Anthropic locked away Mythos as too dangerous. A hobbyist benchmark says: maybe

A model too dangerous to ship, reverse-engineered from its own bug list

Anthropic took its strongest security-research model, Mythos, and its safety-filtered sibling Fable offline, framing them as too capable at finding and exploiting vulnerabilities to release broadly. That left an obvious question nobody outside the company could answer: is Mythos actually a different class of tool, or is the danger talk part marketing?

The developer behind *Will It Mythos?* built a way to probe the edges of it. He took the nine confirmed security bugs Anthropic credited Mythos with discovering, vetted them with Opus 4.7 plus human spot-checks, confirmed they all post-date the other models' knowledge cutoffs, then ran a dozen models through a bare harness. Each model was told to audit a specific file, given the run of the whole repo, and given no clue what it was looking for. A separate judge model — pointed at the bug only to score answers, not to hunt — graded the results against a $100 budget cap per model.

The results refuse to crown a clean winner. The strongest public models cluster at 4 out of 9 bugs: GPT 5.5, Opus 4.8, Gemini 3.5 Flash, DeepSeek V4, and the open mimo-v2.5-pro all land there. GPT 5.5 Pro sits atop the leaderboard only because it blew its $100 budget after four cases (2/4 = 50%), an artifact the author flags directly. Mythos itself isn't a row in the table — it's the implicit 100%, the source of every bug the others are chasing.

Three findings cut against the hype. The cheap Chinese models keep pace with the frontier at roughly a tenth of the cost. A wrapping the models in an agent loop helped no one — "no model performed better with an Agent," while time and tokens climbed. And in a later, unpublished replication, a small open model (Gemma 4 31B) found six of the nine bugs across multiple tries.

"Mixing up corpus selection and the benchmark"

The first wave of comments accused the post of contradicting itself: it says models can find the bugs when told what to look for, then says they weren't told. The author, in the thread, drew the line himself.

"In the benchmark the models were told to look at the file and were allowed to look at the rest of the repo, with no clues about what to look for. During selection of which mythos bugs to include, I needed judge models to be able to determine if contestants found the right bug…" — SwellJoe, the post's author

That distinction matters, because the benchmark's whole claim rests on the models hunting blind.

The case that Mythos is just guardrails switched off

The sharpest counter-thesis in the thread: maybe Mythos isn't smarter, it just isn't refusing.

"I'm pretty convinced that Mythos is just standard LLM with safety features turned off. If current models weren't reluctant to search for vulnerabilities, they might perform as good as Mythos." — a commenter who later got a direct reply from the author

The author had circled the same suspicion, then watched it partly collapse:

"my conclusion, not very strongly held, is: Mythos is both better than other public models and it has fewer guardrails. But, also that the guardrails in current models are probably not strict enough to prevent this work. Only Gemini models when run under Antigravity refused to perform the work." — SwellJoe, the post's author

A different reading came from someone who'd read Cloudflare's writeup of using Mythos: the point isn't raw bug-finding at all.

"Mythos is important and notable because of the harness and self-direction… it was trained to do the end to end analysis autonomously, which is a big deal." — vessenes, citing the Cloudflare report

Another regular noted what this benchmark structurally can't see: per the Anthropic report, Mythos's alleged danger was that it exploited vulnerabilities more reliably, not that it found more of them.

"You guys are getting Fable?"

Running under all of it is grief for the filtered model that's gone. Several developers who got brief access describe a real jump, and they brought receipts: one posted a full session transcript and commits from 36 hours with Fable so others could judge. Concrete wins piled up — autonomously fixing concurrency bugs other models couldn't diagnose from logs, a data-corruption bug in a Qt C++ app that gpt-5.5, GLM-5.1 and DeepSeek V4 all missed.

"Fable on Low was better than Opus 4.8 on High (and cheaper)… Fable nailed these screens fast." — draginol, working in C++ on a shipping game

The dissent is just as concrete. One developer found Fable "just a bigger sink of tokens than anything else," not worth the steps it saved over Opus 4.8. A side-by-side against Codex 5.5 across three real cases — a resource leak, a SPICE model review, an open CS research problem — landed them "in the same place" by solution metrics, with Fable missing an edge case Codex handled.

And a recurring sub-theme almost drowns the model talk: people swearing the existing models keep getting quietly nerfed.

"Around February, Opus 4.6 was excellent… Then it got lobotomized and it's never been the same." — airstrike, who uses Claude Code daily

One reply puts the obvious skeptic's frame on it: this is the audiophile-cable problem, and a blind test between recent versions would probably leave most of the certainty in tatters. Which is the real lesson of the whole exercise. The one developer who actually measured, on bugs no model could have memorized, found the frontier closer and cheaper than the marketing implies — and the locked-away savant still unmeasured.