Anthropic's new Fable and Mythos models land a week before its IPO could price
Claude Fable 5.2 is open to everyone and costs 20% less than its predecessor. Claude Mythos 5.2, the same model with fewer safety limits, stays locked to vetted cybersecurity and life-sciences organizations.
Anthropic picked a loaded week to ship two new flagship models. Claude Fable 5.2 and Claude Mythos 5.2 went out on November 9, Anthropic says, and CNBC reported the company’s IPO could price within days. My read: that timing isn’t a coincidence, and CNBC’s own headline called this a release “ahead of IPO.”
Fable and Mythos are, per Anthropic, the same underlying model wearing two different seatbelts. Fable ships with cybersecurity and biology safety filters that block most high-risk requests, which is why Anthropic can hand it to anyone with a Claude account. Mythos is the unrestricted version. It only goes to organizations vetted through Anthropic’s Cyber and Life Sciences Verification Programs. That’s not a new structure. Anthropic ran the same split for the 5.1 generation in September, but 5.2 is the first time the price dropped alongside a capability bump.
The numbers that matter if you’re paying for tokens
Fable 5.2 and Mythos 5.2 cost $8 per million input tokens and $40 per million output tokens, Anthropic says, which it frames as 20% cheaper than the 5.1 generation. They’re also supposed to generate output up to 30% faster. There’s an optional fast mode too: $16 input and $80 output per million tokens, for up to 2.5x faster generation, though Anthropic’s page notes Mythos fast mode only ships to Verification Program members.
CNBC quoted Dianne Penn, Anthropic’s head of product management for research, saying the company “called for pacing, not pausing.” Nice line. Penn told the outlet Fable 5.2 is really meant for “power users or power use cases,” and that most people should keep using Claude Opus 5.5 for daily tasks and only reach for Fable when a problem resists it. That’s a more modest pitch than the benchmark table on Anthropic’s own site, which shows Fable 5.2 beating Opus 5.5 on nearly every metric Anthropic chose to publish, including a jump from 32.4% to 49.1% on a coding benchmark called FrontierCode 2.0.
The claims I can’t verify myself
Anthropic says a team of agents running Fable 5.2 formalized a complete proof of Fermat’s Last Theorem in 30 hours. It also says the model independently solved two open problems in theoretical black-hole physics, with preprints posted for both. Big claims. I haven’t read the math closely enough to judge whether the results hold up to peer review, and Anthropic’s own footnote says none of its external safety evaluators reviewed those specific claims. My read: a vendor’s benchmark page is not where you’d want to confirm a 30-year-old theoretical physics problem got solved, so treat this as what the company is claiming today, not as settled.
On cybersecurity, Anthropic’s data is more checkable since it ties to a named benchmark: Mythos 5.2 found 10.2% more vulnerabilities than Opus 5.5 and 14.4% more than Mythos 5.1 on Anthropic’s internal CVE Rediscovery test. A security company (unnamed in Anthropic’s post) told Anthropic that one in six bugs Mythos 5.2 caught were missed by every other frontier model it tested. That’s a specific, falsifiable-sounding number, but it’s still Anthropic citing an unnamed third party about its own model, not an independent audit.
Anthropic’s own benchmark table, published alongside the announcement, lines up a few of these claims against the 5.1 generation and Opus 5.5:
| Benchmark | Fable 5.2 | Fable 5.1 | Opus 5.5 |
|---|---|---|---|
| Agentic coding (FrontierCode 2.0, high effort) | 49.1% | 32.4% | 45.5% |
| Agentic coding (FrontierSWE v2) | 69.9% | 56.3% | 62.3% |
| Cybersecurity (CVE Rediscovery, Mythos) | 89.6% | 78.3% | 81.3% |
| Knowledge work (Remote Labor Index, high effort) | 41.3% | 17.9% | 21.3% |
All four numbers come from Anthropic’s own test suite, not an outside lab, so they say more about how Anthropic chose to measure than about how the models perform on a reader’s own task.
Who should care right now
Developers already building on Claude get the easy win: lower per-token pricing and a faster model, with Fable 5.2 live today on the Claude Platform, AWS, Google Cloud and Microsoft Azure. Anyone doing cybersecurity research or biomedical work that needs Mythos’ unrestricted mode has to apply through Anthropic’s verification programs first, and that process isn’t instant.
For everyone else, the more interesting story might be the timing. Anthropic filed its S-1 with the SEC in late October, and CNBC’s sourcing points to a stock pricing within days. My take: a flagship model launch days before that reads like standard tech-IPO choreography, showing growth and product momentum right when investors are deciding what the company is worth. Anthropic didn’t frame it that way in its post. I’m not accusing anyone of anything, just noting that Wall Street sees this release at the same moment everyone else does.
One more thing worth flagging: Anthropic’s benchmark table compares Fable 5.2 against “GPT-6 Astra” and “Gemini 4 Carbon,” names that don’t match any model OpenAI or Google has announced publicly as of this writing. Either those are internal Anthropic codenames for competitor models tested under NDA, or placeholder labels, and Anthropic’s post doesn’t clarify which. I’d treat those specific comparison numbers with more caution than the Fable-vs-Fable 5.1 numbers, since at least those compare Anthropic’s own models to themselves.
Underdog shipped a 32MB voice model the same week that still beat a 41GB rival on hard sentences, which is a reminder that not every AI release this month is chasing the frontier-lab benchmark race Fable and Mythos are built for.
Note on sources: I read Anthropic’s full announcement and CNBC’s report in full, including the Dianne Penn quotes. I did not independently verify the Fermat’s Last Theorem or black-hole physics claims, since that would require reviewing unpublished math, and I say so above. The Verge’s report is secondary and used only to confirm the September 5.1 launch date.