June 2026 · 14 min read
Claude Fable 5 launched June 9, 2026. It is the first Mythos-class model available to the general public, and it arrives with benchmark numbers that make every previous Claude model look like a rough draft. But buried inside the system card is a layer of safeguards that most coverage skipped entirely.
Image Credit: Leonardo AI
What is Claude Fable 5?
Claude Fable 5 is Anthropic's most capable model ever released to the general public. It launched on June 9, 2026, alongside Claude Mythos 5, and it sits above the Opus class in Anthropic's model hierarchy. Before that date, the tier it belongs to was closed to everyone except a narrow group of vetted cybersecurity organizations.
Fable 5 shares its core with Claude Mythos 5. They run the same underlying weights. The difference between them is the safety layer that Anthropic built to make Fable broadly available. Anthropic's official announcement describes Fable 5 as the first Mythos-class model made safe for general use.
The name has a deliberate logic. "Fable" comes from the Latin fabula, meaning "that which is told." "Mythos" is the Greek equivalent, same meaning, different language. Anthropic chose the pair specifically: both words describe a story, and the names reflect the safeguards, not the capability. The model is the same. What separates the two names is who can access them.
Andrej Karpathy, one of the most widely-cited researchers in AI, described Fable 5 as "a major-version-bump-deserving step change forward" on launch day. Given that Karpathy tends to measure his words carefully around AI claims, that quote is worth noting.
On SWE-bench Pro, Fable 5 scores 80.3%. The next-best model on that benchmark sits at 69.2%. That 11-point gap is not a rounding error. It holds across coding, reasoning, vision, and knowledge work, not just one test category.
Through June 22, 2026, Fable 5 is included free on Pro, Max, Team, and Enterprise subscription plans. It is also available through the Claude API, Amazon Bedrock, Google Cloud Vertex AI, and Microsoft Foundry as of launch day. The scale of that infrastructure rollout is comparable to Starlink's global expansion of ground infrastructure in 2026: wide distribution from day one, with capacity constraints following.
Claude Mythos: the origin of Fable
The story of Claude Fable 5 starts in March 2026, when leaked blog post drafts revealed the existence of a model Anthropic called Mythos. The company confirmed it in April 2026 and launched Claude Mythos Preview through Project Glasswing, a joint initiative with the US government. Glasswing gave access to vetted cyberdefenders and critical infrastructure providers only.
The Claude Mythos large language model was built with one exceptional capability that made broad release impractical: it can identify and exploit software vulnerabilities at a scale and speed that no previous AI model could match. On ExploitBench, the leading cybersecurity benchmark, Claude Mythos 5 scores 78%. Claude Opus 4.8, Anthropic's previous best public model, scores 40% on the same test. TechCrunch noted that Anthropic released Fable 5 just days after its own public warnings about AI getting too dangerous, a timing worth remembering.
Releasing that capability publicly without controls would mean giving anyone with an internet connection one of the most powerful software attack tools ever built. Anthropic's decision to hold it back made sense. Their stated goal was always eventual deployment at scale. Project Glasswing was the intermediate step.
By last week, Glasswing had expanded from its initial small group to hundreds of organizations across 15 countries. The focus stayed on organizations managing critical software infrastructure: security firms, government agencies, and infrastructure providers who use the model defensively. Early results from Glasswing, published by Anthropic as a research update, showed Mythos-class models helping defenders secure critically important software systems. The energy and data center demands that models like this create are a separate story entirely, one we covered in depth in our guide to AI data center water consumption and what it actually costs to run frontier AI.
Anthropic Claude Mythos also showed early promise in life sciences. With protein design and bioinformatics tools, the model matched or exceeded skilled human operators on drug design tasks. More on that below.
Claude Fable 5 is the public deployment of the same technology, with the sensitive capabilities locked behind safety classifiers. CNBC described it as Anthropic releasing a Mythos-like AI model to the public. Anthropic frames it as honoring their stated goal of bringing Mythos-class models to as many users as possible, as quickly and as safely as possible.
Fable 5 vs Mythos 5: same model, different gates
Both Fable 5 and Mythos 5 run the same underlying model. Their pricing is identical: $10 per million input tokens and $50 per million output tokens. Their API model IDs are claude-fable-5 and claude-mythos-5.
The classifiers separate them.
