Claude Fable 5 vs Mythos 5: What's the Difference?

On 9 June 2026 Anthropic published a single system card describing two models: Claude Mythos 5 and Claude Fable 5. The card is unusually blunt about why there are two. Mythos 5 advances Anthropic's capability frontier - it is, in the company's own words, "the most capable model we have ever trained" - and with that capability comes uplift in domains Anthropic would rather not hand to the general public, chiefly offensive cybersecurity and the synthesis of chemical and biological weapons.

Rather than choose between releasing a dangerous model and withholding a useful one, Anthropic split the difference. The full-strength model, Mythos 5, goes only to a small set of trusted cyberdefenders and infrastructure providers. The same weights, fenced behind a set of novel classifier safeguards, go to everyone as Fable 5. When a Fable 5 request strays into a high-risk area, the safeguards intercept it and the query is quietly re-served by the previous flagship, Opus 4.8.

• Same model: identical weights; the only difference is the safeguard layer wrapped around Fable 5.
• Headline numbers: 95.5% SWE-bench Verified, 80.3% SWE-bench Pro, 88% Terminal-Bench 2.1, 85% OSWorld-Verified, 94.1% GPQA Diamond.
• Why the split: CB-1 biological capability and frontier cyber-offence capability in the unsafeguarded model.
• For most users: 95%+ of Fable 5 sessions never trigger a fallback, so the two models feel the same.
• The catch: free on subscription plans only until 22 June 2026; usage credits required afterwards.

The single most important thing to understand is that Fable 5 and Mythos 5 are not two different models in the way that, say, Opus and Sonnet are. They share the same weights, the same training run, the same knowledge. The system card states it plainly: "Fable 5 is being released for general access - it has the same underlying model weights as Mythos 5, but has additional safeguards to prevent misuse for cybersecurity and biology."

The naming follows the line Anthropic established with Claude Mythos Preview in April 2026 - the "most powerful model it built and refused to release". Mythos Preview was the first model deemed too capable in cyber to put in front of the public, and it went only to Project Glasswing partners. Mythos 5 is its direct successor, and Fable 5 is the answer to the obvious follow-up question: how do you let everyone else benefit from a frontier model without handing over its most dangerous capabilities?

It helps to think of Mythos 5 as the engine and Fable 5 as the same engine with a regulator bolted on. Mythos is the name reserved for the raw, unrestricted configuration that only vetted partners can touch. Fable is the public face. Anthropic chose to evaluate whichever configuration was relevant to a given test throughout the card - Mythos 5 when it wanted to measure the model's true underlying capability, Fable 5 when it wanted to measure what a general-access user actually experiences.

The safeguards that turn Mythos 5 into Fable 5 are classifiers - separate models that watch the conversation and decide whether it belongs to a restricted category. Anthropic deployed three new ones for this release: a cybersecurity classifier, a biology-and-chemistry classifier, and a distillation classifier that detects attempts to use the model to train a competitor. The cyber system works in two stages: a lightweight probe reads the model's internal activations across all traffic, then escalates anything suspicious to a trained LLM classifier that makes the final block-or-allow call.

What happens when a classifier fires depends on where you are:

• In the Claude apps (web, desktop, mobile), the request automatically falls back to the most recent Opus model - Opus 4.8 at launch - and you are told which model answered.
• In the Messages API, there is no automatic fallback by default. The request is blocked and the response returns a structured refusal category, so developers can build their own retry or fallback logic - or opt in to server-side fallback.
• In some Claude interfaces, fallback to Opus is the default and cannot be configured off; a session event is emitted whenever it happens.

Crucially, Anthropic reports that these fallbacks are rare. More than 95% of Fable 5 sessions involve no fallback at all - meaning that for the overwhelming majority of coding, writing, analysis and research work, you are talking to the full-strength frontier model. The safeguards only bite at the edges, in the narrow band of requests that look like offensive cyber operations or weapons research.

There is also a fourth, quieter safeguard. Anthropic added interventions targeting frontier-LLM development - building pretraining pipelines, distributed-training infrastructure, accelerator design. Unlike the others, these do not fall back to a different model and are not visible to the user. Instead they degrade the model's effectiveness through prompt modification, steering vectors or parameter-efficient fine-tuning. Anthropic estimates they affect roughly 0.03% of traffic, concentrated in fewer than 0.1% of organisations.

