Anthropic launches Claude Fable 5 + Mythos 5 — first publicly available Mythos-class model, free in Pro/Max/Team through June 22

What was launched

The reporting from Anthropic’s official announcement, TechCrunch, The Hacker News, VentureBeat, MacRumors, NBC News, and Yahoo Finance confirms:

The model architecture

• Claude Fable 5 and Claude Mythos 5 are the same underlying model
• Fable 5 ships with safety classifiers active — high-risk queries fall back to Opus 4.8
• Mythos 5 lifts the safeguards for authorized users only (cybersecurity professionals now; biomedical researchers soon)

Pricing

• $10 per million input tokens
• $50 per million output tokens
• Less than half the rate of the prior Mythos Preview (restricted-access)
• Free on Pro, Max, Team, and seat-based Enterprise plans from June 9-22, 2026 (11 days)
• June 23 onward: usage requires credits

Safety classifier behavior (Fable 5)

The classifiers redirect requests involving:

• Cybersecurity: exploitation and offensive cyber tasks
• Biology/Chemistry: bioweapon-related queries; dual-use biological research
• Distillation: attempts to extract model capabilities

When the classifier triggers, the request is routed to Claude Opus 4.8 instead. Anthropic reports more than 95% of Fable 5 sessions avoid fallback entirely — meaning the safety surface affects a small minority of queries, while the vast majority of Fable 5’s higher capability is broadly available.

Benchmarks claimed

• CursorBench: state-of-the-art (per Michael Truell, CEO of Cursor: “Claude Fable 5 is the state of the art model on CursorBench. It’s opened up a class of long-horizon problems that were out of reach.”)
• Cognition’s FrontierBench: highest-scoring (per Scott Wu, CEO of Cognition: “Claude Fable 5 is the highest-scoring model on FrontierBench, Cognition’s frontier coding eval.”)
• Hebbia’s Finance Benchmark: top performance for senior-level reasoning
• Vision tasks: playing Pokémon FireRed with vision-only input (Anthropic-cited example)
• Protein design: ~10× acceleration for drug discovery tasks

Availability

• Claude API: immediately
• Pro / Max / Team / Enterprise plans: included at no cost June 9-22
• AWS Bedrock, Google Cloud Vertex AI, Microsoft Foundry: details not yet specified in the announcement

Why this matters

Three reads.

1. Anthropic is operationalizing the per-domain release-tier framework. The pattern that emerged through Project Glasswing (cybersecurity restricted-access) and OpenAI’s Rosalind Biodefense (life sciences restricted-access) is now formalized. Anthropic has the same model in two configurations: Fable 5 (95%+ of capability publicly available, with high-risk fallback) and Mythos 5 (full capability for vetted partners). This is structurally different from the binary “ship or don’t ship” choice that dominated 2024-2025 frontier-model release discussions.

For other labs, this becomes the reference architecture for shipping high-capability models responsibly: a public surface with classifier-based fallback, plus a restricted-partner surface for users with documented need for the high-risk capabilities.

2. Pricing dropped substantially. Mythos Preview pricing was effectively “talk to us” for the restricted-access partner program. Fable 5 at $10 input / $50 output per million tokens makes the model economically usable for production workloads — comparable to Claude Opus 4.8 ($5/$25) but at ~2× the price for substantially higher capability. For workloads where Fable 5’s frontier-tier reasoning materially improves outcomes (Cursor’s “class of long-horizon problems that were out of reach”), the 2× cost is rational.

3. The 11-day free window is unusual. Free trial periods at this scale of capability uplift are not standard for frontier models. By giving Pro/Max/Team subscribers free Fable 5 access through June 22, Anthropic is forcing comparison against ChatGPT Plus, Gemini Advanced, and Grok subscriptions during a window when Fable 5 is the strongest publicly accessible model on the market. Some fraction of those Pro/Max users were already exploring Apple’s Siri Extensions routing decisions — this gives them substantial reason to default to Claude.

What it means for Claude and Claude Code users

For Pro / Max / Team / Enterprise subscribers: Use Fable 5 between now and June 22. It’s free during that window. After June 23, usage requires credits. For demanding workloads (complex refactors, multi-step research, long-horizon agentic tasks), the capability uplift over Opus 4.8 is real per the Cursor and Cognition CEO quotes.

For Claude Code users: Fable 5 on the API is now the strongest publicly available model — Cursor’s CEO is quoted calling it state-of-the-art on CursorBench, and Cognition’s CEO calls it the highest-scoring model on their FrontierBench coding evaluation. For developers running complex agentic workflows (“class of long-horizon problems that were out of reach” per Truell), the upgrade is meaningful.

For API users: Compare $10/$50 per million Fable 5 tokens against Opus 4.8 at $5/$25. The price-doubling is justified when Fable 5’s capability advantage materially improves output quality on your specific workload. For routine work, Opus 4.8 remains the cost-efficient default.

For workloads in cyber-offensive, bioweapon-related, or distillation domains: Fable 5 will fall back to Opus 4.8. These workloads aren’t the right fit for Fable 5; for authorized cybersecurity professionals, Mythos 5 access is the partner-program path.

How this connects to recent Anthropic positioning

This is the 12th Anthropic structural signal in ~30 days, and the most important one capability-wise since the May 28 Series H + Opus 4.8 launch:

Capital + capability + safety credentialing + IPO process + secondary-market pricing + enterprise distribution + frontier release strategy — all moving in parallel.

The honest caveats

Three caveats:

The benchmark claims are CEO testimonials, not third-party verified. Truell and Wu both have commercial interest in promoting Fable 5 (Cursor and Cognition are major Claude consumers). Independent benchmark verification will surface over the coming 30-60 days. Treat the “state of the art on CursorBench / FrontierBench” claims as Anthropic-published / partner-endorsed pending independent confirmation.

The 95% no-fallback figure is internal evaluation. Real-world fallback rates may differ depending on workload composition. Users running heavy cybersecurity research, biology, or chemistry workloads will see higher fallback rates than the typical session.

Free-window dynamics may pressure Pro/Max usage caps. Anthropic offering free Fable 5 access for 11 days will substantially increase per-user inference demand. Pro and Max users may see capacity tightening during the free window if demand exceeds projections.

What it changes for Pick Right readers tomorrow

If you’re a Claude Pro, Max, Team, or Enterprise subscriber, use Fable 5 between now and June 22. After June 23, usage requires credits — so this is a real 11-day window to evaluate the strongest publicly available Claude model on your specific workloads.

If you’re a Claude Code user, try pointing Claude Code at Fable 5 for the longest-horizon, hardest tasks you’ve been deferring. The Cursor/Cognition CEO quotes specifically call out “class of long-horizon problems that were out of reach” — that framing isn’t typical of incremental model upgrades.

For broader context, see the Claude review, the Claude Code review, the Project Glasswing 10K-vulnerability milestone, the Series H + Opus 4.8 article, the OpenAI Rosalind Biodefense article, and the Anthropic $1T secondary-market valuation article for the broader Anthropic positioning and frontier-model release-tier framework.

Questions or corrections? Email Pick Right. Want the full list? See all news.