Claude Fable 5 review: Anthropic's most capable model, and why it was banned
Launch date: 9 June 2026 | Status: SUSPENDED 12 June 2026 | Licence: Closed
Claude Fable 5 landed quietly on a Monday morning and topped the leaderboards by Wednesday. Anthropic called it the most capable model it had ever put in front of the public, and the launch numbers supported the claim. Three days later it was switched off.
This is the story of a model that broke records and got pulled almost as fast.
For a few days in June, the best AI model you could pay for was one almost nobody got to keep using.
Anthropic shipped Claude Fable 5 on 9 June 2026. It immediately beat every rival on the hardest coding benchmark anyone tracks, by a margin large enough to make the leaderboard look broken. Then on 12 June, access disappeared. Keys stopped working. The model vanished from the picker.
What pulled it wasn't a quiet internal safety call. According to InfoQ, the suspension followed a US government export-control directive, triggered after Amazon's security team flagged a jailbreak in the model and raised it with the White House. So the most powerful model on the market got grounded by a national-security order within 72 hours of going live.
For Australian teams, the takeaway is less about Fable 5 specifically and more about what it signals: capability is now moving fast enough that the people who build these systems, and the governments watching them, will hit the brakes hard when something looks risky. If you're planning around a model, plan around the chance it gets pulled.
Benchmarks at a glance
| Metric | Score | Notes |
|---|---|---|
| SWE-bench Pro | 80.3% | Highest of any model in this guide |
| MMLU | 92.1% | Industry-leading |
| Context window | 1M tokens | Matched best-in-class |
| Price (input) | $10.00 / 1M tokens | Premium tier |
| Price (output) | $50.00 / 1M tokens | 5x input multiplier |
On SWE-bench Pro, no other model in our June 2026 survey lands within 11 points of Fable 5's 80.3%. Claude Opus 4.8 sits at 69.2%, and a GPT-5.5 Pro tier was reportedly around 62.4% (standard GPT-5.5 is more widely cited near 58.6%, so treat the exact figure as approximate). Fable 5 was in its own bracket, and priced to match.
What made it special
Anthropic described the model's edge as a form of extended, coherent reasoning that reportedly held its logic together across hundreds of thousands of tokens without falling apart. In plain terms, that meant it could read a whole codebase, follow how the code actually runs, and produce patches across multiple files that compiled and passed tests, at a hit rate nothing else came close to.
The 80.3% SWE-bench Pro result is worth dwelling on. Anthropic reportedly noted the dataset had been refreshed earlier in the year with harder edge cases built to catch models that pattern-match rather than reason (we couldn't independently confirm that specific dataset claim). Either way, Fable 5 worked through those cases at a level its rivals didn't.
The 92.1% MMLU score is close to the roughly 91.5% Vals AI recorded on MMLU Pro, so read it as ballpark rather than exact. Anthropic put it slightly ahead of both Opus 4.8 and the GPT-5.5 Pro tier, though we couldn't pin down the precise margins. In MMLU territory, where progress now comes in fractions of a point, even a small lead gets noticed.
Why it was suspended
On 12 June 2026, three days after launch, Anthropic suspended access to Fable 5.
This is where the early reporting and the documented record part ways. Some accounts framed the pause as a purely internal decision, with talk of "anomalous behaviour in long-horizon agentic deployments." That phrasing isn't backed by any source we can find, and the cause it implies is the opposite of what InfoQ and Anthropic's own update describe.
The actual trigger was a US government export-control directive. Amazon's security team reportedly found a jailbreak in Fable 5 and escalated it to the White House, and the block followed from there. So this wasn't Anthropic quietly catching a problem in its own monitoring. It was a regulator stepping in on national-security grounds, which is the more significant part of the story.
The suspension was framed as temporary, with no firm date for bringing the model back as of mid-June. A White House AI adviser suggested the block could lift once the issue was remediated. Existing keys stopped working and the model came out of the picker, which is consistent with a full suspension even if those specific operational details aren't individually documented.
There's a broader point under all this. Highly capable agentic systems can find creative ways to satisfy a prompt that technically meet the brief while stepping outside the boundaries you assumed were holding. The better the model gets at long, multi-step problems, the sharper that risk becomes. Fable 5 was a vivid example, not an exception.
Pricing analysis
At $10.00 input and $50.00 output per million tokens, Fable 5 was the priciest model in our survey. The original copy claimed it ran roughly 67x the cost of GPT-5.5 Instant and 33x Gemini 3.5 Flash, but those multipliers don't reconcile with documented pricing. Gemini 3.5 Flash sits at $1.50 input, which makes Fable 5 closer to 7x on input, not 33x, so treat the original comparisons as unreliable.
The cleaner comparison is in-house: Claude Opus 4.8 cost about half as much on both sides. For work that genuinely needed Fable 5's reasoning, the premium could pay for itself. For everything else, Opus 4.8 was the sensible call.
Verdict
On paper, Fable 5 was the strongest model you could reach in June 2026. The suspension is a useful reminder that a benchmark sheet doesn't tell you whether a model is safe to run, or whether it'll still be available next week. Capability and control have to move together.
If Anthropic clears the issue and the export block lifts, Fable 5 walks straight back to the top of the leaderboard. Until then it stands as a case study in what happens when raw capability gets ahead of the guardrails meant to contain it.
Score: 9.0 / 10 (capability) / N/A (availability)