Claude Fable 5 is Anthropic’s first publicly-available Mythos-class model, released on Tuesday June 9, 2026. Mythos has existed since an April 2026 limited preview to a handful of partners, then expanded last week to hundreds of organizations across 15 countries focused on critical-infrastructure operators. Fable 5 is the version of that same model technology that any developer with a Claude API account or a paying Claude subscription can access today, with one critical caveat: Fable 5 ships with hard safety limits that block responses in cybersecurity, biology, chemistry, and distillation areas, falling back to Claude Opus 4.8 for those requests. Anthropic’s own data shows 95% of Fable sessions run entirely on Fable’s own responses; the safeguard fallback is real but rare. Pricing is $10 per million input tokens and $50 per million output tokens, double Opus 4.8. Through June 22, Fable 5 is included in Pro, Max, Team, and seat-based Enterprise plans at no extra cost. Starting June 23, Anthropic moves Fable 5 to usage-credit billing on those plans, with stated intent to restore it as a standard subscription feature "as soon as possible."
The release is significant for three reasons that go beyond a normal model launch. It’s the first time a Mythos-class model has been available to the public at all. It comes with a precedent-setting 30-day mandatory data-retention policy that overrides prior zero-retention enterprise agreements. And it lands a week after Anthropic’s public plea to major global AI labs to establish coordinated brakes on frontier-AI development, including warnings about recursive self-improvement. This post covers what Fable 5 actually is, how it differs from Mythos itself, what the guardrails block, the precedent-setting retention policy, the pricing and access rollout, the third-party benchmark performance, the strategic context, and what teams building on Claude should be doing right now.
What Fable 5 actually is
Mythos is Anthropic’s most advanced model class, sitting above Opus 4.8 in capability. It launched in April 2026 as a limited preview specifically because Anthropic was concerned about the model’s superhuman ability to identify and exploit cybersecurity vulnerabilities, which made wide release inadvisable until safeguards were in place. The initial preview was restricted to a handful of partners working on defensive cybersecurity. On June 2, 2026, Anthropic expanded Mythos access to several hundred organizations across 15 countries, focused on critical-infrastructure operators (utilities, finance, healthcare, government) where the defensive applications outweighed the misuse risk.
Fable 5 is the same underlying technology with public-safe guardrails. The model itself is Mythos under the hood; what’s added is a classifier-and-fallback system that detects prompts in the high-risk areas (cybersecurity, biology, chemistry, distillation) and either blocks the response outright or routes the query to Claude Opus 4.8, which has had its safety story stress-tested for far longer. The result is a Mythos-class capability surface for everything except the explicitly-gated domains.
For developers using Claude today, the practical framing: Fable 5 is the new top of the Claude lineup, sitting above Opus 4.8 in capability for everything Fable will actually answer. For the queries Fable defers, you get Opus 4.8, which is the same model you’d be using anyway if Fable didn’t exist. Worst case is Opus 4.8 baseline; best case is Mythos-class output.
The capability claims
Anthropic’s own framing at release: "Fable’s capabilities exceed those of any model we’ve ever made generally available" and the model "is state-of-the-art on nearly all tested benchmarks of AI capability, showing exceptional performance in software engineering, knowledge work, vision, scientific research, and many other areas."
The third-party endorsements that landed in the release window are more substantive than usual for a model launch. Hex (analytics platform) said Fable 5 was the first model to score 90% on Hex’s core analytics benchmark of complex, long-running analytical tasks: "On the hardest questions, it shows strong judgement and attention to nuance." Base44 (vibe-coding platform) noted Fable is meaningfully better at "one-shotting full apps" and has excellent tool-calling. Genspark (AI-powered workspace and agent platform) said Fable beat every other model it evaluated, with significantly better performance on UI design and game coding. Rakuten (shopping rewards) emphasized the highest-effort behavior: "Fable reflects on and validates its own work. For us, that’s what makes highly autonomous operations possible. The extra thinking pays for itself."
The pattern across the testimonials is consistent: Fable 5 is meaningfully stronger than Opus 4.8 on the tasks third parties were already running against the Claude line. The 90% on Hex’s analytics benchmark is a specific public datapoint worth noting; the other vendors describe their improvements qualitatively. Independent third-party benchmarks haven’t been published yet at scale and are worth waiting for if benchmark precision matters to your evaluation decisions.