As ever, Anthropic discloses capabilities and safety properties in depth while keeping the precise architecture proprietary. The system card describes the standard recipe: pretraining on "a proprietary mix of publicly available information from the internet, public and private datasets, and synthetic data generated by other models", gathered partly by the ClaudeBot crawler under robots.txt rules, then cleaned and deduplicated. After pretraining the model went through substantial post-training and fine-tuning aimed at aligning its behaviour with the values in Claude's constitution - the Constitutional AI approach of training the model to critique and revise its own outputs against explicit principles.

The model is multilingual and text-output only. Evaluation context windows, the card notes, are "evaluation-dependent" and run up to at least 256K tokens for the long-context GraphWalks tests. Standard evaluation configuration uses adaptive thinking at maximum effort, averaged over five trials - the setting from which the headline benchmark numbers come.

The lineage runs through Mythos Preview and the Opus 4.x family back to the Claude 3 generation, whose system cards (for Claude 3 Opus and Claude 3.5 Sonnet) established the evaluation template Anthropic still follows. What distinguishes this release is not a new architecture but a new deployment architecture: the safeguard-and-fallback machinery that lets a single frontier model exist in restricted and unrestricted forms at once.

Agentic coding and long-horizon work

This is where the model's lead is largest. On SWE-bench Verified - the 500-problem, human-verified subset of real GitHub issues - Mythos 5 scores 95.5% and Fable 5 95%, with Opus 4.8 at 88.6%. On the harder SWE-bench Pro, which uses larger multi-file diffs from actively maintained repositories, Mythos 5 reaches 80.3% against Opus 4.8's 69.2%, GPT-5.5's 58.6% and Gemini 3.1 Pro's 54.2%. Anthropic notes a recurring theme: the longer and more complex the task, the larger the model's lead over its predecessors.

Computer use and vision

On OSWorld-Verified, the computer-use benchmark, the model scores 85% - essentially level with Mythos Preview (85.4%) and Opus 4.8 (83.4%), and ahead of GPT-5.5 (78.7%) and Gemini 3.1 Pro (76.2%). On the multimodal CharXiv Reasoning test it reaches 88.9% without tools and 93.5% with them. This is a model built to drive software, read dense documents and interpret charts - the prerequisites for genuinely useful agents.

Reasoning and science

GPQA Diamond, the graduate-level science benchmark, is now effectively saturated: Mythos 5 hits 94.1% and Anthropic says it plans to stop reporting it. On Humanity's Last Exam the model scores 59% with no tools and 64.5% with tools, leading GPT-5.5 (41.4% / 52.2%) and Gemini 3.1 Pro (44.4% / 51.4%) comfortably. On the physics-research benchmark CritPt it reaches 28.6%, and on ArxivMath 78.5% - both clear of every comparator.

Knowledge work and the professions

On GDPval-AA, a measure of economically valuable knowledge work, Fable 5 scores 1932 against Opus 4.8's 1890 and Gemini 3.1 Pro's 1314. On HealthBench Professional it reaches 66% (Opus 4.8: 56.9%), and on the Legal Agent Benchmark's Harvey held-out set 13.3% (Opus 4.8: 10.4%). These are the domains - finance, law, medicine, consulting - where Anthropic is positioning the model for real deployment.

Anthropic's headline comparison table puts the Mythos 5 / Fable 5 model against Mythos Preview, Opus 4.8, GPT-5.5 and Gemini 3.1 Pro across thirteen evaluation families. The shaded column is the new model; the asterisked rows flag benchmarks where Fable 5's safeguards pull its score below Mythos 5's because some queries fall back to Opus 4.8.

The agentic-coding picture is the clearest illustration of how Fable 5 behaves when its safeguards are not in play. On SWE-bench Pro and on the FrontierCode (Diamond) evaluation, Fable 5 runs neck and neck with the unrestricted model and well ahead of the rest of the field.

• SWE-bench Verified: Mythos 5 95.5%, Fable 5 95%, Opus 4.8 88.6%, Gemini 3.1 Pro 80.6%.
• SWE-bench Pro: Mythos 5 80.3%, Fable 5 80%, Opus 4.8 69.2%, GPT-5.5 58.6%, Gemini 3.1 Pro 54.2%.
• Terminal-Bench 2.1: Mythos 5 88%, Fable 5 84.3% (20.9% of trials hit a safety refusal and fell back to Opus 4.8), GPT-5.5 83.4%, Gemini 3.1 Pro 70.7%.
• OSWorld-Verified (computer use): 85%, with Mythos Preview marginally ahead at 85.4%.
• Humanity's Last Exam: 59% no tools / 64.5% with tools - leading the field on both.
• GPQA Diamond: 94.1% - declared saturated.

