Editorial note: ChatForest does not have access to Claude Fable 5 and has not tested it. This article is based on Anthropic’s official announcement, the Claude API documentation, benchmark data from third-party evaluators, and published reporting from TechCrunch, CNBC, VentureBeat, and others. We do not claim hands-on testing.
At a glance: Claude Fable 5. Launched June 9, 2026. API model ID: claude-fable-5. $10 per million input tokens, $50 per million output tokens. 1M token context window. Available on Claude API, Claude Platform on AWS, Amazon Bedrock, Vertex AI, and Microsoft Foundry. Part of our AI Tools & Companies reviews.
On April 7, 2026, Anthropic launched Claude Mythos Preview — then announced it would not make the model publicly available. The reason: Mythos had autonomously discovered thousands of zero-day vulnerabilities across every major operating system and browser, found 271 security flaws in Firefox, and demonstrated expert-level autonomous cyberattack capability. Anthropic said at the time: “We do not plan to make Mythos Preview generally available."
Sixty-three days later, Anthropic launched Claude Fable 5.
Fable 5 is built on the same underlying model as Mythos. It is the Mythos-class capability that Anthropic spent two months figuring out how to release safely. The answer was three safety classifiers. Here is what changed — and what didn’t.
What Fable 5 Is
Claude Fable 5 is Anthropic’s most capable publicly available model. Anthropic describes it as built for “the most demanding reasoning and long-horizon agentic work." It shares its architecture with Claude Mythos 5 (the successor to Mythos Preview, available through Project Glasswing). The difference between Fable 5 and Mythos 5 is not capability — they are the same underlying model. The difference is three safety classifiers applied to Fable 5 that restrict certain categories of requests.
The practical result: most users will never encounter a classifier. Anthropic reports that its safety classifiers, combined, trigger in fewer than 5% of sessions on average, and that they are designed conservatively enough that some harmless requests still get caught.
Specifications
Specs below are as published in Anthropic’s own documentation.
| Specification | Claude Fable 5 |
|---|---|
| API model ID | claude-fable-5 |
| Context window | 1M tokens (default) |
| Max output tokens | 128k per request |
| Input pricing | $10 per million tokens |
| Output pricing | $50 per million tokens |
| Data retention | 30 days (no zero-data-retention option) |
| Thinking mode | Adaptive thinking (always on) |
| Raw thinking returned | No |
Pricing is less than half what Anthropic charged for Claude Mythos Preview access. The 1M-token context window matches Claude Mythos Preview.
Platforms at launch:
- Claude API (direct)
- Claude Platform on AWS
- Amazon Bedrock
- Google Cloud Vertex AI
- Microsoft Foundry
Subscription availability: Free on Pro, Max, Team, and seat-based Enterprise plans from June 9–22, 2026; usage credits required after June 22. “When sufficient capacity allows us to do so—we aim to restore Fable 5 as a standard part of subscription plans," per Anthropic.
What It Can Do
Software Engineering
The headline number: 80.3% on SWE-bench Pro — a benchmark measuring autonomous performance on real-world software engineering tasks. Anthropic’s own figure is produced with its own agent scaffolding rather than a neutral evaluation harness, so treat it as a vendor-reported result rather than an independently audited one.
For comparison: GPT-5.5 scores 58.6% and Gemini 3.1 Pro scores 54.2% on the same benchmark, per Anthropic’s release. Fable 5’s lead is more than 21 percentage points ahead of GPT-5.5 on this benchmark.
The Stripe case study from Anthropic’s launch materials: “In a 50-million-line Ruby codebase, the model performed a codebase-wide migration in a day that would otherwise have taken a whole team over two months by hand.” In Stripe’s own words, Fable 5 “compressed months of engineering into days.”
Vision
Fable 5 beat Pokémon FireRed using a minimal, vision-only harness. This is not a programming benchmark. It requires sustained visual reasoning, planning, and adaptation across a multi-hour autonomous task.
Finance and Enterprise Knowledge Work
On Hebbia’s Finance Benchmark, Fable 5 has the highest score of any model; on IMC’s trading-analysis evaluations, Fable 5 “aced” the evaluations “nearly across the board," per Anthropic’s launch materials.
Enterprise knowledge work benchmark (GDPval-AA): Fable 5 scores 1,932 versus GPT-5.5 at 1,769 and Gemini 3.1 Pro at 1,314.
Scientific Research
In life sciences testing, Claude Mythos 5 (same underlying model, no classifiers) conducted autonomous genomics research; its custom ML solution “outperformed a recent model published in the journal Science—despite being 100 times smaller," per Anthropic. In internal drug-design testing, Anthropic reports the model accelerated aspects of the drug design process by around 10 times, and that 9 of the 14 protein targets in the study yielded strong candidates for drug design.
Long-Context and Persistent Memory
In persistent memory testing using Slay the Spire as an evaluation platform, Anthropic reports Fable 5’s persistent file-based memory improved its performance three times more than for Opus 4.8. The model maintains focus and coherent state across millions of tokens.
The Full Benchmark Table
Figures below are as published by Anthropic; independent evaluators have noted the SWE-bench Pro figure uses Anthropic’s own scaffolding rather than a neutral harness, so treat these as vendor-reported rather than independently audited. (Anthropic’s GPQA Diamond comparison for Fable 5 could not be confirmed against a readable primary source and third-party reporting on it conflicts, so it is omitted here.)
| Benchmark | Claude Fable 5 | GPT-5.5 | Gemini 3.1 Pro |
|---|---|---|---|
| SWE-bench Pro | 80.3% | 58.6% | 54.2% |
| GDPval-AA (enterprise knowledge) | 1,932 | 1,769 | 1,314 |
| GDPpdf (document reasoning, no tools) | 29.8% | 24.9% | 16.7% |
The Three Safety Classifiers
The central engineering decision on Fable 5: apply safety classifiers that block specific request categories before generating output. When a classifier triggers, the API returns HTTP 200 with stop_reason: "refusal" — not an error — and reports which classifier declined the request.
