A $4 GitHub Clone and a One-Day, 50-Million-Line Migration: Fable 5's First Week

On June 9, Anthropic released Claude Fable 5 — the first model from its Mythos tier that anyone with an API key can use. For over a year, "Mythos" was the model Anthropic talked about but didn't ship: the internal frontier system that benchmark footnotes kept referencing, available only to a handful of government partners under Project Glasswing. Fable 5 is that model, with safety classifiers bolted on for the high-risk domains, released to everyone.

When I wrote about Opus 4.7 in April, the pattern was familiar: a better Opus, same price, retune your prompts. Fable 5 breaks the pattern. It's a new tier above Opus, it costs twice as much ($10 input / $50 output per million tokens, versus $5/$25 for Opus 4.8), and the API behaves differently in ways that will surprise you if you treat it as a drop-in upgrade.

Three days is not enough time for a verdict. It is enough time to see what people reach for when you hand them a model that can run autonomously for hours without falling apart. That's the interesting part.

The Headline Stat Is From Stripe

The number everyone is quoting: Stripe pointed Fable 5 at a codebase-wide migration across 50 million lines of Ruby and it finished in a single day. Their estimate for doing it by hand was a full team working for two-plus months.

That's the category Fable 5 was built for — not "write me a function" but "here is a task that takes a team weeks, go." Cursor called it "the state of the art model on CursorBench" and said it "opened up a class of long-horizon problems that were out of reach for earlier models." GitHub reported it handling complex coding tasks "with a level of autonomy and reliability that exceeded previous benchmarks." GitLab wired it into its Duo Agent Platform to take an issue from interpretation through planning, editing, testing, fixing failures, and opening a merge request.

The common thread: the unit of work is shifting from the prompt to the project. Earlier models could do impressive things inside a turn. Fable 5 holds coherence across the kind of timescales where Opus would drift, forget a constraint, or quietly declare victory.

The Demos: Delightfully Unhinged

The enterprise stories are impressive. The community demos are fun, and they tell you more about what the model is actually like to use.

• It beat Pokémon FireRed with vision alone. Earlier Claude models needed an elaborate helper harness — RAM readers, coordinate overlays, structured game state — to play Pokémon at all. Fable 5 completed the game looking at raw screenshots, the way you would. It also autonomously plays Factorio, which anyone who's lost a weekend to that game will recognize as a legitimate long-horizon planning benchmark.
• A working GitHub UI clone from a screenshot. One developer fed Fable 5 a screenshot of GitHub in Kilo Code and got back a functional replica. Ten minutes, $4.07.
• A real-time black hole. A single prompt produced a Three.js gravitational lensing visualization — doppler-boosted accretion disk, chromatic aberration, 60fps, one file.
• A macOS-style operating system in one HTML file. Draggable windows, a working terminal, a file explorer, and a playable Minecraft clone, with audio synthesized in-browser. One prompt.
• A production game engine in a week. The Spawn 5.0 engine — custom physics, GPU lighting for 1,000+ dynamic lights, real-time global illumination on WebGPU, million-particle effects — built across 1,687 prompts in roughly seven days.
• An SRE that closes the loop. My favorite for practical reasons: Fable 5 diagnosed a hidden infrastructure bug by reading pod logs, querying database errors, and comparing container image digests — then opened the fix PR itself.

And then there's the science, which deserves its own paragraph because it's quietly the wildest part of the announcement. Anthropic says internal protein design experts accelerated drug design work roughly 10x. In blinded comparisons, its molecular biology hypotheses were preferred about 80% of the time over Opus-class output — and one hypothesis about an E. coli protein mechanism was later independently corroborated. In a genomics test, it autonomously assembled single-cell data across 138 animal species and built a custom ML model that outperformed a recent Science publication while being 100x smaller. One more: given persistent file-based memory, its Slay the Spire performance improved 3x over Opus 4.8 — the model is markedly better at writing notes to itself and actually using them.

What's Different Under the Hood

If you're calling the API, Fable 5 is not "Opus but smarter." Plan for these:

• Thinking is always on. You don't configure it — sending thinking: {type: "disabled"} or a token budget returns a 400. You control depth with the effort parameter (low through xhigh and max). Notably, low effort on Fable 5 often beats max on previous models.
• New tokenizer, ~30% more tokens. The same content tokenizes to roughly 30% more tokens than on Opus-tier models. Your max_tokens settings and cost baselines don't transfer — re-measure before you migrate anything.
• Turns run long. A single request on a hard task can run many minutes. If your integration assumes responses come back in seconds, you need streaming, generous timeouts, and progress UX.
• A 1M-token context window and 128K output, with the long-context coherence to actually use it.
• 30-day data retention is required. Fable 5 isn't available under zero-data-retention agreements. If your org runs ZDR, every request 400s — a real constraint for some regulated environments, and the most contentious part of the launch terms.

The Guardrails Are the Story for Security People

Here's the part relevant to how I actually use these models. Fable 5 ships with AI-based safety classifiers targeting cybersecurity and biology — the two domains where Anthropic decided Mythos-class capability shouldn't be generally available. Exploit development, offensive cyber tasks, and agentic hacking get blocked. So does bioweapons-adjacent biology.

The design choice worth noticing: when a classifier triggers, the request doesn't hard-fail. On Claude's consumer surfaces it falls back to Opus 4.8, which answers instead. On the API you get an HTTP 200 with stop_reason: "refusal" and a category, and you can configure server-side fallbacks yourself. Anthropic says over 95% of sessions never touch a classifier, external red-teaming found no universal jailbreaks across 1,000+ hours, and testers got zero compliance on single-turn cyberattack requests.

If you do authorized security work, the practical read is the same one I gave when Opus 4.7 introduced its cyber safeguards: benign-adjacent work will occasionally trip false positives — security tooling, defensive analysis, vulnerability research. The classifiers are deliberately tuned conservative. Build the fallback path into your integration rather than discovering it in production, and watch for the trusted-access programs Anthropic says are coming for verified researchers. Meanwhile, the unrestricted variant — Claude Mythos 5, the same weights without the cyber classifiers, which Anthropic describes as having "the strongest cybersecurity capabilities of any model in the world" — stays gated behind Project Glasswing.

The community reception splits exactly along this line. Developers doing long-horizon coding call it transformational. Researchers in the filtered domains report friction, confusing fallbacks, and false positives. Both are accurate. That tension — ship the capability, gate the dangerous slice, accept the false-positive cost — is the whole bet of the release.

Should You Pay 2x Over Opus?

My early framework, subject to revision as I log more hours:

Stay on Opus 4.8 (or Sonnet) for anything interactive, anything routine, anything where you're iterating with the model in the loop. The 2x price plus the ~30% tokenizer overhead means Fable 5 is effectively ~2.6x Opus cost for the same work, and for turn-by-turn tasks Opus is already excellent.

Reach for Fable 5 when the task is the unit of work, not the turn. Multi-hour autonomous runs. Migrations and refactors you'd otherwise staff. Research tasks where it picks directions, kills its own wrong beliefs, and comes back with something novel — Anthropic's own researchers described Mythos 5's reasoning as "senior research scientist grade," and the early evidence mostly backs the claim. Give it the full spec up front, give it a memory file, and let it run.

The Stripe math is the argument in one line: a team-months task at API prices is a rounding error against the payroll it replaced. The question for the rest of us is how many of our tasks are actually shaped like that. Three days in, the demos suggest the answer is "more than we assumed" — people just hadn't been handing models team-sized work, because until this week there wasn't much point.

---

Sources: Anthropic's announcement, TechCrunch, The Hacker News, CNBC, and the awesome-claude-fable-5 community collection.