Claude Fable 5.1 and Mythos 5.1 Are the Same Model. That’s the Story.

Anthropic shipped two new frontier models yesterday: Claude Fable 5.1 and Claude Mythos 5.1. Same announcement page, same release date, same number after the name.

Plot twist: they're the same model.

Not "same family." Not "siblings." Identical weights. The difference between them is which set of rules gets applied to you when you use them. Fable 5.1 is generally available. Mythos 5.1 is for "vetted cyberdefenders and life scientists" — a trusted-access program that, right now, only US organizations can join, coordinated with the US government.

The HN thread hit ~1,000 points in an evening, and most of it wasn't about the benchmarks. It was people trading stories about how often Fable had thrown them out and demoted them to Opus. Which turns out to be the actual story of this release. Let me get into it.

The benchmarks are genuinely good, so let's do those first

Fable 5.1 is a big step up from Fable 5, and Anthropic posted honest-looking comparisons this time:

• Terminal-Bench-Science 0.1 (agentic scientific research): 52.6%, versus 29.0% for Claude Opus 5 and 22.4% for GPT-5.6 Sol. That's not an increment, that's a different animal.
• Terminal-Bench 4.0 (agentic coding): 55.8% for Fable 5.1, 60.9% for Mythos 5.1, versus Opus 5 at 52.3%.
• Humanity's Last Exam: 60.9% without tools, 65% with.
• OSWorld 2.0 (computer use): 41.7% strict, 77.9% partial — the partial/strict gap tells you it still fumbles the last mile of GUI work.
• AutomationBench (business workflows): 31.4%, up from 17.1% for Fable 5. Still barely a third. Business rules remain where agents go to die.

The anecdote I keep thinking about comes from Millennium, the investment firm. They had a crash hitting roughly one run in a million that nobody — human or model — could explain for years. Fable 5.1 disassembled an external vendor library, matched it against a core dump, and traced the bug into code nobody at Millennium had written. A senior MongoDB engineer describes the same shape of thing: it ran for hours unattended overnight with verification loops and finished the next phase of a prototype by morning.

That's the frontier-model pitch in 2026. Not "it writes your code faster." It's "it does the multi-day forensic archaeology nobody wants to pay a human for."

Fine. Cool model. Now the interesting part.

The benchmark gap between the two "models" is a safety tax, and Anthropic printed it on the label

Here's the sentence from the announcement that deserves way more attention than it got:

"Claude Fable 5.1 and Claude Mythos 5.1 are the same underlying model; the gap between them reflects the tasks on which our earlier, less precise cyber safeguards intervened."

Read that again. The 5-point gap on Terminal-Bench 4.0 between Fable 5.1 and Mythos 5.1 isn't capability. It's the classifier. When the safeguard layer decides your task looks "cyber," it intervenes — and per the eval footnotes, intervened tasks score zero, with the work silently routed to a fallback model instead (cyber tasks go to Opus 4.8, bio tasks to Opus 5).

Sit with that for a second. Your effective model quality now depends on what a classifier thinks you're doing. Ask a security engineer. Actually, you don't have to — the HN thread did it for you. One commenter who works in defensive cybersecurity: "I haven't been able to use it for anything even with being in the CVP. It downgrades to Opus every time." Another: "almost anything related to my job is flagged for 'cyber.'" The trusted-access program didn't save them; the false positives hit the exact people the program was built for.

This release is, in large part, an apology for that. The headline safety number: 60% fewer false positives in cybersecurity. The policy shift underneath it: Fable 5.1 can now discover software vulnerabilities — it just can't develop exploits for them. Which is the correct line, honestly. Finding the hole and weaponizing the hole are different acts, and it's good to see a lab draw it at "can find, can't weaponize" instead of the previous stance of flagging anyone who said the word "port scan."

The two-names thing is the part I find genuinely novel. It's one set of weights wearing two different muzzles. Mythos 5.1's "safeguards are specifically designed to support work in cybersecurity and the life sciences" — looser in the domains where the customers are vetted, tighter everywhere else. Fable is the cage everyone gets; Mythos is a wider cage behind a credentials check and, de facto, a government list.

Remember June? Commerce put export controls on Mythos 5 and Fable 5, then partially walked it back to "entities identified in Annex A." Whatever your politics, the precedent landed: frontier access is now a licensing question. Five months later, Anthropic has internalized the shape of the world and productized it. You don't buy a model anymore. You buy an access tier. The weights are the same; the permission slip is the product.