For context on where Fable 5 sits in the broader Claude lineup, our Claude Opus 4.8 launch coverage covers the model that Fable defers to on safety-gated queries, and our Opus 4.7 vs 4.8 comparison covers the recent Opus dot-release history.
The guardrails: what Fable 5 blocks
The four explicitly-gated areas in Fable 5:
Cybersecurity. The original reason Mythos was restricted to limited preview. Mythos-class models have demonstrated superhuman ability to find and exploit vulnerabilities, which is useful for defensive red-teaming but dangerous as a general capability surface. Fable 5 classifiers detect cybersecurity-shaped prompts and either block the response or fall back to Opus 4.8.
Biology. Bioweapon-uplift concerns, both Anthropic’s stated reason for caution and the consistent industry-wide concern about frontier-model deployment. The gating likely covers more than weapons specifically (depending on classifier scope), so general biology questions may also route to Opus 4.8.
Chemistry. The chemical-weapon-uplift parallel of the biology concern. Same classifier-and-fallback pattern.
Distillation. This is the technical-AI concern: Fable 5 blocks prompts designed to elicit training-data-quality outputs that could be used to train competing models. Anthropic has previously called out distillation by foreign labs (most prominently in a February 2026 statement about Chinese AI labs mining Claude outputs). The gating here is about model-property protection rather than safety in the traditional sense.
Anthropic’s stated 95% pass-through figure (95% of Fable sessions run entirely on Fable’s own responses without ever triggering a fallback) suggests the classifiers are well-tuned for typical workloads. For software engineering, analytics, knowledge work, vision, and most scientific-research queries, you’ll get Fable answers throughout. For the gated areas, you get Opus 4.8 underneath, which is good-but-not-Fable performance.
The classifiers themselves got substantial pre-release stress-testing. Anthropic ran an external bug bounty that produced "no universal jailbreaks in over 1,000 hours of testing." External red-teaming organizations also failed to find universal jailbreaks. This doesn’t mean novel attacks won’t emerge after release (they almost always do), but the bar to reach Fable’s underlying Mythos capability without the safeguards is meaningfully high.
The 30-day retention policy
The most operationally significant detail in the Fable 5 release is the data-retention policy, which sets a possible industry precedent.
With Fable 5 and Mythos 5, Anthropic is requiring a 30-day retention period on all traffic to the model, regardless of whether the enterprise has previously had a zero-retention agreement with Anthropic. The data isn’t used for training. Anthropic’s stated purpose is to "defend against complex and novel attacks, including new jailbreaks" and to "identify and reduce false positives" in the classifier system.
The precedent this sets is significant. Enterprise zero-retention agreements have been a standard part of Anthropic’s commercial offering for organizations with regulatory or contractual constraints on AI vendor data handling. Overriding those agreements for a specific model class introduces a new constraint: access to the most capable Anthropic models comes with mandatory data retention as a safety measure. For some organizations (regulated industries with strict data-residency or zero-retention requirements), the retention policy may be a deal-breaker that locks them out of Fable 5 use entirely.
The broader industry implication: if Anthropic establishes this pattern (zero-retention agreements override-able for specific high-risk model classes), other frontier-model providers may follow. The trade-off frontier customers will face going forward: maximum capability with mandatory data retention, or zero-retention with a less capable model tier. That’s a new dimension of commercial AI agreement that didn’t exist a week ago.
Pricing and access rollout
Pricing. Fable 5 (and Mythos 5) cost $10 per million input tokens and $50 per million output tokens. That’s double Claude Opus 4.8’s pricing ($5/$25 per million). Anthropic has positioned the cost as worth it for the Mythos-class capability; the testimonials from third-party launch partners emphasize that the extra capability "pays for itself" for highly autonomous operations. Whether the price-to-capability ratio works for any specific workload depends on the workload.
Subscription access (through June 22). Pro, Max, Team, and seat-based Enterprise plans get Fable 5 access at no extra cost through June 22. This is a finite-window promotion; teams that want to evaluate Fable 5 on subscription plans have approximately two weeks from release to do so.
Subscription access (starting June 23). After the promotional window, Fable 5 moves to usage-credit billing on those plans. Subscribers will need to purchase additional credits to use Fable 5 beyond what their existing plan covers. Anthropic has said they intend to restore Fable 5 "as a standard subscription feature as soon as possible," but no specific timeline is committed.