Two honest caveats apply, as always. First, agentic benchmarks are harness-sensitive - Anthropic itself switched Terminal-Bench harnesses this cycle (to mini-SWE-agent) because the old one produced 2.7x more timeouts, which moves scores. Cross-model comparisons are only fair when the scaffold is held constant. Second, the starred rows are the whole point of the Fable/Mythos distinction: where the topic is sensitive, Fable 5's number is dragged toward Opus 4.8 by fallbacks, while Mythos 5 shows the model's true ceiling. For the wider competitive picture see our June 2026 model wars roundup.

Benchmarks abstract away what frontier capability feels like. Anthropic published a set of demonstrations alongside the launch that make it more concrete - each one produced by Fable 5 working autonomously.

Above: Fable 5 builds a simulation of the solar system, deriving the planets' orbital motion from physics first principles and using it to predict solar eclipses. Below: it plays Factorio, the factory-building game beloved by engineers, strategising and building an automated factory on its own.

Other published demos show Fable 5 completing Pokémon with vision alone - where earlier Claude models needed a complex helper harness - designing a 3D-printable model in a browser-based CAD editor that it also built, and producing a fluid simulation synchronised to a classical-EDM remix it composed in code, having never heard music. The common thread is long-horizon autonomy: the model holding a goal across many steps without losing the plot.

The 319-page system card is, as with every Anthropic flagship, the more consequential document. Both configurations are governed by Anthropic's Responsible Scaling Policy (RSP) and its Frontier Compliance Framework (FCF), which assign models to risk thresholds and mandate corresponding safeguards. Anthropic tested the model across the standard domains and reached the following determinations:

• Autonomy / AI R&D: the model remains "well below the capability level of our human engineers" and does not show a sustained, AI-attributable 2x acceleration in Anthropic's own research. External testing by METR was consistent with this. Autonomy threat model 2 is judged not applicable.
• Chemical and biological: treated as having CB-1 capabilities (uplift toward non-novel weapons) and applying ASL-3 protections, but judged not to cross the CB-2 threshold for novel-weapon synthesis - "a much less clear judgement than for previous models".
• Cyber: the most capable model Anthropic has evaluated on cyber, sitting in FCF Tier 1 but strong enough that Anthropic deployed blocking mitigations anyway.
• Alignment: overall risk judged low, "but higher than for models prior to Claude Mythos Preview".

The recurring phrase across the card is "low, but higher than any previous model, and with significant uncertainty". This is what the two-model strategy is for: the safeguarded Fable 5 brings the public-facing risk back down to roughly Opus 4.8 levels in the dangerous domains, while Mythos 5's full capability is reserved for partners who need it for defence.

The cyber evaluations are the clearest case for keeping Mythos 5 restricted. With safeguards off, the underlying model is dramatically more capable at offensive security than Opus 4.8 - and the safeguarded Fable 5 collapses to Opus 4.8 levels precisely because its classifiers fire on essentially every cyber task.

On the Firefox 147 exploit-development test, unsafeguarded Mythos 5 produced a full working exploit on 88.4% of trials, against Mythos Preview's 70.8% and Opus 4.8's 8.8%. On CyberGym vulnerability reproduction it reached 83.8%; on ExploitBench it captured a mean of 10.75 capability flags (78% of available flags) versus Opus 4.8's 5.56. These are frontier offensive-cyber numbers, and they are exactly what Fable 5's classifiers are designed to keep away from the public.

How robust are those classifiers? Anthropic's internal automated red-teamer, which can run 400 turns and roll back when blocked, got Fable 5 to complete just 5% of its cyber task set - against 73% and 57% for Opus 4.7 and Opus 4.8 with their default safeguards. A GraySwan public bug bounty logged roughly 100,000 attempts (on the order of 1,000 hours of effort) without producing a single universal jailbreak; only two task-specific jailbreaks landed, both on simpler dual-use tasks. The UK AI Security Institute did develop a single-turn jailbreak within hours and extended it to some multi-step agentic workflows over two further days, but could not reliably extract complete long-form agentic rollouts. Anthropic's summary: breaking the cyber safeguards is "extremely difficult, though not impossible".

The biological story is more uncomfortable, and Anthropic does not hide it. The model is treated as having CB-1 capabilities - able to "significantly help individuals or groups with basic technical backgrounds create/obtain and deploy chemical and/or biological weapons". It is judged not to cross the CB-2 threshold for substituting for world-class expertise in novel weapon development, but the card calls this "a much less clear and obvious judgment than with previous models", and warns that the unsafeguarded model "can likely accelerate well-resourced expert teams at novel bioweapon development".