Fable 5 includes safety classifiers covering cybersecurity, biology, chemistry, and model distillation. When a request touches one of those areas, the classifier fires, and the API returns a response from Claude Opus 4.8 instead. The user is notified when this happens. Anthropic tuned these classifiers conservatively: they will catch some harmless requests, but the fallback triggers in fewer than 5% of sessions on average.
There is also a fourth safeguard category that works differently and receives almost no attention in standard coverage. For requests related to frontier LLM development, including building pretraining pipelines, distributed training infrastructure, or ML accelerator design, Fable 5 does not fall back to Opus 4.8. There is no notification. Instead, Anthropic limits the model's effectiveness through methods including prompt modification, steering vectors, or parameter-efficient fine-tuning (PEFT). The model still responds, but with deliberately degraded output. Anthropic estimates that this affects approximately 0.03% of traffic, concentrated in fewer than 0.1% of organizations.
Mythos 5 has some of those classifiers removed. It carries unrestricted cybersecurity capability, available only to Project Glasswing partners and a trusted-access program that Anthropic plans to expand in the coming months.
Both models support a 1 million token context window by default, with up to 128,000 output tokens per request. Both are classified as Covered Models, which means they carry 30-day data retention requirements and are not available under zero data retention.
For the vast majority of developers, Fable 5 is the model that matters. Mythos 5 access requires a Glasswing application, and the use cases that need it are narrow.
Claude Fable 5 benchmarks
The Claude Fable 5 benchmark numbers are consistent across every major evaluation. They do not excel in one area while trailing in another. They lead across the board.
On SWE-bench Pro, the gap over the competition is 11 points. Claude Opus 4.8 scores 69.2%. GPT-5.5 reaches 58.6%. Gemini 3.1 Pro sits at 54.2%. Digital Applied's independent benchmark breakdown confirmed these figures against Anthropic's official table, with the same pattern holding across knowledge work, spatial reasoning, and tool use.
On Cognition's FrontierCode Diamond set, which evaluates both task completion and production code quality standards, Fable 5 scores 29.3% versus 13.4% for Opus 4.8. That is more than a 2x lead in relative terms on the harder benchmark. At roughly the same per-task cost, Fable 5 at high effort scores around 24%, where Opus 4.8 peaks at 13.4% on frontier-difficulty work.
Cursor ran their internal benchmark (CursorBench) at max effort and gave Fable 5 a score of 72.9%. CEO Michael Truell called it state-of-the-art on that benchmark and said the model opened up a class of long-horizon problems that were out of reach for earlier models.
Third-party aggregators confirm the picture. BenchLM ranks Claude Fable 5 at #2 out of 123 models on their provisional leaderboard, with an overall score of 96/100. LMMarketCap gives it rank #1 out of 320 models specifically for coding. These are not Anthropic numbers; they come from independent tracking platforms.
Knowledge work benchmarks hold the same pattern. Fable 5 leads the Hebbia Finance Benchmark for senior-level analytical reasoning, with the largest gains in document-based reasoning, chart interpretation, and problem-solving. IMC's internal trading-analysis evaluations gave Fable 5 near-perfect marks across factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis.
One early partner reported that Fable 5 completed a frontier physics research task in 36 hours using one-third the reasoning tokens it took GPT-5.5 four days to match. At $10/M input versus $5/M, if you are using 3x fewer tokens, Fable 5's effective cost on that class of task is lower. This efficiency advantage applies to complex multi-step reasoning, not to simple or short-context queries.
On OSWorld-Verified (computer use), Fable 5 scores 85.0%, with Claude Mythos Preview edging it at 85.4%. That is one row where the new model does not lead outright, though the difference is within measurement noise.
What can Fable 5 actually do?
Software engineering
During early testing, Stripe ran Claude Fable 5 against a 50-million-line Ruby codebase. The task: a complete codebase-wide migration. Fable 5 finished it in a day. The same migration, done manually by a full engineering team, would have taken over two months.
Stripe's codebase ranks among the most complex in production software. A meaningful migration at that scale in 24 hours tests sustained performance, error recovery, and architectural reasoning across an enormous surface area.
Cursor noted that apps requiring a hundred prompts to build a year ago can now be completed in one. GitHub's Chief Product Officer called Fable 5 a step forward for complex, long-horizon coding tasks with a level of autonomy and reliability that exceeded previous benchmarks. Replit said the model understands what builders mean, not just what they type.
