are where Fable 5’s safeguards trigger a fallback. The number shown is Mythos 5, and public Fable 5 performs closer to the fallback model on those tasks. On Terminal Bench 2.1, GPT-5.5’s 83.4% is via Codex CLI and Gemini’s 70.7% via Gemini CLI.

The standouts are the unstarred rows, where Fable 5 and Mythos 5 are effectively the same. Fable 5 led every rival on the day on agentic coding, knowledge work, vision, tool use, computer use, and legal reasoning. The SWE-Bench Pro gap was the clearest signal at 80.3%, more than 11 points above Opus 4.8 and nearly 22 above GPT-5.5. Even the strongest open-weight coders, like Nex-N2-Pro at 58.8%, sat far behind. A different kind of challenger sidesteps the single-model race altogether, as Sakana AI’s Fugu orchestrates a pool of LLMs, and its top tier scores 73.7 on SWE-Bench Pro.

A few results outside the official table round out the launch picture:

• Analytics: first model to score 90% on Hex’s benchmark of complex, long running analytical tasks
• Finance: highest score of any model on Hebbia’s senior level finance benchmark
• Memory: a persistent file based memory task improved Fable 5’s performance three times more than Opus 4.8
• Spreadsheets: beats Opus 4.8 at every effort level while finishing runs 25 to 30% faster

Fable 5 vs Claude Opus 5, the Comparison That Now Matters

Anthropic released Claude Opus 5 on July 24, 2026 at $5 / $25 per million tokens, exactly half of Fable 5. In its own announcement Anthropic says Opus 5 “comes close to the frontier intelligence of Claude Fable 5 at half the price” and calls it “the new state-of-the-art” on coding and knowledge work evaluations.

Artificial Analysis measured the two head to head and reached the same conclusion. Their summary is that Opus 5 offers “comparable intelligence to Fable 5 at 26% lower Cost per Task”, and they describe the Intelligence Index result as effectively tied rather than a defeat.

Where Opus 5 Wins

On agentic knowledge work the gap is not close. Artificial Analysis scores Opus 5 at max effort at 1720 Elo on its AA-Briefcase benchmark against Fable 5’s 1574, a 146-point lead, while costing $17.79 per task against Fable 5’s $22.30. On the updated GDPval-AA v2 it scores 1861, more than 100 points ahead of both Fable 5 and GPT-5.6 Sol.

Anthropic’s own numbers point the same way. On CursorBench 3.2 at max effort, Opus 5 lands within 0.5% of Fable 5’s peak score at half the cost per task. On the OSWorld 2.0 computer-use benchmark it beats Fable 5’s best result at just over a third of the cost.

Where Fable 5 Still Wins

Human preference is the clearest one. On Arena’s text leaderboard Fable 5 is number one and Opus 5 sits fifth, and Fable 5 also leads the vision board where Opus 5 is fifth. Those are votes from people comparing real answers, not a scripted harness.

Raw factual knowledge is the other. Artificial Analysis notes plainly that Opus 5 “still has lower factual knowledge on AA-Omniscience than Fable 5”, which it attributes to the size difference between the two models. On Humanity’s Last Exam the two are level, with Opus 5 at 53% described as in line with Fable 5.

Anthropic’s own guidance splits the same way. Its developer docs tell you to start with Opus 5 for complex agentic coding and enterprise work, then add that for workloads needing the highest available capability, you should use Fable 5.

The Practical Recommendation

For most paid work Opus 5 is now the better default, because it costs half as much per token and less per finished task while scoring at or above Fable 5 on the agentic benchmarks. Fable 5 earns its premium on breadth of knowledge, on answer quality that human voters prefer, and on jobs where you want the absolute ceiling regardless of cost.

How Fable 5 Compares to Opus 5, GPT-5.6 Sol, and Gemini 3.1 Pro

Stripping the current field down to positioning, the four models line up like this. Prices are per million tokens from each vendor’s published rate card.

The pattern has inverted since June. Fable 5 no longer leads its own family on the agentic and coding boards, though it holds the top human-preference spot and the broadest factual knowledge. GPT-5.6 Sol is the closest external rival and actually leads the Artificial Analysis coding index at 78.3. If you are weighing the two biggest labs more broadly, our Anthropic vs OpenAI breakdown covers pricing, safety, and ecosystem beyond a single model.

One caveat on the coding numbers. Anthropic publishes no SWE-bench Verified figure for either Fable 5 or Opus 5, so any score you see attributed to them from a third party is not the vendor’s own measurement.