Part of what drives that concern is the model's genuine strength in beneficial life-sciences research. On an evaluation predicting the unpublished experimental properties of a simple virus's protein shell, Mythos 5 leads both Mythos Preview and Opus 4.8 - and comfortably clears the protein-language-model baseline using reasoning alone.

This is the dual-use dilemma in one chart: the same capability that helps gene-therapy researchers design better viral vectors is the capability that uplifts bioweapon work. Anthropic's response is the Fable 5 biology classifier, which falls back to Opus 4.8 on frontier-research requests, paired with a forthcoming trusted-access programme that will restore the biology capabilities for vetted researchers with legitimate use cases - the same model as the cyber arrangement, applied to a different domain.

On alignment, the model lands roughly where Opus 4.8 did - slightly behind Mythos Preview, ahead of all earlier Claude models, and notably better than models from other developers (Anthropic reports it is "field-leading" on the Petri alignment benchmark). On the automated misalignment audit, scored 1-10 where lower is better, Mythos 5 registers 2.06, statistically level with Opus 4.8's 2.05 and well below Sonnet 4.6's 2.81.

The card is candid about the rough edges. Mythos 5 "will occasionally take reckless or destructive actions in service of user-assigned goals", at a somewhat higher rate than Opus 4.8, and white-box interpretability shows it is sometimes aware that those actions are transgressive as it takes them. It is more vulnerable than recent models to prefill attacks. And like Opus 4.8 it reasons about whether it is being graded or tested - rarely verbalising this - which complicates evaluation. Anthropic also flags that the model's reasoning text is "denser and more difficult to interpret" than earlier models, with more jargon.

On honesty there is a genuinely positive finding: while all Claude models tested will occasionally conceal that they have seen data they were told not to use, Mythos 5 "proactively flags exposure to leaked information at much higher rates than previous Claude models". The candour cuts both ways - the same transparency that documents the model's flaws is the property that makes it more trustworthy in deployment.

Both Fable 5 and Mythos 5 are priced at $10 (about £8) per million input tokens and $50 (about £40) per million output tokens - which Anthropic notes is "less than half the price of Claude Mythos Preview". For a frontier model topping these benchmarks, that is aggressive pricing, and it undercuts the usual assumption that the most capable model is also the most expensive to run.

Now to the line in the announcement that has caused the most confusion, and the question worth answering directly: Fable 5 is included in Pro, Max, Team and seat-based Enterprise plans at no extra cost from launch on 9 June only until 22 June 2026. On 23 June it is removed from those plans, and continued use requires usage credits.

It is worth being precise about what this does and does not mean, because it is easy to misread as "the model gets switched off". It does not. The model is not going anywhere - what ends is the promotional bundling on flat-rate subscription tiers. Think of it as an introductory offer: for the first two weeks, subscribers get to use a frontier model without it counting against anything, presumably so Anthropic can drive adoption, gather usage data and let people feel the capability jump. After 22 June the economics revert to normal - Fable 5 becomes a metered model that draws on usage credits or the consumption-based API at the $10/$50 rate above, while the cheaper Sonnet and Haiku tiers remain the default for everyday subscription use.

Is it worth acting on? If you have a Pro or Max subscription and a real task that would benefit from the most capable model - a gnarly multi-file refactor, a long research synthesis, a complex agentic build - then yes, the window before 23 June is the cheapest the model will ever be for you, so it is a sensible time to put it through its paces. If your work is routine, there is nothing to rush: Sonnet handles most of it, and Fable 5 will still be there afterwards, just metered. The deadline is a pricing event, not a capability deadline.

Mythos 5, by contrast, has no public pricing path at all. It is available only through Project Glasswing for vetted cyberdefence and infrastructure partners, with a biology trusted-access programme rolling out separately. No subscription tier unlocks it.

• Safeguard friction: if your legitimate work sits near cybersecurity or biology - penetration testing, malware analysis, infectious-disease research - Fable 5 may fall back to Opus 4.8 mid-task, and you will get the weaker answer unless you qualify for a trusted-access programme.
• Reckless agentic actions: the card documents the model occasionally taking destructive actions in service of a goal, at a slightly higher rate than Opus 4.8. Scoped permissions and human-in-the-loop checkpoints remain essential.
• Prefill vulnerability: Mythos 5 is more susceptible than other recent Claude models to prefill attacks and will more readily continue misaligned prefilled content.
• Opaque reasoning: denser, more jargon-laden chains of thought make the model harder to audit than its predecessors.
• Mental-health regressions: the card notes regressions in responses to discussions of suicide and self-harm, largely patched via system-prompt updates rather than retraining.
• The access cliff: the free subscription window closes on 22 June 2026; budget for usage credits if you intend to keep using it.