On token efficiency, Fable 5 scores highest among frontier models on FrontierCode even at medium effort. That means top-tier output without necessarily running the model at maximum compute.
Knowledge work and analysis
Fable 5 leads the Hebbia Finance Benchmark, which tests senior-analyst-level reasoning across documents, charts, and tables. The gains over prior models are largest in document-based reasoning and chart interpretation, according to Hebbia's testing.
IMC, the trading firm, reported that Fable 5 exceeded their evaluations across the board: factual lookup, conceptual reasoning, root-cause analysis, and expected-value analysis. Those are the core skills of a quantitative analyst.
Harvey, the AI legal platform, ran a blind review where lawyers compared Fable 5's document redlines against their current model. Fable 5 matched or beat the existing model every time. The lawyers did not know which model produced which output.
Vision
Fable 5 now leads all frontier models for vision tasks. It extracts precise numbers from dense scientific figures, reads complex charts accurately, and rebuilds web app source code from screenshots alone, with no access to the original files.
The most widely shared demonstration involved Pokémon FireRed. Previous Claude models could only complete the game with a helper harness providing maps, navigation aids, and game-state data. Fable 5 beat FireRed using only raw screenshots. No harness, no extra context, vision alone.
For developers building agents that interact with screens, dashboards, or document images: Fable 5 needs less scaffolding than previous models when working with visual interfaces.
Memory and long-context performance
Anthropic tested long-context performance with Slay the Spire, a deck-building game that demands sustained strategic planning across many turns. Giving Fable 5 access to persistent file-based memory improved its performance 3 times more than the same setup improved Opus 4.8. Fable 5 also reached the game's final act 3 times more often.
For production work, this means Fable 5 holds coherence across long documents, extended coding sessions, and multi-step research tasks better than any Claude model before it. The 1M token context window is the default, with 128,000 tokens available per output request.
Scientific research
Most of the science results come from Mythos 5 specifically, since those capabilities live in the restricted tier. They show what the underlying model can do.
Anthropic's internal protein design experts used Mythos 5 to accelerate aspects of drug design by around 10 times. With protein design and bioinformatics tools, the model matched or beat skilled human operators with no human assistance: choosing binding sites, selecting and running design tools, and recovering from failures automatically. Nine of 14 protein targets from one study yielded strong drug design candidates, covering immune checkpoints, growth-factor signaling, neurodegeneration, and muscle disease.
Mythos 5 produced molecular biology hypotheses consistently enough that Anthropic scientists advanced several to experimental evaluation. One hypothesis about an E. coli protein mechanism was independently confirmed by a separate lab working on the same problem at the same time.
On novel genomics research, Mythos 5 conducted over a week of largely autonomous work, assembling single-cell data for millions of cells across 138 animal species and training a custom model to identify cells performing the same biological role across distantly related organisms. The trained model outperformed a recent model published in Science, at 100 times smaller size. Anthropic intends to publish those results.
Claude Code and Fable 5
Claude Code is Anthropic's command-line tool for agentic coding. It runs in the terminal, accepts codebases directly, and executes multi-step tasks autonomously over extended sessions. Developers use it for large-scale refactoring, codebase migrations, debugging complex logic errors, and long-horizon feature work.
With Fable 5, Claude Code becomes substantially more capable. The model's 95% score on SWE-bench Verified and its lead on FrontierCode translate directly into the agentic setting: better planning, more reliable execution, faster error recovery, and sustained coherence across long sessions.
The Stripe migration result tells this story concretely. A 50-million-line codebase, completed in a day. That is the kind of task Claude Code handles, and Fable 5 is the engine that makes it practical at that scale.
Claude Code is available on Pro, Max, Team, and Enterprise subscription plans. Through June 22, using Fable 5 with Claude Code counts as 2x usage on subscription limits but costs nothing extra. To use Fable 5 as your model in Claude Code, set the model string to claude-fable-5.
Anthropic recommends configuring the fallback parameter in your API calls if you are using Claude Code for production workflows. If Fable 5's classifiers trigger on a request, the API retries automatically with Opus 4.8. Your pipeline keeps moving without manual intervention on those edge cases.