Classifier 1: Cybersecurity Blocks vulnerability-discovery and offensive cyber tasks, preventing Fable from making progress on exploitation and broader cyberattack activities. Anthropic subjected Fable 5 to more than 1,000 hours of external red-teaming and found no universal jailbreaks. The cybersecurity classifier is the primary reason Mythos-class capability can be publicly released — without it, Fable 5 would have the autonomous vulnerability discovery capability that made Mythos Preview a restricted model.
Classifier 2: Biology and Chemistry Falls back to Opus 4.8 on most biology and chemistry requests, due to dual-use concerns about gene therapy design and pathogenic capabilities. Anthropic reports its safety classifiers overall — cybersecurity, biology/chemistry, and distillation combined — trigger in fewer than 5% of sessions on average, and has said the classifiers are “still stricter than would be ideal — for example, sometimes benign requests will trigger," with a stated aim to reduce false positives over time.
Classifier 3: Distillation Blocks attempts to extract Claude’s capabilities for training competing models. This addresses capability proliferation — the risk that a sufficiently capable model can be used to train other capable models outside sanctioned processes.
Requests refused before any output is generated incur no billing charges. If you retry on a fallback model, Anthropic’s fallback credit system refunds the prompt-cache cost so you are not charged twice for the same prompt.
Fallback and API Integration
For developers integrating Fable 5, the primary change from prior Claude models is handling refusals. Three approaches:
- Server-side fallback (beta): Pass the
fallbacksparameter; the API retries on another Claude model automatically. - Client-side fallback: Use SDK middleware (TypeScript, Python, Go, Java, C#) to retry on the client.
- Manual fallback: Build the retry logic yourself, on any platform.
Anthropic’s documentation notes that a refused request on Fable 5 “can usually be served by another Claude model” — the most common fallback target is Claude Opus 4.8.
Thinking mode: Adaptive thinking is always on for Fable 5 and Mythos 5. The thinking: {"type": "disabled"} parameter is not supported. Raw chain-of-thought content is never returned; use thinking.display: "summarized" to receive a readable reasoning summary, or "omitted" (the default) for an empty thinking block. Use the effort parameter to control thinking depth and cost.
Data retention: 30-day retention for all Fable 5 and Mythos 5 traffic. Zero data retention is not available for these models. Both are designated Covered Models with human access logging.
Claude Mythos 5 — The Classifier-Free Version
Claude Mythos 5 (claude-mythos-5) is the same underlying model as Fable 5 with the safety classifiers removed in certain areas. It remains restricted to approved customers through Project Glasswing — Anthropic’s collaboration with critical-infrastructure operators, open-source maintainers, and the US government.
Mythos 5 replaces Claude Mythos Preview in the Glasswing program. In life sciences testing, Mythos 5 conducted autonomous genomics research that produced models outperforming a recent Science journal publication. Anthropic has separately deployed Mythos 5 to a small group of cyberdefenders and infrastructure providers.
Customers without Glasswing access should use Fable 5. Anthropic’s documentation states it “offers the same capabilities” — meaning Fable 5 reaches the same capability ceiling, but with classifiers that can interrupt specific request types.
Context: From “Too Dangerous to Release” to General Availability
Claude Mythos Preview was announced April 7, 2026, as a model Anthropic would not make publicly available. The core concern was cybersecurity: a model capable of autonomously discovering and exploiting vulnerabilities at scale was deemed too dangerous for open access, even if that same capability was valuable to defenders.
Fable 5 represents Anthropic’s answer to the question that Mythos Preview posed: how do you release the most capable model you’ve ever built while limiting the most dangerous applications?
The answer is classifier-based gating. The cybersecurity classifier is the technical mechanism that separates Fable 5 from Mythos Preview’s unrestricted capability. Whether classifier-based controls are a durable solution — or a temporary constraint until someone develops reliable jailbreaks — is the live question in the AI safety community.
What has changed is the premise: in April, Anthropic’s public position was that Mythos-class capability should not be publicly accessible at all. In June, Anthropic’s position is that Mythos-class capability can be publicly accessible with the right technical controls. That is a significant policy shift, wrapped in a model launch.
Who Should Use Fable 5
Software engineering teams with complex, long-running tasks — the SWE-bench Pro gap versus competitors is the largest performance differential on any major benchmark category, and the Stripe case study is a credible real-world data point.
Finance and enterprise knowledge work — Fable 5 leads on reasoning and document analysis benchmarks relevant to financial work.
Long-horizon agentic applications — the 1M-token context and improved persistent memory performance make Fable 5 the current leader for applications requiring sustained autonomous operation.
Researchers in biology, chemistry, or cybersecurity — expect classifier friction. Plan for fallback handling if your work touches any of the three restricted areas.
Users who were waiting for Mythos-class access — Fable 5 is now the most capable generally available Claude model, priced well below what Mythos Preview cost in restricted access programs.
ChatForest covers AI tools, models, and the ecosystem around them. This article is research-based; we have not tested Claude Fable 5 directly. For the technical integration details, Anthropic’s official documentation is the authoritative source.