API access. Fable 5 is available immediately through the Claude API and through consumption-based Enterprise plans (which are billed differently than seat-based Enterprise). For developers building product features, the API is the right path; the subscription is more relevant for end-user-facing usage like Claude.ai chat.
Mythos 5 access. The base Mythos model also got a 5 upgrade, deployed today to the organizations that previously had Mythos access. Mythos 5 has the same pricing as Fable 5 but without the public-facing safety classifiers.
The strategic context
Three things landed in the news cycle adjacent to the Fable 5 release that are worth holding in your head.
Anthropic’s recursive self-improvement (RSI) warning. A week before the Fable 5 release, Anthropic published a plea urging major global AI labs to establish coordinated brakes on frontier-AI development. The specific concern: systems may soon achieve recursive self-improvement (the ability to autonomously improve themselves without human intervention), and the timing of that transition is poorly predictable. Releasing Fable 5 (Anthropic’s most capable model) days after publishing that warning reads as a deliberate juxtaposition: Anthropic is committed to deploying capability while signaling concern about the trajectory.
Anthropic’s IPO preparation. Multiple reports indicate Anthropic is preparing to enter public markets, alongside OpenAI (which filed confidentially for IPO last week) and SpaceX. The model lineup at IPO matters for the company’s positioning and valuation; Fable 5 being the most powerful publicly-available model in the industry is a meaningful talking point.
Industry-wide AI cost pressure. Enterprise teams are increasingly critical of AI bills as advanced reasoning models split single requests into multiple inference calls. Opus 4.8 has already contributed to runaway costs at some customers; Fable 5 at 2x the price will exacerbate the issue for teams that adopt it broadly. The pricing is a real deterrent to widespread use, which may be by design from a safety-and-control standpoint.
For broader context on the agentic coding-tools landscape that Fable 5 lands in, our Claude Code vs OpenAI Codex comparison covers two of the major coding-agent products that Fable 5 will power. Our Project Polaris coverage covers Microsoft’s competing in-house coding model that replaces GPT-4 in GitHub Copilot starting August 2026.
What teams building on Claude should do right now
Six concrete actions worth taking before June 22 when the free-on-subscription window closes:
- Evaluate Fable 5 on the actual workloads that have been Opus 4.8 bound for capability reasons. If your team has been running into ceiling effects with Opus 4.8 on analytics, knowledge work, or long-running autonomous tasks, Fable 5 is the upgrade to test first. Run real prompts, not benchmarks.
- Audit your prompts and workflows for the gated areas (cybersecurity, biology, chemistry, distillation). If a substantial fraction of your workload falls into those domains, Fable 5’s fallback behavior means the upgrade is less material than it would be otherwise; you’ll be on Opus 4.8 for the gated queries.
- Test the 95% pass-through claim against your specific workload. Anthropic’s data is aggregate; your specific prompt mix may trigger more or fewer fallbacks. The fallback behavior is detectable in the API response, so logging it across a representative sample will show you the real Fable-to-Opus split for your use case.
- Review the 30-day data-retention policy against your compliance constraints. If your team has been relying on zero-retention agreements with Anthropic for regulatory reasons, Fable 5 isn’t usable for you without working through whatever process change the retention introduces. Engage your compliance team now if so.
- Plan the cost story before June 23. Free Fable 5 through June 22 is great for evaluation; after that the usage-credit model means real costs. If Fable 5 makes the workload meaningfully better, the math may still work; if it doesn’t, plan to fall back to Opus 4.8 (or even Sonnet) for the bulk of inference and use Fable 5 only where it pays.
- Watch for novel jailbreak disclosures. The 1,000-hour external bug bounty without universal jailbreaks is impressive but isn’t the same as proving none exist. The first month or two of public Fable 5 use will surface the attacks Anthropic’s pre-release testing didn’t catch. If your workload is safety-sensitive, plan for jailbreak-detection monitoring rather than assuming the classifiers are airtight.
The deeper takeaway is that Fable 5 changes the top of the Claude lineup in a way that matters operationally. For teams that have been on Opus 4.8 expecting incremental improvements every few months, Fable 5 is a more substantial step than the typical Opus dot-release. For teams that have been holding off on Claude entirely because the gap to GPT or Gemini felt manageable, Fable 5 widens that gap meaningfully on the workloads where third-party testers have measured it. The hardest decision in front of most Claude users today is how aggressively to adopt Fable 5 before the pricing transition on June 23 makes the calculus less generous.
