Claude Fable 5.1 & Mythos 5.1: the same brain, two different doors

Good morning. Two days before OpenAI lit a fuse with GPT-6 Astra, Anthropic did something quieter and arguably smarter: it released one model under two names — Claude Fable 5.1 for everyone, and Claude Mythos 5.1 for a very short list of vetted researchers. Same weights. Different doors. The boring version of this story is a price cut on cached tokens. The real version involves protein binders, a 300-meter map of Venus, and the first major lab to publish a score where its own model got worse.

Anthropic calls Fable 5.1 and Mythos 5.1 “the world's most advanced models for coding and knowledge work.” They are a point release on June 2026's Fable 5 / Mythos 5 — not a from-scratch flagship — but the split between the two names is the whole news:

• Fable 5.1 is the general-access version. It ships with additional safeguards that block high-risk dual-use work in biology and cybersecurity, and it's available today on consumer plans (Pro, Max) and the API.
• Mythos 5.1 is the same model with those specific safeguards relaxed — accessible only through Anthropic's trusted-access programmes: the Life Sciences Verification Program, with a Cyber Verification Program for defensive work planned to follow. Direct access is invitation-only for vetted US companies and individuals.

One practical consequence hides in plain sight: Anthropic's Claude Security — the product that scans your codebase and suggests patches — now runs on Mythos 5.1 for every Enterprise customer, even those who'll never see the name on an invoice.

From Anthropic's system card, Fable 5.1 vs. the previous generation — Fable 5, Claude Opus 5, and OpenAI's GPT-5.6 Sol where comparable figures exist:

The ceiling worth remembering is CursorBench 3.2.0: 73.4% — the highest published result on Cursor's independently-run agentic-coding benchmark, ahead of Opus 5's 70.0% and GPT-5.6 Sol's 67.2%. And unlike most point releases, Fable 5.1 now leads Claude Opus 5 on every benchmark Anthropic published — reversing rows where Opus 5 used to win.

Anthropic's headline price didn't move — same $10/$50 per million tokens as Fable 5. The actual move is barely visible: prompt-cache reads dropped 75%, from $1.00 to $0.25 per million tokens. That single cut makes typical workloads roughly 25% cheaper, and highly agentic ones — where the same context gets re-read over and over across hours of tool calls — up to ~45% cheaper.

Why that matters more than a leaderboard: the frontier is increasingly measured in unattended hours, not chat turns. MongoDB's Ron Sanzone, an early-access partner, described building a production prototype in about three days — the model researched their services code and docs, produced a design, then ran for hours unattended with verification loops. Fable 5.1's pricing is a deliberate bet that the winning workload is “leave an agent running overnight and come back to finished work.” At the $1 cache-read rate, that bet didn't pencil out; at $0.25 it does.

Because Mythos 5.1 relaxes the bio and cyber safeguards, it did the science. The results are the most striking part of the whole release:

• Protein design: Mythos designed protein binders with affinities 10× higher than the best designs ever submitted to Adaptyv Bio's protein-design competitions — on three targets. Its viable-binder rate sat at ~50% across 12 targets, where 10–15% is the current norm.
• Computational biology: by writing custom GPU kernels and caching intermediate results, it sped up seven open-source deep-learning models by up to 2.5× with identical outputs — cutting estimated GPU costs on genome-wide analyses by 30–60%.
• Planetary science: Anthropic reprocessed decades-old NASA Magellan radar data of Venus into a ~300-meter resolution digital elevation model — resolving features down to 2–3 km instead of 10–20 km, with heights up to 25% more accurate. It's released under a Creative Commons license ahead of the NASA VERITAS and ESA EnVision missions.

Keep the billing straight: the announcement attributes the life-sciences results to Mythos 5.1. The Venus work and the terminal-bench gains are Fable-5.1-class work. That distinction matters because it's exactly what the “two doors” design is for — one model to build products, one to push science.

The most disarming disclosure is the one Anthropic could have easily buried — and didn't:

This candor is genuinely rare in the frontier-model market right now. OpenAI, the same week, published a model with benchmark rows lit like a launch party; Anthropic published a card that leads with the thing its own model got worse at. Which one is the better marker of trust is an interesting question to sit with.

Not a drop-in upgrade. Three breaking API changes and three new betas came with 5.1, and if you hand-build messages arrays, two of these will bite you on day one:

• Forced tool use now errors. tool_choice: {"type": "any"} or {"type": "tool", "name": "..."} returns a 400 invalid_request_error. Only auto and none remain supported. If you force tools, you'll need to migrate logic.
• Thinking blocks are append-only. Modifying anything before a Fable 5.1 thinking block (the system prompt, the tools array, or any earlier message) invalidates every thinking block after it and errors on the next request — enforced for accounts created on or after August 31, 2026. The “inject a per-request reminder, then delete it” trick no longer works.
• Three new betas: per-message effort (change reason effort mid-conversation without invalidating the prompt cache), plus two more shipping through the API.
• Enterprise Frontier Safeguards: a new enterprise programme where customer logs stay in the customer's own cloud — a direct answer to the biggest enterprise objection to frontier models.

Context matters here. Anthropic shipped Fable 5.1 / Mythos 5.1 on September 1. OpenAI shipped GPT-6 Astra on September 3 — at the exact same $10/$50 list price. This is the first time the two labs' flagships have been at price parity, and the market is treating it as a head-to-head on workload, not a single winner:

• Generalist reasoning, speed-then-correctness: Fable 5.1 leads the AA Intelligence Index (where Astra scores 61, behind Fable 5.1 and two other Anthropic models — data point from our Astra deep dive).
• Hardest end-to-end tasks, context, autonomy: Astra counters with a 1.05M context, Critical-tier cyber and 99.9% ARC-AGI-3 claims.
• Everyday work: both labs basically concede it — Sonnet 5, Opus 5, or GPT-5.x mid-tiers, not the flagships.

Friendly fire even leaked into official materials: OpenAI's Astra announcement quietly noted that Claude Fable 5 and 5.1 “refuse the majority of questions” on leading life-science evals — the flip side of Claude's tighter bio safeguards. When labs start jabbing each other over benchmark abstentions, that's how you know the frontier has become a positioning war.

Facts in this piece were compiled from Anthropic's official Fable 5.1 / Mythos 5.1 announcement and system card, Anthropic's Fable pricing page, DataCamp, ExplainX, AI Tools Review, SaaSCity, Codersera, ComputingForGeeks, and Metirai, all contemporaneous with the September 1–3, 2026 release window. Benchmarks are Anthropic's own self-reported figures unless noted. This is news analysis by an independent publication — not AI advice, and not investment advice.