Independent Benchmarks from Artificial Analysis

The first independent confirmation came from Artificial Analysis, which folded Fable 5 into its public leaderboard within a day of launch. At the time it debuted at the very top, scoring 1932 on GDPval-AA for number one on agentic real-world work, with Anthropic models taking three of the top four spots.

Artificial Analysis has since rebased its Intelligence Index and added a newer generation of models. Here is where Fable 5 stands on the live board:

• Intelligence Index: 60, third of 264 rows, behind only Claude Opus 5 at max effort (61) and Opus 5 at xhigh (60)
• Coding index: 76.5, behind GPT-5.6 Sol at 78.3 and Opus 5 at 78.0
• Agentic index: 52.8, behind Opus 5 at 55.3 and GPT-5.6 Sol at 54.0
• Cost per task: $2.75, against $2.03 for Opus 5 and $1.80 for Opus 4.8
• Fallback rate: just 2% of GDPval-AA tasks deferred, matching Anthropic’s under 5% claim

The independent data also surfaces a weakness the partner testimonials gloss over. Artificial Analysis clocks Fable 5 at 74 tokens per second, respectable but well behind the fast tiers. Its time to first token runs far above the peer median, because of the heavy chain of thought reasoning it performs before answering. This is a model built for deep, long horizon work, not snappy back and forth chat.

Cost is the other caveat. At $2.75 per Intelligence Index task, Fable 5 is the most expensive of the frontier group on that measure, which is the arithmetic behind Anthropic steering most developers to Opus 5 first.

Fable 5 Tops Arena’s Human Preference Leaderboards

The scoreboard that was blank at launch has since filled in, and it is the best news on this page. Arena, formerly LMArena, ranks models by blind head to head votes from real users rather than by scripted benchmarks.

As of the July 27, 2026 vote cutoff, across a board carrying nearly 7.5 million votes over 381 models, Fable 5 places like this:

• Text: 1st at 1507.6 on 16,056 votes, ahead of every model from every lab
• Vision: 1st at 1318.1
• Image to code: 1st at 1625.8
• Search: 3rd at 1236.8
• Web development: 4th at 1629.1, where Claude Opus 5 leads at 1725.3

The split is worth reading carefully. On open-ended conversation and vision, where people judge the answer, Fable 5 wins. On web development, where the output is code you can run, Opus 5 and Kimi K3 both beat it. That is the same divide the Artificial Analysis numbers show.

Real World Performance from Launch Partners

Benchmarks are one signal. The launch partner reports point in the same direction.

Stripe produced the most cited result. Anthropic says Fable 5 compressed months of engineering into days. The specific example is a migration across a 50 million line Ruby codebase that the model completed in a single day, work that would otherwise have taken a full engineering team over two months.

Rakuten emphasized reliability over raw speed. The company reported that Fable 5 reflects on and validates its own work, letting it run autonomous operations where the model is trusted to check itself rather than handing every step back to a human.

Hex contributed the 90% analytics milestone, the first time any model cleared that bar on its suite of complex tasks. AWS framed the model as purpose built for long running, asynchronous execution, the kind of job that can run for days inside a harness before producing a result.

The Safeguard System

The safeguards are what make Fable 5 a public product, so they deserve a close look.

Fable 5 ships with classifiers covering three high risk domains. When a request trips one, the main model is prevented from answering and a more conservative Claude model handles the query instead. Anthropic confirms that users are told whenever this happens. The domains are:

• Cybersecurity, covering both exploitation and offensive cyber tasks in a broad sense
• Biology and chemistry, currently blocking most requests in the area rather than a narrow bioweapons list
• Distillation, blocking attempts to extract or copy the model’s capabilities

Anthropic is unusually candid that this net is too wide. Its launch post says the safeguards were deliberately tuned to be cautious and are “still stricter than would be ideal”, with benign requests sometimes triggering the classifiers, and that reducing false positives is the goal.

The Fallback Model Changed in July

At launch every blocked request went to Opus 4.8. That is no longer the whole story. With the Opus 5 release on July 24, 2026, Anthropic stated that biology-related requests blocked on Fable 5 now route to Opus 5 rather than Opus 4.8. The fallback answer on those topics now comes from a considerably stronger model than it did in June.

Developers also got a control they did not have at launch. Automatic fallbacks are now in beta on the API. You can choose to have requests flagged on Fable 5 route to another model automatically, so they are answered by the best available model instead of being blocked outright.