Frequently Asked Questions
What is Claude Fable 5?
Claude Fable 5 is Anthropic’s first publicly-available Mythos-class model, released on June 9, 2026. Mythos is Anthropic’s most advanced model class, previously restricted to a limited preview (April 2026) and then expanded to several hundred critical-infrastructure organizations across 15 countries (June 2, 2026). Fable 5 is the same underlying model technology with public-safe guardrails that block responses in cybersecurity, biology, chemistry, and distillation areas. For queries in those gated areas, Fable falls back to Claude Opus 4.8.
How is Fable 5 different from Mythos 5?
Mythos 5 is the unrestricted Mythos-class model, available to the limited population of pre-approved organizations (primarily critical infrastructure operators) that had Mythos access before today. Fable 5 is the public-safe variant with hard safety classifiers and Opus 4.8 fallback for gated domains. Underneath, both use the same underlying model technology. The pricing is identical ($10 per million input tokens, $50 per million output tokens).
What does Fable 5 cost?
$10 per million input tokens and $50 per million output tokens, which is double Claude Opus 4.8’s pricing. Through June 22, Fable 5 is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans (a roughly two-week promotional window from release). Starting June 23, Fable 5 access on those plans moves to usage-credit billing. API access (Claude API and consumption-based Enterprise) was paid at the standard rate from day one.
What areas does Fable 5 refuse to answer?
Cybersecurity (the original reason Mythos was preview-only, due to superhuman vulnerability identification), biology (bioweapon-uplift concerns), chemistry (chemical-weapon-uplift concerns), and distillation (training-data-quality outputs that could be used to train competing models). For prompts in these areas, Fable 5 either blocks the response outright or falls back to Claude Opus 4.8. Per Anthropic’s own data, 95% of Fable sessions run entirely on Fable’s own responses without triggering a fallback.
What is the 30-day retention policy and why does it matter?
Fable 5 and Mythos 5 traffic is subject to a 30-day data retention period that overrides any prior zero-retention agreements an enterprise customer may have had with Anthropic. The data isn’t used for training; Anthropic says it’s used to “defend against complex and novel attacks, including new jailbreaks” and to “identify and reduce false positives” in the classifier system. The precedent matters for the industry: access to the most capable Anthropic models now requires accepting mandatory retention as a safety measure. For regulated industries with strict data-residency or zero-retention requirements, this may make Fable 5 unusable.
How does Fable 5 compare to OpenAI’s and Google’s frontier models?
Third-party benchmarks haven’t been published at scale yet, so direct comparisons are positioning rather than empirical. Anthropic’s framing is that Fable 5 is “state-of-the-art on nearly all tested benchmarks.” Specific third-party endorsements at launch: Hex (90% on its core analytics benchmark), Base44 (better one-shotting full apps), Genspark (beat every other model evaluated). The honest summary: Fable 5 is meaningfully better than Claude Opus 4.8 on the workloads testers measured, and Anthropic believes it’s competitive with or ahead of OpenAI GPT and Google Gemini frontier models, but independent verification will take a few weeks.
Should I migrate from Opus 4.8 to Fable 5?
For most workloads that have been bumping into Opus 4.8 capability ceilings, yes, at least for evaluation through June 22 while the free-on-subscription window applies. For workloads that don’t need Mythos-class capability (most routine Claude usage), Opus 4.8 remains the right choice at half the price. For workloads heavy in the gated domains (cybersecurity research, biology, chemistry), Fable 5 won’t add value because the fallback to Opus 4.8 kicks in. For latency-sensitive or high-volume workloads where the 2x price matters, model the cost story carefully before committing.
What does Fable 5 mean for Anthropic’s broader strategy?
Fable 5 lands in a particularly loaded news cycle for Anthropic. A week before release, the company published a plea urging major global AI labs to establish coordinated brakes on frontier-AI development, including warnings about recursive self-improvement. Multiple reports indicate Anthropic is preparing for IPO. Fable 5 is the demonstration that Anthropic can ship its most capable model class publicly with safeguards that survive 1,000+ hours of red-teaming; that demonstration matters for both the safety positioning and the IPO narrative. The Mythos-class lineup is now Anthropic’s most visible competitive differentiation against OpenAI’s GPT and Google’s Gemini frontier models.
