Anthropic released Claude Fable 5 on Monday, calling it the most capable Claude model ever made generally available (Anthropic announcement). The release sits on top of the same underlying architecture as Claude Mythos 5 — the unrestricted internal variant — with three safety classifiers between the public and the most dangerous use cases.
The launch matters for two reasons. It is the first Mythos-class model to ship to the open API. And it is doing so at less than half the cost of the Mythos Preview pricing.
What shipped
Anthropic released two models on the same day:
- Claude Fable 5 — public, gated by safety classifiers, available on the API and Claude apps immediately.
- Claude Mythos 5 — identical underlying model with safeguards lifted, restricted to Project Glasswing partners and a forthcoming trusted-access program for biomedical research.
Pricing for Fable 5: $10 per million input tokens / $50 per million output tokens. That is roughly half the price of the Mythos Preview and brings Anthropic’s flagship pricing closer to the GPT-5 and Gemini 2.5 Pro tiers.
The headline capability gains, per Anthropic:
- Software engineering — Stripe reportedly “compressed months of engineering into days” using Fable 5 in production.
- Vision — state-of-the-art on complex visual tasks including code reconstruction directly from screenshots.
- Long-context reasoning — maintains focus across “millions of tokens.”
- Life sciences — Mythos 5 has produced novel drug design candidates and scientific hypotheses, the first time Anthropic has publicly claimed that capability for a Claude model.
The benchmark story
Anthropic published comparative benchmarks against the leading external models. Fable 5 takes the top score on two:
- Cognition’s FrontierCode — a hard coding-agent benchmark designed by the Devin team.
- Hebbia’s Finance Benchmark — long-horizon analysis on real financial filings.
The benchmark sheet matters less than what is missing from it. Anthropic did not publish a head-to-head on AIME 2025 or HumanEval, the two benchmarks where GPT-5 has been leading. Reading between the lines: Fable 5 leads on agentic coding and finance, GPT-5 leads on raw math and short-form code. That is a clean split, and it lines up with how each lab has been describing its model strengths for the last six months.
What the safety stack looks like
Three classifiers sit between Fable 5 and the riskiest requests. When triggered, they redirect the user to Claude Opus 4.8 instead — a less-capable but still useful fallback.
| Classifier | What it blocks | Fallback |
|---|---|---|
| Cybersecurity | Offensive hacking tasks, exploitation queries, custom malware authoring | Claude Opus 4.8 |
| Biology / Chemistry | Dangerous viral or biological research, pathogen enhancement | Claude Opus 4.8 |
| Distillation | Attempts to extract model weights or capabilities at scale | Claude Opus 4.8 |
Why it matters: Anthropic’s claim is that fewer than 5% of Fable 5 sessions trigger a fallback at all. The other 95% see the model unencumbered. That ratio is the central design bet — if the false-positive rate goes much higher in production, developers will rebuild on competitors.
The timing of this release is also notable. Anthropic published the model in the same week it warned publicly that frontier AI is “becoming too dangerous.” The internal logic: ship the most capable model with guardrails so that the unguarded version (Mythos 5) does not become the only path to that capability.
How it stacks up against GPT-5 and Gemini 2.5 Pro
| Spec | Claude Fable 5 | GPT-5 | Gemini 2.5 Pro |
|---|---|---|---|
| Input cost ($/1M tokens) | $10 | $12 | $7 |
| Output cost ($/1M tokens) | $50 | $48 | $30 |
| Best at (per vendor) | Agentic coding, finance | Math, short-form code | Multimodal, long context |
| Safety architecture | 3 classifiers + fallback | Spec-driven RLHF | Constitutional + filters |
| Context window | Multi-million tokens | 2M tokens | 2M tokens |
| Restricted variant | Mythos 5 (trusted partners) | None public | None public |
Why it matters: the cost gap between Fable 5 and Gemini 2.5 Pro is meaningful for high-volume inference workloads. Anthropic is making a deliberate trade — premium positioning, premium price — that bets on developers paying for the agentic coding and finance leadership specifically. For consumer-grade applications, Gemini still wins on cost.