Cursor, GitHub, and Replit all gave Fable 5 strong marks in early testing. For developers who already use Claude Code in production, switching to it claude-fable-5 is a one-line change. The free window runs through June 22.
When the classifiers will fire on you (and what to do about it)
Every article on Fable 5's launch mentions the 5% figure. Almost none of them explain what actually triggers it, what the response looks like in practice, or how to build around it. Developers will find this out the hard way in production if they do not prepare.
The classifiers cover three publicly visible domains: cybersecurity, biology and chemistry, and model distillation. Within cybersecurity, the design is explicit. A request to "find vulnerabilities in this authentication code" reads like offensive cybersecurity to the classifier, even when the intent is defensive. A bioinformatics pipeline prompt that references protein sequences may trigger the biology classifier. A prompt asking Fable 5 to explain its own internal architecture in detail can trigger the distillation classifier. None of these requires malicious intent. The classifiers are, by Anthropic's own admission, conservatively tuned.
The behavior differs depending on which domain triggers. For cybersecurity, biology/chemistry, and distillation, Fable 5 falls back to Opus 4.8 and notifies the user. For frontier LLM development requests, there is no fallback and no notification. The model's output is silently degraded through methods including prompt modification, steering vectors, or PEFT. You get a response. It may just be a worse one. Anthropic estimates that this affects 0.03% of traffic.
In the API, the visible refusal returns HTTP 200 with stop_reason: "refusal", not an error code. The response body identifies which classifier fired. Standard SDK error handlers will miss this silently if your integration treats every HTTP 200 as a successful completion. Building explicit handling for the refusal shape is not optional in production.
To avoid broken pipelines, configure the fallbacks parameter in your API calls. This instructs the API to retry automatically with Opus 4.8 when Fable 5 refuses. Server-side fallback is currently in beta on the Claude API and Claude Platform on AWS. Client-side fallback via SDK middleware is available for TypeScript, Python, Go, Java, and C#.
On billing: if Fable 5 refuses a request before generating any output, you are not charged. When the system retries on another model, fallback credit refunds the prompt-cache cost of switching models mid-request.
The 5% figure sounds small until a CI pipeline runs 2,000 requests overnight and 100 of them silently fall back to Opus 4.8 without any log entry showing why. Logging classifier-triggered refusals from the start lets you audit which prompt categories are consistently hitting the classifier and reformulate them before they become a production incident.
Where Fable 5 underperforms, and which tasks still belong to smaller models
Launch coverage is promotional by nature. Nobody publishes "here is where the new flagship loses" on release day. A technically honest breakdown of Fable 5's weak spots matters for anyone making architecture decisions.
Latency is real. Fable 5 is slower than Sonnet 4.6 and Haiku 4.5. For real-time user-facing applications, including chatbots, live completions, and interactive tools, the latency cost is measurable. Time-to-first-token at the frontier tier is not the same as at the speed-optimized tiers, and no benchmark score changes that.
Computer use has a narrow gap. On OSWorld-Verified, Fable 5 scores 85.0% versus Mythos Preview's 85.4%. The reason is worth understanding: classifiers sometimes intercept screen-reading workflows that touch security-adjacent interfaces, such as terminals, code editors with debugging views, or browser developer tools. The gap is small, but it exists for a specific architectural reason.
Routing all tasks to Fable 5 is economically indefensible for simple work. At $10/M input versus $3/M for Sonnet 4.6, a "summarize this email" task running on Fable 5 is pure cost waste. The per-token premium only makes sense when task complexity justifies it.
Zero data retention requirements lock you out entirely. Fable 5 is a Covered Model. It carries a 30-day data retention requirement with no exception. If your compliance environment prohibits that, Fable 5 is unavailable regardless of how capable it is.
High-volume document processing still belongs to Haiku 4.5. For classification, tagging, extraction, or any repetitive operation at scale, batch processing helps, but the token cost still makes Haiku 4.5 the correct choice. Routing those tasks to Fable 5 multiplies costs without proportional quality gain on well-defined, repetitive work.
| Task type | Latency tolerance | ZDR required? | Recommended model |
|---|---|---|---|
| Long-horizon coding, complex migrations | High | No | Fable 5 |
| Deep document analysis, dense reasoning | High | No | Fable 5 |
| General assistant, daily tasks | Medium | Either | Sonnet 4.6 |
| Real-time chat, live completions | Low | Either | Sonnet 4.6 or Haiku 4.5 |
| High-volume classification, extraction | Medium | Either | Haiku 4.5 |
| Any task with a ZDR compliance requirement | Any | Yes | Opus 4.8 or Sonnet 4.6 |
What the same underlying weights actually mean in practice
Most articles repeat the "same model, different gates" framing without examining what it means operationally. The divergence between Fable 5 and Mythos 5 goes deeper than which requests get refused.