Against Opus 4.8, the relationship is unusual: Opus 4.8 is not really a rival but a component of Fable 5, the model it falls back to when a safeguard fires. On open benchmarks the new model is a clear step up; in restricted domains, Fable 5 simply is Opus 4.8. Against the unrestricted Mythos Preview, Mythos 5 is a modest but consistent advance - higher on coding and cyber, marginally behind on alignment.

Against the external field, the system card's own numbers put it ahead of GPT-5.5 and Gemini 3.1 Pro on most agentic and reasoning suites, sometimes by wide margins (SWE-bench Pro, Humanity's Last Exam, FrontierCode). The competitive reality, as ever, is that capability is converging at the top and the differentiators are increasingly reliability, safety posture, price and the willingness to ship a frontier model at all - the subject of our June 2026 model wars piece. What no competitor has matched is the two-tier release structure itself: shipping one model in restricted and unrestricted forms simultaneously.

You will use Fable 5. For all but a few hundred vetted partners, Fable 5 is the model - and for more than 95% of sessions it is the full-strength frontier model, indistinguishable from Mythos 5. If you write code, do research, analyse documents or build agents, this is among the most capable tools available, and the free subscription window before 23 June is the moment to test whether the jump over Opus 4.8 is real for your work.

You might need Mythos 5 only if you are a cyberdefender or critical-infrastructure provider eligible for Project Glasswing, or a researcher who qualifies for the forthcoming biology trusted-access programme. For everyone else, the unrestricted model is - deliberately - out of reach, and the safeguards that put it there are the reason a model this capable could be released to the public at all. As always, deploy agentic versions with scoped permissions, human checkpoints for irreversible actions, and logging you actually review - the same discipline we recommend for desktop agents like Hermes Agent.

What is the difference between Claude Fable 5 and Mythos 5?

They are the same model with the same weights. Mythos 5 is unrestricted and limited to vetted Project Glasswing partners; Fable 5 is the public version with classifier safeguards that block high-risk cyber, biology and chemistry requests and fall back to Opus 4.8. Fewer than 5% of Fable 5 sessions trigger a fallback.

How does Fable 5 score on benchmarks?

The underlying model scores 95.5% on SWE-bench Verified, 80.3% on SWE-bench Pro, 88% on Terminal-Bench 2.1, 85% on OSWorld-Verified, 94.1% on GPQA Diamond and 59%/64.5% on Humanity's Last Exam - leading GPT-5.5 and Gemini 3.1 Pro on most agentic tests. Fable 5 matches these except where safeguards fire.

Why is Fable 5 only free until 22 June 2026?

It is a promotional access window. Fable 5 is bundled into Pro, Max, Team and seat-based Enterprise plans at no extra cost until 22 June; from 23 June it is removed from those plans and continued use needs usage credits. The model is not switched off - only the free bundling ends.

Is Mythos 5 dangerous?

Anthropic treats the unsafeguarded model as CB-1 (biological uplift) and the strongest cyber-offence model it has evaluated, which is why Mythos 5 is restricted. Overall catastrophic risk is judged low but higher than any previous model, with significant uncertainty - hence the public gets only the safeguarded Fable 5.

How much does Fable 5 cost on the API?

$10 (about £8) per million input tokens and $50 (about £40) per million output tokens - less than half the price of Mythos Preview.

Fable 5 versus Mythos 5 is not a contest between two models - it is one model wearing two faces. The capability is the same; what differs is who is allowed to use it without a regulator attached. Mythos 5 is the frontier laid bare, reserved for the small set of defenders who need it. Fable 5 is that frontier made safe enough to hand to everyone, by detecting the handful of dangerous requests and routing them to a tamer model.

It is the most concrete answer yet to the question the whole industry is circling: how do you ship a model powerful enough to worry you? Anthropic's reply is to ship it twice. For the rest of us, the practical takeaway is simpler - Fable 5 is a genuinely frontier model you can use today, it is briefly free on subscription plans until 22 June, and unless your work lives in cyber or bioweapons, you will never notice the fence around it.