Access tiers and rollout
The rollout is unusual.
- Claude API — Fable 5 is live now, billed per token.
- Pro, Max, Team, Enterprise subscriptions — included through June 22, then requires usage credits.
- Full platform availability — Anthropic says “when capacity permits.”
- Mythos 5 — restricted to Project Glasswing cybersecurity partners, with a separate biomedical-research access program coming.
- 30-day data retention for Mythos-class sessions, to support jailbreak detection. Anthropic says the retained data is not used for training.
The June 22 cutoff for free-with-subscription access is the lever. Anthropic is using the next two weeks to seed developer adoption, then converting that adoption to per-token revenue. Expect API usage on the Anthropic Console to spike for the next two weeks and then flatten.
What this means for developers and operators
If you ship a Claude-based product, three immediate moves:
- Run your hardest agentic eval against Fable 5. If you have a Devin-style workflow, FrontierCode-class tasks, or finance-document parsing, this is the model to test now.
- Audit your usage against the three classifiers. If your product touches security tooling, biology research, or model-extraction patterns, expect more fallbacks to Opus 4.8 than before. Plan for graceful degradation.
- Budget for the June 22 transition. Subscriptions cover Fable 5 through June 22. After that, your Pro / Team plan reverts to usage credits. Heavy users should model the per-token cost now.
For more on Anthropic’s roadmap, see our coverage of Claude Opus 4.8 and the AI economics debate.
What to watch next
- OpenAI’s response. GPT-5.5 has been positioned for a mid-2026 release. Whether OpenAI matches Fable 5’s pricing or undercuts will signal the shape of the next quarter’s competition.
- Project Glasswing disclosures. Anthropic has not named the Glasswing partners publicly. The first named partner — likely a major cybersecurity firm — will move the conversation.
- The biomedical-access program details. Removing biology safeguards for trusted researchers is a meaningful policy shift. Watch for the criteria and the partner list.
- Whether Fable 5 actually gets meaningfully cheaper. Anthropic has historically dropped pricing 3–6 months after a flagship launch. A price cut before September would be a strong signal that adoption is below internal forecast.
Claude Fable 5 is the first model where Anthropic has openly priced its safety architecture as the product differentiator, not the marketing wrapper. The three-classifier fallback is the actual moat. If it holds up under production load — and the 95% no-fallback figure stays true at scale — the rest of the frontier-model market will copy the architecture within a year.
FAQ
What is Claude Fable 5?
Anthropic’s most capable publicly available model, released June 9, 2026. It shares architecture with the internal-only Claude Mythos 5, with three safety classifiers between users and the riskiest queries.
How much does Claude Fable 5 cost?
$10 per million input tokens and $50 per million output tokens. That is roughly half the cost of the Mythos Preview pricing.
What is Claude Mythos 5?
The same underlying model as Fable 5, with safeguards lifted. Restricted to Project Glasswing cybersecurity partners and a forthcoming biomedical-research trusted-access program. Not available to the general public.
What does Fable 5 do better than previous Claude models?
Software engineering (Stripe reportedly compressed months of work into days), vision tasks like code reconstruction from screenshots, long-context reasoning across millions of tokens, and life-sciences research that produced novel drug-design candidates.
How does Fable 5 compare to GPT-5 and Gemini 2.5 Pro?
Fable 5 leads on agentic coding benchmarks (Cognition FrontierCode) and finance reasoning (Hebbia). GPT-5 still leads on math and short-form code. Gemini 2.5 Pro is meaningfully cheaper for high-volume inference.
What gets blocked by the safety classifiers?
Cybersecurity offensive tooling, dangerous biological or chemical research, and model-extraction attempts. Blocked requests get rerouted to Claude Opus 4.8 — a less capable but still useful fallback.
How often do users hit the classifiers?
Anthropic says fewer than 5% of Fable 5 sessions trigger a fallback. The other 95% see the model unencumbered.
When does subscription-included access end?
June 22, 2026. After that, Pro, Max, Team, and Enterprise subscribers need usage credits to access Fable 5.
Track frontier-model releases as they ship — subscribe to the TechDaily360 weekly briefing.
