Mythos & Fable 5:
the complete research report.

a.Claude Fable 5 — the public one

The first Mythos-class model the public can use. Anthropic's words: "a Mythos-class model that has been made safe for general use."

Available immediately via API (claude-fable-5), in Claude apps, Claude Code, and Amazon Bedrock.

b.Claude Mythos 5 — the restricted twin

The identical underlying model with the safeguards lifted.

Only available to approved users: Project Glasswing cybersecurity partners and, soon, vetted biology researchers.

Anyone on the old Mythos Preview gets auto-upgraded to Mythos 5.

Software engineering

Plus: 72.9% on CursorBench at max effort — 8.6 points ahead of GPT-5.5.

Analytics, documents & vision

📈 First model ever past 90% on Hex's core analytics benchmark — a 10-point jump over Opus.

📄 OfficeQA Pro: 57.9% vs Opus 4.8's 48.1%.

👁️ GDP.pdf (hard document vision): 29.8% vs Opus 4.8 (22.5%), GPT-5.5 (24.9%), Gemini 3.1 Pro (16.7%). Anthropic calls it the new state of the art for vision.

Long context & agentic work

🧵 1M token context window, 128K max output.

🧵 GraphWalks BFS at 1M depth: 79.4 F1 vs Opus 4.8's 68.1.

🌐 BrowseComp multi-agent: 93.3%.

🛠️ Toolathlon: 61.7% in 19.8 average turns vs Opus 4.8's 59.9% in 24.5 turns — better results in fewer steps.

🔬 Matches or beats GPT-5.5 on frontier physics research using one-third the tokens.

💼 Real-World Finance v2: preferred in 74% of comparisons (Elo 1,374 vs 1,222) + top score on Hebbia's finance benchmark.

The access facts

🔑 API model ID: claude-fable-5 — live everywhere from day one.

📦 Batch pricing halves it: $5 in / $25 out — Opus money for Mythos-class output.

🎁 Included in Pro, Max, Team and Enterprise plans June 9–22 at no extra cost. From June 23: usage credits, pending capacity.

🎛️ New effort parameter: low, medium, high, and a new xhigh tier for the hardest jobs.

☁️ Available on Amazon Bedrock at launch.

🔒 All Mythos-class traffic: 30-day retention, never used for training, human access logged.

Three classifier categories trigger a fallback

1. Cybersecurity — offensive hacking, vulnerability exploitation, defense evasion.

2. Biology & chemistry — bioweapons-adjacent and dual-use research.

3. Distillation — attempts to extract the model's capabilities.

When triggered, the request is answered by Claude Opus 4.8 instead — and you're told it happened. Developers see stop_reason: "refusal" with a category, and can configure server-side fallback.

The honest caveats — from Anthropic itself

⚠️ The safeguards are "deliberately tuned to be cautious" and "stricter than would be ideal" — benign requests will sometimes trip them.

⚠️ The UK AI Safety Institute made "initial progress" toward a jailbreak within a brief testing window.

⚠️ Anthropic concedes it's "likely impossible to completely prevent universal jailbreaks."

⚠️ The alignment assessment found Mythos 5's misalignment levels "similar to that of Opus 4.8" — better tools, same fundamental challenges.

Who gets the unrestricted model

🏛️ Project Glasswing — Anthropic's collaboration with the US government: cyber-defenders and critical-infrastructure providers get Mythos 5 with the cyber safeguards lifted, immediately.

🌍 The program is expanding to roughly 150 organizations in more than fifteen countries.

🧬 Biomedical researchers get a trusted-access program "in the coming weeks" with bio/chem safeguards removed.

📈 A broader trusted-access program is promised later.

What the essay actually says

It maps three futures: capability gains plateau; AI keeps accelerating with humans steering; or AI achieves recursive self-improvement — "fully autonomously designing and developing its own successor."

That last one, they write, "is not inevitable" but "could come sooner than most institutions are prepared for."

The proposal: verification systems so frontier labs can prove to each other they've slowed down.

The stats buried in it

🤖 As of May 2026, over 80% of Anthropic's merged code is authored by Claude.

📈 Anthropic engineers ship roughly 8× more code per quarter than in 2024.

Their framing of why: "The doing... now costs almost nothing in human time, even if it still has costs in compute."

If you use Claude for work

Switch your hardest tasks to Fable 5 before June 22 while it's free on your plan. Test it on the jobs the old model fumbled — that's where the gap shows.

If you run agents or automations

Point one workflow at claude-fable-5 and watch turn counts drop. Keep Opus 4.8 as your fallback — that's literally Anthropic's own architecture inside Fable 5.

Use the effort dial: low for routine runs, xhigh for the jobs that used to fail.

If you're watching costs

Batch mode is the cheat code — $5/$25 is Opus pricing for Mythos-class output. For overnight or queued work, there's no reason to pay the live rate.

If you're thinking about the bigger picture

The 80%-of-code stat is the real headline. The frontier lab is already mostly AI-written. The question isn't whether agentic work arrives — it's whether your systems are ready to use it.

❓ Capacity. Anthropic expects high demand and staged the rollout — how constrained will access be after June 23?

❓ False-positive rates over time. The classifiers are "deliberately cautious" — will they loosen as data accrues?

❓ The trusted-access expansion. Who counts as "trusted" for bio research, and how fast does it widen?

❓ Competitive response. GPT-5.5 just got jumped on the hardest benchmarks — OpenAI and Google's counters are presumably imminent.

❓ Subscription economics. "Usage credits" after June 22 is vague — real-world costs for heavy users are TBD.