The price cut is the real headline, and it's about agents, not intelligence

Buried under the model news: cache reads now cost $0.25 per million tokens, down 75%. Net effect: about 25% cheaper than Fable 5 for typical workloads, and up to ~45% cheaper for "highly agentic" work. List pricing is unchanged at $10/M input and $50/M output — still not cheap, still roughly a mortgage for a heavy agent workload.

Why does the cache discount matter so much? Because a long-running agent's transcript is mostly it re-reading its own homework. Every tool call, every file it read an hour ago, every step of a plan — it all rides along as input context on the next call. For agentic work, cache reads are the cost base. Cut those by 75% and the economics of "run an agent for six hours" change shape. Anthropic's own numbers say an agent-heavy workload that cost 100 indexed points on Fable 5 costs about 55 on 5.1, with zero model changes on your side.

The market noticed instantly. Cognition announced they're moving Devin's Opus 5 traffic to Fable 5.1 on launch day, specifically citing the cache pricing. Dan Shipper at Every: roughly twice as fast as Opus 5 using half as many tokens.

The other economics lever is quieter: the effort ladder. Fable 5.1 runs at low, medium, high, xhigh, or max effort, and at low/medium it matches or beats Fable 5's high-effort results for a fraction of the cost. Defaults differ by surface — max effort is opt-in, Claude Code ships at high, Claude.ai and Cowork at medium. Translation: the same model is now sold to you at five price-performance points, and most users should actively choose a lower one. "How hard should the model think" has become a budget dial, like a GPU power limit.

Put the two together and the strategy is obvious: Anthropic isn't selling intelligence this week. It's selling cost-per-finished-task. The intelligence was table stakes; the release is about making the meter run slower.

One more thing for the enterprises: EFS

Enterprise Frontier Safeguards stores your data on infrastructure you control, not Anthropic's — full zero-retention privacy while keeping the adversarial-use protections. Rolling out in phases this fall; until then, eligible customers get plain zero data retention. Boring, genuinely useful, and clearly aimed at the banks and hospitals that loved the model and hated the data flow. That's all I'll say about it.

The retreat nobody announced

Now the strategic read. Anthropic fixed pricing, false positives, and data retention in one release. What they did not fix: the bottom of their lineup.

Haiku hasn't been updated in about a year. It's priced roughly 33% above Gemini 3.7 Flash and several multiples above GPT-5.6 Luna. The HN consensus, in one comment: "What's the point in paying them for Haiku-class models? You can run those on your own graphics card." People are migrating Haiku and Sonnet workloads to Grok, DeepSeek, and about twenty open-weights models that are good enough and nearly free.

And they're right to. This is the part of the market I've been writing about all summer — GLM-5.3 Flash doing Terminal-Bench heroics on cheap Chinese GPU time, Qwen3.8 shipping open weights that hold their own against closed flagships, DeepSeek V4 Flash undercutting everyone. The Haiku/Sonnet tier has collapsed into open source. There is no putting that genie back.

Anthropic's answer, visible in this release, is to stop fighting there. If the cheap tier is a commodity, let it be someone else's commodity. Own the expensive tier completely: the model that does the multi-day forensic bug hunt, wrapped in enterprise privacy, behind a safety classifier, at a price that only makes sense when the alternative is a $400K engineer. It's a coherent bet. It might even be right.

But notice what you gave up in the trade. The model that never touches your task because a classifier guessed wrong. The fallback model quietly finishing your "cyber" request with last generation's weights. And a world where the same weights ship under different names with different rules, and the assignment between them is decided by a vetting process you can't see.

Same weights, different cages

That's the real template this release sets, and I'd bet money it spreads. Expect more "same model, multiple governance profiles" launches from every frontier lab by next year — it's too elegant a solution to their actual problem (one eval pipeline, many regulators) not to copy. The benchmark charts will keep going up and to the right. The interesting differences will live in the terms of service.

Which brings me to the escape hatch, because there is one and it's sitting on your desk. Open-weights models — GLM, DeepSeek, Qwen, Kimi, MiniMax — run on the Mac you already own, no access list, no classifier second-guessing your job title, no per-token mortgage. They're not Fable 5.1. They don't need to be. For most real work — the automation, the scraping, the file wrangling, the coding — "good enough and yours" beats "frontier and rented."

That's the idea behind CopperRiver, the desktop AI assistant I work on: open-source models doing real agent work locally on your Mac, plans from $9/month, and nobody's safety router deciding your terminal command looks suspicious. If the two-cage future annoys you even slightly, it's worth a look at agent22.app/copperriver.

Fable 5.1 is a great model. Just remember what you're actually buying: not the weights — everyone gets those. The cage.