Claude Fable 5 Review: Jaw-Dropping and Way Too Expensive

Fable 5 launched 9 June 2026. I've been running it on Pro since day one, in the free window before it requires usage credits. These are my notes from that testing.

Update, 13 June 2026: the day after I published this, the US government suspended access to both Fable 5 and Mythos 5. Right now you can't run either one. Everything below still holds for whenever it comes back, but the buying decision is moot for the moment. I added a section near the bottom on what happened.

On 9 June 2026, Anthropic released Claude Fable 5. It's the first model in what they call the Mythos class, a tier above Opus. The announcement leads with autonomy: plans across stages, delegates to subagents, sustains complex agentic work for up to 12 hours at a stretch. And there's a second version, Claude Mythos 5, the same underlying model without the public safety guardrails, available only to approved organizations like Project Glasswing cybersecurity partners and select biology researchers.

Some of what this model can do is genuinely impressive, and for most of my actual work the cost doesn't make sense.

What Fable 5 actually costs

The number you need first: $10 per million input tokens, $50 per million output. That's exactly double Opus 4.8, which runs $5 and $25.

I'm on the 22-euro Pro plan. Fable 5 was included free from launch through 22 June, which is why I can test it now. From 23 June it needs usage credits on top of the base plan, so it effectively leaves Pro unless you pay extra.

To understand what that doubling feels like in practice, look at Simon Willison's notes from launch day. He spent $110.42 in a single day of testing. Ninety-nine dollars of that was one project, which he called "slow, expensive" but also "something of a beast." He also measured a simple pelican SVG across the five effort levels: about 9.67 cents at low, about 72 cents at max.

One widely shared report from launch week described a user burning over $1,000 in tokens in a single day on a $200 plan. Developer backlash on X and Hacker News covered both the token costs and the mandatory 30-day data retention policy for Mythos-class traffic, including coverage from Decrypt.

Fable 5 also has safety classifiers covering cybersecurity and bio/chem topics. Flagged requests are silently rerouted to Opus 4.8. Anthropic says under 5% of sessions trigger this fallback.

What it can actually do

The headline capability is autonomy. Fable 5 can plan a task across stages, spin up subagents to handle pieces of it, and keep running without human intervention for hours. Anthropic's announcement shows a solar system physics simulation that predicts eclipses. Stripe reportedly used it to run a 50-million-line Ruby migration in about a day, where the original estimate was two months.

The coding demos have been wild. Ethan Mollick one-shotted several playable games from single prompts in Claude Code: a Snake variant, an exploration game, a Rilke-themed walking game. The Matt Shumer Minecraft demo reportedly built a functional clone in custom ThreeJS in 20 to 55 minutes for under $50 in tokens. Fable 5 is also the first Claude model to finish Pokemon FireRed using vision alone.

The demos are consistent enough that the capability is genuine, not a cherry-picked launch stunt.

The five effort levels (low, medium, high, xhigh, max) give you some control over how much it spends. But as Willison's numbers show, even low effort on a simple task costs more than Opus on the same job, and max effort compounds fast on anything complex.

The model lineup now has five tiers

Anthropic's lineup now runs Haiku, Sonnet, Opus, Fable, and Mythos. Five tiers. The Superpowers skill I use in Claude Code already routes simple steps to Sonnet automatically (I wrote about this setup in my Claude Code skills field report). That's the right instinct applied to a three-tier world. In a five-tier world, the routing question gets sharper.

For my daily work, mostly content workflows, browser tools, and mid-size automation, Opus 4.8 on high effort is already more than I need for most tasks. I wrote about that in my Opus 4.8 review and the conclusion still holds. The jump to Fable 5 doesn't change what I produce on those tasks. It changes the bill.

When do you actually need this tier?

My answer, based on testing and the evidence from launch week: multi-hour autonomous agent runs where you need it to sustain attention and plan across dozens of steps without help, migrations or refactors at the scale of tens of millions of lines, full million-token context jobs where you need the whole thing active at once, and one-shot builds of genuinely complex interactive systems, the kind where a junior developer would need weeks.

For blog posts, small tools, everyday coding, or any task where you're going to review and redirect after each step anyway, Sonnet or Opus already finishes the job.

I keep costs manageable by matching the model to the task. Sonnet handles simple checks and formatting steps. Opus handles planning and hard reasoning. The Superpowers skill does this automatically, and Caveman keeps context tight so Opus can run longer without hitting limits. These habits matter more now, not less.

The token trap, one more time

I called this for Opus 4.7: you build workflows on a stronger, hungrier model, get used to the quality, and gradually lock yourself into a cost structure that's hard to walk back. The pattern is playing out at enterprise scale with Opus. Microsoft dropped Claude Code internally because token spend blew through the annual AI budget. Uber burned through its 2026 coding budget in about four months.

Fable 5 doubles the per-token price on top of that pattern. If you're already feeling the Opus budget pressure and you start defaulting to Fable 5 for everything, the math gets painful fast.

Fable 5 has genuinely different capabilities, and for the tasks it's built for those are worth paying for. But "more capable" and "worth paying for in this specific case" are different questions, and the second one is the one that actually matters for how you work.

Then the government pulled it

One day after I published this, on 12 June, the US government issued an export control order suspending access to both Fable 5 and Mythos 5. Everyone loses it: paying customers, foreign nationals, even Anthropic's own staff. The other Claude models (Haiku, Sonnet, Opus) are unaffected, so this is specifically a Mythos-class problem. (Anthropic's statement is here.)

The stated reason is a jailbreak, and the method is almost funny given everything above. Officials say they found a way around the safeguards by asking the model to read a specific codebase and fix any software flaws it finds. That's it. That's the same "point it at a big repo and let it work" capability the whole launch was built on, the thing Stripe used for the 50-million-line migration.

Anthropic pushed back on most of it. They call the exploit narrow rather than universal. They point out the same capability "is widely available from other models (including OpenAI's GPT-5.5)" and that it's the kind of thing defenders use every day to keep systems safe. And they argue that if this is the bar, the same standard would "essentially halt all new model deployments" across the industry.

Their safety case is that they did the work upfront: limits on cybersecurity misuse, thousands of hours of red-teaming with the government, the UK AI Safety Institute, and outside groups, plus that 30-day data retention policy developers were already annoyed about, which is partly there to catch jailbreaks faster. The defense-in-depth pitch is that a working jailbreak should end up either narrowly scoped or expensive enough not to matter at scale.

I don't have a clean take on who's right here. The capability the government is worried about is the exact capability I spent this whole post describing: the autonomy, the codebase-scale work. That's what makes Fable 5 worth its price, and apparently also what got it pulled. Worth watching how this resolves, because the answer sets a precedent for every model above this line that ships next.

What to take from this

For multi-hour autonomous migrations or one-shot complex builds, Fable 5 is the right tool. For daily content work or small automation, the per-token cost isn't justified.

The practical answer for most people right now: stay on Opus 4.8 for daily work. As of 12 June that's not even a choice for Fable 5, since the suspension took it off the table for everyone. If it comes back and you have a specific task that fits the use case, spin it up, do the job, and measure what it costs. Don't let it become the default.

The lineup is broader than it's been before. Five tiers means more options, which is useful only if you match the tier to the task. For where the Sonnet/Opus line sits in my day-to-day, see my Sonnet vs Opus comparison. The same cost and quality reasoning that works for Sonnet versus Opus now applies one level higher, to Opus versus Fable.