Fable 5: The Mythos Model Just Went Public

In April I wrote about Mythos Preview, the model Anthropic said might never be released in its current form. It went to twelve founding partners and a closed group of organizations maintaining critical infrastructure. Everyone else waited, and the compounding clock ran only for the people inside the gate.

Today the gate opened. Anthropic released Claude Fable 5, a Mythos-class model, to everyone with an API key or a paid plan.

I build systems with AI every day, so a release like this changes my week, my clients' systems, and the planning conversations I have with companies. This post covers what Fable 5 is, what it does better in practical terms, and where the access story I wrote about lands now.

What Anthropic Shipped

Fable 5 is the public version of the Mythos-class model. The capability is the headline, but the release mechanism is the interesting part: instead of holding the model back, Anthropic ships it with a safeguard layer. Queries that touch high-risk areas like offensive cybersecurity or biology get answered by Claude Opus 4.8 instead. Everything else runs on Fable 5 directly.

Anthropic reports the fallback triggers in under 5% of sessions. For most work, you are simply using the strongest model they have ever released to the public.

The pricing breaks a pattern, and this part is easy to miss. Every previous Claude release was folded into the existing monthly plans on day one and stayed there. Fable 5 is the first that is not. Paid plans include it free only through June 22. After that it leaves the subscription, and every Fable 5 token is billed at usage rates on top of whatever plan you already pay for.

Mythos 5 itself, without the safeguard layer, stays restricted to Project Glasswing partners and select researchers. The frontier still has a gated tier. But the public tier just jumped to within arm's reach of it.

What It Does Better in Practice

Benchmark deltas are easy to find. The more useful question for anyone running AI in a business is where the gains show up in working systems. Four areas stand out.

1. Long projects finish instead of drifting

The clearest practical gain is long-horizon work. Prior models could run multi-step tasks, but on long chains they drift: they lose the goal, redo finished work, or declare victory early. Fable 5 holds focus across tasks spanning millions of tokens.

The reference point from the launch: Stripe used Fable 5 to migrate a 50-million-line codebase in one day. Their manual estimate for the same work was two months.

Takeaway

The unit of delegation is shifting from a task to a project.

For company workflows, that changes what you can hand off. A document backlog, a data migration, a quarter of reconciliation cleanup: work that used to need a person re-aiming the model every hour can now run as a supervised job with review at the end.

2. Vision you can hand a screenshot

Fable 5 can rebuild the working source code of a web app from screenshots alone. As a stunt, it also completed Pokémon FireRed using nothing but raw game screenshots, with no helper tools reading the game's memory.

The business translation: legacy systems with no documentation, scanned paperwork, dashboard exports, charts in old PDFs. A lot of company knowledge exists only as pixels. Models that read pixels at this level can pull that knowledge into systems that use it.

3. It catches its own mistakes mid-task

Anthropic describes Fable 5's reasoning as operating at senior research scientist grade, with stronger reflection and self-validation. It posted the top score on Hebbia's Finance Benchmark for senior-level reasoning and leads Cognition's FrontierCode evaluation for production code quality.

In April I wrote that agent systems are chains of steps, and small accuracy gains compound across the chain. Self-validation is that effect concentrated: a model that notices a wrong step at step 3 saves the whole run, because errors that survive early steps multiply through everything downstream.

4. Less tuning for the people who build with it

A quieter change for builders: the API surface keeps getting simpler. There are no sampling knobs to tune on Fable 5, and the model decides when and how much to think on its own. You describe the outcome and give it room.

That sounds minor. In practice it removes a whole category of configuration that teams used to maintain, test, and argue about. The model carries more of the judgment that used to live in settings.

I Ran My Own Tests. Then a Harder One.

Launch-day claims deserve testing, so on day one I ran my own comparison: Fable 5 against Opus 4.8 and Sonnet 4.6, the models most teams are using right now. I started with three short, scorable tasks, three trials per model, identical prompts. 27 runs total.