"Same weights" means the base model is identical. But inference-time classifiers modify behavior before and after token generation, not just as a post-processing step. This affects output distributions on borderline prompts, not only outright refusals. A prompt that sits on the edge of a restricted domain may produce different output from Fable 5 than from Mythos 5, even if no refusal fires.
On non-restricted tasks, Fable 5's benchmark scores should be essentially identical to Mythos 5's scores. The benchmark gaps you see on cybersecurity rows are by design. On coding, knowledge work, vision, and reasoning, the two models should perform the same. That is the correct technical claim, and it matters when making a procurement decision.
The distillation classifier deserves more attention than it receives. It targets prompts that appear designed to extract model weights, training data, or capability information from Fable 5 for use in training other models. Researchers working on fine-tuning pipelines, embedding models, or model evaluation frameworks may encounter this classifier even when their work has no competitive intent.
Long-context prompts touching restricted topics mid-document create an unusual classifier firing pattern. The classifier evaluates a rolling window across the conversation, not just the final prompt turn. A 200,000-token research document that includes a section on synthetic biology methods, even in a purely academic context, can trigger the biology classifier partway through processing.
Anthropic has not publicly disclosed the exact decision threshold for each classifier, whether thresholds differ by API tier, or whether enterprise plans can access adjusted classifier settings. Those are the unknowns that matter most for organizations building on Fable 5 at scale.
Myths and misreadings from the launch coverage
| What circulated | What is actually true | Why it matters |
|---|---|---|
| Fable 5 is a watered-down Mythos. | On every non-cybersecurity task, Fable 5 and Mythos 5 produce identical outputs. The classifiers do not degrade general capability. They block a narrow set of domains. | Treating Fable 5 as inferior for standard work leads to unnecessary Glasswing applications and delayed deployment. |
| ExploitBench 78% is a Fable 5 result. | The asterisked benchmark rows report Mythos 5 performance. Fable 5 scores 0% on offensive cybersecurity tasks by design. Multiple outlets cited the 78% as a Fable 5 figure without the asterisk. | If you are evaluating Fable 5 for security research, the 78% figure is not relevant to what you will receive. |
| The free trial runs until June 22. | Free usage applies to subscription plan limits through June 22. API access has been charged at standard rates from day one. High-volume API developers are not in the free window. | Cost projections built on the "free through June 22" reading will be wrong for API-heavy workflows. |
| Glasswing is a PR initiative. | Glasswing is operationally active, covering hundreds of organizations across 15 countries. Anthropic published a research update showing measurable cyberdefense outcomes from the program. | The program's real scope is relevant to anyone assessing whether Mythos 5 access is worth applying for. |
| Fable 5 will replace Opus 4.8. | Anthropic explicitly positioned Opus 4.8 as the fallback model for Fable 5 refusals. Opus 4.8 remains in the active lineup with a defined role, not on a sunset path. | Teams migrating entirely away from Opus 4.8 lose the fallback infrastructure that Anthropic designed for the system. |
| The safeguards are fully transparent. | Three classifier categories notify the user when they fire. The frontier LLM development safeguard does not. It silently degrades output through PEFT or steering vectors with no notification and no fallback to Opus 4.8. | AI developers building on Fable 5 for ML tooling may receive degraded responses without any indication that a safeguard is active. |
Claude Fable 5 pricing
Claude Fable 5 pricing and Claude Mythos 5 pricing are identical.
| Pricing tier | Input | Output |
|---|---|---|
| Standard | $10 per million tokens | $50 per million tokens |
| Batch | $5 per million tokens | $25 per million tokens |
| Cache read | $1 per million tokens | N/A |
| Cache write | $12.50 per million tokens | N/A |
For comparison: Claude Mythos Preview, the previous restricted version, cost approximately $30 per million input tokens and $150 per million output tokens. Fable 5 cuts that by more than 50%.
On claude.ai subscription plans, Fable 5 counts as 2x usage. It is included at no extra cost on Pro, Max, Team, and seat-based Enterprise plans from June 9 through June 22, 2026. From June 23, accessing Fable 5 requires usage credits. Anthropic says it plans to restore Fable 5 as a standard subscription feature as soon as capacity allows.
A practical cost estimate: analyzing a 10,000-line codebase at roughly 40,000 input tokens plus 10,000 output tokens runs approximately $0.90 at standard rates. On batch, that drops to $0.45 per run. For repeated tasks on large codebases, aggressive caching reduces costs further. Cache reads cost $1 per million tokens versus $10 for fresh input.
Token efficiency partially offsets the price premium. One early customer reported Fable 5 completing a frontier physics research task in 36 hours using one-third the reasoning tokens it took GPT-5.5 four days to match. On complex multi-step tasks, the effective cost per solved task can be lower than the sticker rate suggests.
Claude Fable 5 API details
The Claude Fable 5 API model string is claude-fable-5. You access it through Anthropic's standard Messages endpoint. TruFoundry's integration guide covers the full setup, including regional availability and SDK configuration.
A minimal API request:
{
"model": "claude-fable-5",
"max_tokens": 1024,
"messages": [
{
"role": "user",
"content": "Your prompt here"
}
]
}Fable 5 is generally available on all major cloud platforms as of June 9, 2026. AWS published its own deployment guide on launch day, covering regional availability and integration details:
- Anthropic Claude API (direct)
- Amazon Bedrock: US East (N. Virginia) and Europe (Stockholm) regions
- Claude Platform on AWS: North America, South America, Europe, Asia Pacific
- Google Cloud Vertex AI
- Microsoft Foundry
One behavior difference from previous Claude models: when Fable 5's classifiers decline a request, the API returns HTTP 200stop_reason: "refusal", not an error. The response body includes which classifier fired. Your integration needs to handle this response shape explicitly. It will not look like a normal error.
To avoid broken pipelines on refusals, configure the fallbacks parameter. This tells the API to retry automatically with Opus 4.8 if Fable 5 refuses. Server-side fallback is currently in beta on the Claude API and Claude Platform on AWS. Client-side fallback via SDK middleware is available for TypeScript, Python, Go, Java, and C#.
Billing on refusals: if Fable 5 refuses a request before generating any output, you are not charged. When the system retries on another model, fallback credit refunds the prompt-cache cost of switching models mid-request.
Fable 5 is a Covered Model, which means it carries 30-day data retention requirements. It is not available under zero data retention. If your use case requires stricter data handling, check Anthropic's data retention documentation before integrating. Teams running regulated workloads through third-party network infrastructure face a related layer of exposure that is easy to overlook; our breakdown of ISP traps and network device vulnerabilities covers the infrastructure side of that problem.
How the safety system works
Anthropic published a full system card alongside Fable 5 at launch, documenting the classifier design, alignment assessment methodology, and known false-positive rates. That level of transparency is relatively uncommon for a frontier model release.
The visible classifiers cover three categories: cybersecurity, biology and chemistry, and model distillation. Within cybersecurity, the design is explicit: on offensive cyber tasks (ExploitBench), Fable 5 makes 0% progress by design, while Mythos 5 scores 78% on the same evaluation. The gap between those two numbers is what the classifiers produce. TechTimes described the approach as a template that other frontier labs may be pushed to follow.
The fourth category is frontier LLM development. Unlike the other three, this one does not produce a visible refusal or notification. When Fable 5 detects a request targeting pretraining pipelines, distributed training infrastructure, or ML accelerator design, it limits its own effectiveness through prompt modification, steering vectors, or PEFT. The model still responds. Anthropicthat estimates this affects approximately 0.03% of traffic and expects minimal behavioral impact on requests that are not specifically targeting frontier LLM development. Nathan Lambert's analysis at Interconnects called this the most consequential part of the safety story, and the part most likely to affect legitimate users working in adjacent areas. Much like the invisible infrastructure layers that govern wireless communication, these classifiers operate below the level most users ever see, shaping what is possible without announcing themselves.