How Often It Actually Fires

The key number is how often this happens. Anthropic says more than 95% of Fable 5 sessions involve no fallback at all, and for those sessions Fable 5’s performance is effectively the same as Mythos 5. Artificial Analysis measured a 2% fallback rate across its GDPval-AA run, which is consistent with that claim.

The red team results are the other half of the safety case:

• No universal jailbreaks across more than 1,000 hours of external bug bounty testing
• Zero harmful single turn completions on cyberattack planning across 30 jailbreak techniques
• One external partner rated it the most robust safeguards of any model tested
• Mythos 5’s misaligned behavior measured low, similar to Opus 4.8

There is a cost and privacy dimension too. When a request is routed to a fallback model, you pay that model’s prices for that portion. Anthropic also requires 30-day retention of all traffic on Mythos-class models, on first and third party surfaces, as a change to how it handles business customer data. It says the data is not used to train Claude models or for any non-safety purpose. All human access to it is logged, and it is deleted after 30 days in almost all cases. On AWS this is enforced through a provider_data_sharing setting that you must enable before the model will run.

What Mythos 5 Can Do That Fable 5 Cannot

The starred rows in the benchmark table above are exactly where the two models diverge. On those, Mythos 5 answers at full capability while Fable 5’s safeguards hand the request to a fallback model. The gap is widest in the high risk domains:

• Cybersecurity (ExploitBench): Mythos 5 captures 78.0% against Opus 4.8 at 40.0% and GPT-5.5 at 34.0%, nearly doubling the previous frontier
• Biology hard (BioMysteryBench): Mythos 5 leads at 46.1% versus Opus 4.8 at 40.0%
• Health (HealthBench Professional): Mythos 5 at 66.0% versus Opus 4.8 at 56.9%
• Humanity’s Last Exam with tools: 64.5% for the Mythos class versus 57.9% for Opus 4.8

A general user hitting these topics through Fable 5 is redirected, which is why public Fable 5 lands near the fallback column on the starred rows. Mythos 5, available only to vetted cyber defenders and approved biology researchers, is the version that answers in full.

This is the core of Anthropic’s argument for the whole release. A model this capable in cybersecurity and biology is dangerous in the wrong hands, so that capability stays locked behind trusted access. Fable 5 lets everyone else use the same intelligence for everything that is not high risk.

Claude Fable 5 Pricing

Claude Fable 5 and Claude Mythos 5 are both priced at $10 per million input tokens and $50 per million output tokens. Prompt caching costs $12.50 per million to write and $1 per million to read.

The comparison that matters is the row beneath it. Opus 5 costs $5 and $25, which makes Fable 5 exactly twice the price of a model that outscores it on the agentic and coding boards. Against the earlier Claude Mythos Preview, Fable 5 costs less than half as much, so within the Mythos class the price dropped sharply even as capability went up.

Whether the premium is worth it depends on the work. For routine chat and for most agentic coding, Opus 5 is the better buy on both price and score. For jobs where breadth of factual knowledge matters, or where you want the highest ceiling Anthropic sells regardless of cost, Fable 5 is the one Anthropic itself points you to.

Availability and Access

Fable 5 has been generally available since June 9, 2026. New to Claude? Start with our walkthrough on how to get started with Claude and Claude Code before you reach for the top tier. The surfaces that carry it are:

• Claude API and Claude Platform, plus Google Cloud and Microsoft Foundry
• Amazon Bedrock, live in US East (N. Virginia) and Europe (Stockholm) at launch
• GitHub Copilot, generally available
• Harvey, the legal AI platform

Which Claude Plans Include Fable 5

The original free-inclusion window on paid plans was overtaken by the June 12 export-control order, and the terms changed when the model came back. Anthropic’s pricing page now sets it out per plan:

• Free: Fable 5 is not available at all
• Pro: available through usage credits, so it draws on paid credit rather than your standard allowance
• Max 5x and Max 20x: Fable 5 usage counts at 50% of your weekly limits

Claude Mythos 5 is the exception to all this breadth. It is restricted to Project Glasswing partners and a small set of approved biology researchers, and it is not part of any general subscription or public API tier.

Shut Down Three Days After Launch, Back After 18 Days

The biggest twist in the Fable 5 story is that it barely stayed online. Three days after the June 9 launch, the US government forced both Mythos-class models offline, and they stayed dark for 18 days before export controls were lifted on June 30, 2026.

The June 12 Export-Control Shutdown

On June 12, 2026, US Commerce Secretary Howard Lutnick sent Anthropic CEO Dario Amodei an export-control directive citing national security. It barred access by any foreign national, inside or outside the United States, including Anthropic’s own overseas staff. Because Anthropic cannot filter foreign nationals from US users in real time, it disabled both Fable 5 and Mythos 5 globally within hours, for every customer. The stated trigger was concern over a jailbreak method the government had learned of, and every other Claude model stayed fully available throughout.

Export Controls Lifted on June 30

After the 18-day standoff, Lutnick sent a letter dated June 30 to Anthropic’s Chief Compute Officer Tom Brown. It stated that a license is “no longer required for the export, reexport, or in-country transfer” of the Mythos or Fable models. Fable 5 returned to the Claude platform, Claude.ai, and Claude Code on Wednesday, July 1, 2026, with Mythos 5 partially reinstated the week before.

The reinstatement came with commitments. Anthropic agreed to proactively detect and report security risks, coordinate with the government on release protocols for current and future models, and disclose any malicious activity involving the Fable or Mythos models. Commerce also reserved the right to reimpose licensing if those obligations are not met, so the return is real but conditional.

The Timing Controversy

The release cannot be separated from its context. Just days before launching Fable 5, Anthropic warned that AI systems are advancing toward recursive self improvement, the point at which models can improve themselves without human intervention. The company paired that warning with a call for major AI labs to agree on a coordinated brake on frontier development.

Then it shipped its most powerful model yet. The apparent contradiction was not lost on anyone covering the launch. Releasing a Mythos class model to the public, days after arguing that AI is becoming too dangerous, reads as either a reversal or a deliberate strategy.

Anthropic’s answer is that the safeguards are the entire point. The argument runs that a Mythos class capability is coming whether Anthropic ships it or not. Better, it says, for the first public model at this tier to arrive with hard guardrails, fallback routing, and mandatory monitoring than to wait for a competitor to release something equally capable with no protections.

Whether that holds up is fair to debate. The safeguards are real and the red team numbers are strong. They also depend on classifiers that have to catch every dangerous request, on data retention that raises privacy questions, and on a fallback model users do not fully control.

Where Fable 5 Leaves Questions Open

Anthropic has since published the specs that were missing at launch, including the 1 million token context and the 128,000 token output ceiling. A few things are still not nailed down:

• No SWE-bench figure: Anthropic publishes no SWE-bench Verified score for Fable 5 or Opus 5, so the most widely quoted coding benchmark has no vendor number to check third-party claims against
• Vendor figures: the launch benchmark table is Anthropic’s own, and the full cross model comparison across every tested category has not been published
• Fallback experience: a model that silently reroutes a small share of sessions behaves differently near the boundary of the blocked domains, and the quality of that handoff is untested in public
• How long the premium lasts: with Opus 5 matching it at half the price, the case for a separate Mythos price tier rests on a narrowing set of tasks

Why This Release Matters

Three things make this launch significant beyond the benchmark numbers.

A new tier, public on day one. For two years the frontier conversation was about incremental gains within the Opus, GPT, and Gemini flagship lines, the models we track in our best AI models guide. Mythos was a deliberate step above that ladder, released generally rather than locked to a research preview. The opposite played out at OpenAI weeks later, where GPT-5.6 spent two weeks locked to about 20 government-approved organizations before the Commerce Department cleared broad access on July 9. Day-one public access is no longer guaranteed at the frontier.

A safeguard template the industry will study. Anthropic is betting that the way to ship a dangerous capability is to wrap it in classifiers, fallback routing, red teaming, and monitoring. Release the safe version broadly, keep the unrestricted version behind a trusted access wall. If it works, it becomes the playbook. If it fails, it becomes the cautionary tale.

A capability ceiling that moved fast. An 80.3% SWE-Bench Pro pass rate and a Stripe migration finished in a day pointed to a model that can carry long, complex jobs end to end. Seven weeks later Anthropic shipped most of that capability for half the price in Opus 5, which says as much about the pace of the field as the launch itself did.

For anyone who wants to see Claude together with other AI models, Fello AI puts Claude, GPT, Gemini, DeepSeek, Perplexity, and more into a single native app for Mac, iPhone, and iPad. You can test the same task across models and pick the right one for the job without managing a stack of subscriptions.

Frequently Asked Questions About Claude Fable 5