1. A planted-error audit: an expense report where exactly one number is wrong, constructed so that checking only the totals points you at the wrong answer. Finding the true error requires verifying every underlying row.
2. A production bug hunt: a bank-reconciliation function that intermittently loses real transactions. The root cause is a de-duplication key that collapses distinct same-amount, same-day transactions, with plausible decoy issues nearby.
3. A chart with no data labels: a rendered bar chart where the models had to estimate six values from pixels alone.

Every model dodged the audit trap in every trial. Every model found the dedup bug and prescribed the same fix. On the chart, Fable 5 and Opus 4.8 read all eighteen values exactly; Sonnet 4.6 read one bar as 24 instead of 23 in two of three trials. Across 27 runs, that single unit was the entire measurable difference.

Takeaway

On short, bounded tasks, the public models have converged. The difference lives in the long tasks.

The quick-question taste test cannot tell these models apart anymore, and that includes Sonnet at a third of the price. So I gave them a harder problem, the kind Anthropic says is Fable 5's actual territory.

The Real Test: Audit a 150,000-Line Codebase

I pointed each model at one of my own production backends: a multi-tenant TypeScript service of about 150,000 lines, far more than any model can read in full. The instruction was the same for all three: audit it read-only, find every defect that could cause incorrect behavior, data loss, or a security problem, report everything, and suggest the highest-value upgrades. One pass each, same prompt, same tools.

This is the task that can't be faked with a clever trick. A model has to decide what to read, hold a map of the system in its head, and reason about how pieces interact across files. And here the three models came apart.

They agreed on a solid floor: all three caught the same handful of real issues, including a customer-facing integration that was stubbed out and silently failing, and in-memory state that quietly resets every time the server restarts. Any of the three would have earned its keep. But each one also surfaced a serious problem the other two completely missed, and the misses fell along clear lines.

Takeaway

On the easy tests they tied. On a real codebase, they had different blind spots.

Fable's standouts were the ones I'd least want to ship without: a fail-open secret and a fee rule that was only half-enforced. Both require understanding what the code is supposed to do, not just spotting a suspicious pattern. That is the difference the benchmarks point at, showing up in a real review rather than a leaderboard. Every issue here was caught in a pre-release audit, which is exactly when you want it caught.

The Access Story, Two Months Later

My April post argued that access was becoming the advantage: Glasswing partners were compounding learnings on a model nobody else could touch, and the public might wait a long time.

The capability went public faster than the language suggested. "May never be released in its current form" turned out to mean: released in a modified form, eight weeks later. The safeguard fallback is what made that possible. It let Anthropic ship the capability broadly while keeping the highest-risk uses on a tighter leash.

The honest read is that both things are true. The equalizer era got an extension: an API key once again buys capability close to the frontier, and for two weeks a $20 subscription does too. And the partners who spent two months operating Mythos-class systems still hold that head start, plus continued access to the unrestricted model.

The gate moved. It did not disappear. But the distance between public and frontier just collapsed from a chasm to a step, and that changes the planning math for every team that was telling itself it could wait.

What I Would Do With the Free Window

Fable 5 is included on paid plans at no extra cost through June 22, and then it is gone from the subscription for good. This is the first Claude model that monthly plans do not cover. You get two weeks to find out what it changes for your work, and after that every run is a line item.

1. Skip the quick-question taste test. My head-to-head runs came back nearly identical across Fable 5, Opus 4.8, and Sonnet 4.6 on short tasks. You will not feel the difference there.
2. Point it at something big and messy you actually own: a real codebase to audit, a long document set to reconcile, a migration you keep putting off. That is where the models separated for me.
3. Hand it something visual: a screenshot of a legacy tool, a scanned form, a chart from a PDF. See what it extracts.
4. On a high-stakes review, run more than one model. They have different blind spots, and the second model is cheap insurance against the first one's miss.

Keep Reading

The April piece on Mythos Preview and why access was becoming the new advantage:

And if the free window has you wondering where to point a stronger model first: