Anthropic Just Shipped Fable 5 (Mythos) — Its Most Capable Model
Anthropic Just Shipped Fable 5 (Mythos) — Its Most Capable Model
What Claude Fable 5 can do, why it routes some questions to Mythos, and how it compares to Opus 4.8 and GPT-5.5.
Jun 09, 2026
When a company ships a new product, it gives it one name. One box, one label, one thing you can buy. Anthropic’s latest gives you two names for the same underlying model, and that decision, more than any benchmark, is the story worth understanding.
The two labels are Mythos and Fable 5. Architecturally, they’re the same model: same training, same weights, same ceiling. What separates them isn’t capability — it’s access. Mythos is the model with no restrictions on its most sensitive abilities. Fable 5 is that same model with a hard safety layer in front of it.
Every’s week of testing Fable:
Let me show you what this thing can do before we talk about why it’s fenced in, because the “why” only makes sense once you’ve seen the “what.”
The team at Every — a publication that road-tests these models hard — got about a week with Fable 5 before launch. They didn’t ask it for clever sentences. They asked it for whole projects, each from a single prompt.
One person handed it a grainy, muddy 2007 audio lecture and asked for a way to actually enjoy it. Fable didn’t just clean up the audio. It built a working web app — clear, brightened sound, a full transcript, each sentence lighting up in sync as the speaker talked, usable on a phone.
Another asked it to render Borges’s “Library of Babel” — the famous fictional library that contains every book that could ever be written — as an explorable 3D video game. You could pull books off infinite shelves and climb the spiral staircases.
A third fed it hundreds of customer survey responses and an entire web-analytics stack, and asked what was wrong with the business. It came back with the single biggest problem and a clean, testable experiment to fix it — an idea nobody on the team had thought of.
These aren’t autocomplete tricks. Each is the kind of job that used to eat hours, days, sometimes months. On Every’s hardest internal coding exam their “Senior Engineer” benchmark — Fable 5 scored 91 out of 100, brushing up against the range of actual human engineers. For comparison, the previous flagship, Claude Opus 4.8, managed 63. GPT-5.5 scored 62.
Few items build using one prompt:
A glance at the scoreboard:
Anthropic’s own comparison table is, frankly, a sweep — but the interesting signal is in where the gains cluster, not the headline percentages.
On SWE-Bench Pro — fixing real bugs in real open-source software, start to finish, it hit 80.3%, against Opus 4.8’s 69.2% and GPT-5.5’s 58.6%. Roughly four broken features out of five, repaired on its own.
On Terminal-Bench — actually living inside a command line, installing and configuring and debugging like an engineer at a keyboard, it scored 88%, the kind of number that backs up Every’s verdict that this is the best coding model they’ve tested.
On spatial reasoning — reading floor plans, understanding how physical things fit together, it nearly tripled Opus 4.8’s score (38.6% versus 14.5%). On knowledge work, the dry-sounding category that actually means “the tasks people get paid to do all day,” it edged out every rival. On Humanity’s Last Exam, a deliberately brutal test of expert-level knowledge across dozens of fields, it pulled clear of the pack again.
There’s one honest asterisk worth keeping: on computer use — clicking, typing, and operating apps like a person would, its 85.0% was actually nudged out by an earlier version of the model, which scored 85.4%. Essentially a tie, but a reminder that “best at everything” is marketing, and the truth is “best at almost everything, by a lot.”
The asterisks are the strategy:
Several rows on that table carry a small star: cybersecurity, biology, health. Those stars are where the two models stop being interchangeable — and where the dual-name decision actually lives.
Unrestricted, the model scores 78% on ExploitBench (finding and weaponizing software vulnerabilities) versus Opus 4.8’s 40%, and posts large jumps on hard biology problems. That’s precisely the capability profile that creates dual-use risk. A model that out-paces human experts at vulnerability discovery is, in different hands, an offensive tool. The same fluency that makes it a strong research assistant in the life sciences is fluency you want to be deliberate about distributing.
So Anthropic separated the release along that exact seam:
Mythos is the full model with the safeguards removed, and access stays gated to a small, vetted set of partners. This isn’t improvised — an earlier Mythos build was judged too dangerous for general release back in April and restricted to a handful of partners for evaluation. Today’s launch is the productised, public-facing resolution of that call.
Fable 5 is the general-availability model: identical weights, one behavioral rule. When a request lands in the cybersecurity or biology risk surface, Fable doesn’t answer at full capability. It falls back — routing the query to the older, more conservative Opus 4.8 and returning that response instead. That’s why, on the starred benchmarks, Fable’s effective scores regress toward the Opus 4.8 column. The table shows the model’s raw capability; Fable defines the slice of that capability the public can actually invoke. The delta between those two numbers is the safety policy, rendered as data.
The economics and the operating model
Fable 5 is slow and token-hungry, particularly at high effort settings, and the pricing reflects its tier — $10 per million input tokens and $50 per million output tokens, roughly 2x Opus 4.8 and north of 3x Sonnet 4.6. It will consume usage limits fast. This is not an interactive-chat workhorse; it’s an asynchronous agent you commission.
That maps onto Every’s most telling adoption finding. Advanced practitioners — already fluent at decomposing and delegating to AI — described it as paradigm-shifting on their hardest problems. Less experienced users struggled to find a use that justified it. The constraint has shifted off the model and onto the workflow around it: the quality of the problem framing and the rigour of the review loop now set the ceiling, not the model’s raw ability.
The operating model that gets results is straightforward to state and harder to practice. Treat it as a contractor, not a copilot. Hand it a complete, well-scoped brief; let it run; review the output with the same scrutiny you’d apply to a human’s.
For quick edits and tight back-and-forth, a faster, cheaper model is still the better tool — the latency and cost of Fable only pay off when the assignment is large enough to absorb them.
What it all means?
Strip the labels and the launch reads like this: a lab built a model capable enough that it concluded the public shouldn’t have all of it — not out of distrust, but because the abilities that make it exceptional at building software are the same abilities you’d least want unsupervised in the wild.
Hence two names. Mythos, the complete model, kept behind partner-only access. Fable, the safer telling, shipped to everyone else with a fallback wired into its most dangerous reflexes.
It’s a genuinely unusual posture — a company shipping a quieter version of its own best work on purpose. It’s also, arguably, exactly what you’d want from an organisation holding something this capable. The dual name isn’t a contradiction. It’s a stated position: we built more than we’re ready to release in full, so we’re releasing most of it.
For now, most of it is Fable — and for the teams that learn to delegate the big, hard, well-framed jobs to it, “most of it” is already more capability than the surrounding workflows know how to absorb.