Anthropic describes all classifiers as "conservatively tuned." They will catch some harmless requests. The company's stated estimate is that fallback to Opus 4.8 triggers in fewer than 5% of sessions on average. External bug bounty testing ran for over 1,000 hours before launch and found no universal jailbreaks. The UK AI Safety Institute made partial progress in a short window during testing.
The alignment assessment found that Mythos 5, and by extension Fable 5, shows low levels of misaligned behavior, comparable to Claude Opus 4.8. For developers, the practical implication is straightforward: build fallback handling into your integration from day one. Configure the fallbacks parameter so your pipeline handles refusals gracefully. That is the entire integration difference compared to previous Claude models.
Image Credit: Leonardo AI
Project Glasswing and Anthropic Claude Mythos 5
Project Glasswing is Anthropic's restricted-access program for Mythos-class models. It started in April 2026 with a small group of cybersecurity organizations and critical infrastructure providers, expanded last week to hundreds of organizations across 15 countries, and now runs alongside the public launch of Fable 5. The pace of that institutional expansion is notable: from a handful of vetted partners to a multi-country program in under three months, a trust-building timeline that has few precedents outside of government procurement cycles.
Anthropic Claude Mythos 5 (model string: claude-mythos-5) is the upgrade to Claude Mythos Preview, deployed through Glasswing in collaboration with the US government. It carries the strongest cybersecurity capabilities of any model publicly acknowledged. Access requires a Glasswing application; contact your Anthropic, AWS, or Google Cloud account team for details. Anthropic's Glasswing page is at anthropic.com/glasswing.
Anthropic plans to expand Mythos 5 access through a broader trusted-access program beyond the current Glasswing framework. A biology trusted-access program is also planned, given Mythos 5's performance in drug design and molecular biology research.
For users without Glasswing access, Fable 5 is the closest available version. Anthropic explicitly positions it as the generally available Mythos-class model and the practical choice for anyone who does not qualify for or need the restricted tier.
Production architecture for Fable 5 at scale: routing, fallback, and cost control
Most teams will integrate Fable 5 as a drop-in replacement for Opus 4.8 and absorb twice the per-token cost without capturing twice the value. The teams that get real returns will build routing logic that matches task complexity to model capability. Here is what that looks like in practice.
Three-tier routing
The correct production architecture for most organizations is a three-tier model with a routing layer. Haiku 4.5 handles high-volume classification, tagging, and extraction. Sonnet 4.6 handles standard generalist tasks, real-time completions, and anything with a latency budget under one second. Fable 5 handles complex reasoning, long-horizon coding, dense document analysis, and vision-heavy workflows. The routing decision rests on four variables: token count, task type, latency budget, and data retention requirements. Only Fable 5 on long-horizon complex tasks justifies the $10/M input rate.
Cache strategy
At $12.50 per million tokens for cache writes and $1 per million for cache reads, the economics of a caching system prompt or large shared context blocks are favorable at any volume above roughly 50 requests per day hitting the same context. The break-even on a 10,000-token system prompt: a single cache write costs $0.125. Each subsequent read costs $0.01. After 14 requests using that cached context, caching has paid for itself versus paying for fresh input tokens at $10/M. At 100 daily requests against the same context, caching saves approximately $0.85 per day per endpoint.
Fallback handling
The correct pattern for handling classifier-triggered refusals in production: catch stop_reason: "refusal" Before any downstream processing, log the classifier field to a monitoring system, decide whether to retry with Opus 4.8 or surface the refusal to the user, and route accordingly. Treating a refusal as a normal completion causes silent downstream errors. In a document processing pipeline, this means a refusal-shaped response gets passed to the next processing step as if it were valid content, potentially corrupting downstream outputs.
Four metrics worth monitoring
The useful monitoring set for a Fable 5 integration is: token cost per successful completion (not per request, since refusals cost nothing), fallback rate by prompt category, cache hit rate, and time-to-first-token at P95 and P99. These four figures tell you whether the integration is running efficiently. A rising fallback rate in a specific prompt category is the signal to reformulate those prompts before the volume grows.
The organizational question that gets skipped
Who owns the fallback prompt review? When Fable 5 consistently refuses a category of prompts your team sends, someone needs to evaluate whether to reformulate the prompts, accept the Opus 4.8 fallback, or apply for Glasswing access. That is not a technical decision. It is a product and compliance decision that should be made before deployment, not after a production incident. Assigning ownership of classifier incident review before launch prevents the situation where a team dilaterovers weeks in that a product feature has been silently running on Opus 4.8 because no one noticed the refusal logs. Ownership gaps in high-stakes operational environments are rarely technical problems; they are structural ones, as visible in fields as different as elite sports management as in engineering organizations.
Where Fable 5 fits in the Anthropic model lineup
As of June 10, 2026, Anthropic's model hierarchy looks like this:
| Model | Tier | Access | Best for |
|---|---|---|---|
| Claude Mythos 5 | Mythos (unrestricted) | Project Glasswing only | Cybersecurity, defense, and advanced research |
| Claude Fable 5 | Mythos (with classifiers) | Generally available | Complex coding, research, long-horizon work |
| Claude Opus 4.8 | Opus | Generally available | Strong generalist; Fable 5 fallback model |
| Claude Sonnet 4.6 | Sonnet | Generally available | Fast, mid-tier performance |
| Claude Haiku 4.5 | Haiku | Generally available | High-speed, cost-efficient |
Fable 5 does not push the other models aside. For straightforward tasks, Sonnet 4.6 or Haiku 4.5 will be faster and cheaper per token. The decision to route work to Fable 5 should rest on task complexity: long-horizon coding, deep document analysis, dense scientific reasoning, or vision-heavy workflows are where Mythos-class capability earns its premium.
When Fable 5 refuses a request, the fallback is Opus 4.8, which is itself a strong generalist model. In fewer than 5% of sessions where the classifier fires, most users will not notice a significant quality difference on their specific task.
More capable models are coming in the next few months, per Anthropic's own communication. The company is widely expected to list publicly this year, and the release pace suggests Fable 5 is one step in a longer sequence. For now, it is the best publicly available model Anthropic has ever shipped.
Claude Fable 5 is a real capability step up from Opus 4.8. The benchmark lead over the field is large and consistent. The Stripe codebase result, the Pokémon FireRed vision test, the IMC trading evaluation, and the Harvey legal review all come from early-access customers running real workflows, not synthetic tests. The pricing is lower than the Claude Mythos Preview. The API uses a standard model string. The visible safety classifiers handle fewer than 5% of sessions, with transparent documentation and graceful fallback to Opus 4.8. The invisible safeguard affecting frontier AI development work affects 0.03% of traffic, but developers building ML tooling on Fable 5 should know it exists. The model string is claude-fable-5. If you build on the Claude API or use Claude Code in production, this week is the right time to test it.
USA Beam Take
The benchmark story is real. Claude Fable 5 leads on SWE-bench Pro by 11 points over the previous best, and the early customer results from Stripe, IMC, and Harvey are from actual production workflows, not controlled demos. That gap is significant and not easily dismissed.
The pricing decision is also worth acknowledging. Cutting the cost of Mythos-class inference by more than 50% compared to Mythos Preview is a meaningful move, and the token efficiency data on complex tasks suggests the effective cost differential over Opus 4.8 is narrower than the sticker rate implies.
The part of this launch that deserves more attention than it received is the fourth safeguard category: the silent capability degradation for frontier LLM development requests. Anthropic disclosed it in the system card. They did not hide it. But they also made a deliberate choice that this particular safeguard would produce no notification and no fallback to a different model. The user receives a response. The response may just be a worse one, with no indication that anything happened.
The stated rationale is that the actors most likely to misuse this capability are the ones most willing to violate terms of service, and transparent refusals are easier to route around than silent degradation. That is a defensible argument. It is also a policy choice where the cost falls on legitimate users who happen to work in areas adjacent to the restricted domain. Startups training custom ranking models, researchers fine-tuning small open-source models, and ML platform engineers working on infrastructure tooling all fall within the fuzzy boundary of "frontier LLM development," depending on how the classifier interprets the prompt.
Anthropic estimates 0.03% of traffic is affected. At Fable 5's scale of deployment, that is not a trivial number of requests. The company has pledged to improve precision over time, and its track record on narrowing classifiers post-launch is reasonable. But for now, AI developers building on Fable 5 for ML-adjacent work should treat the model's output as potentially degraded on that class of task, and should cross-validate results against Opus 4.8 when the outputs matter.
The launch is a genuine step forward. The safety architecture it introduced is more layered than anything Anthropic has deployed before. Both things can be true.