Anthropic restored Claude Fable 5 on July 1, 2026, nineteen days after a US export-control directive forced the company to switch off its flagship model for every customer in the world. The model returns to the same seat it left: the highest coding score on record, 80.3% on SWE-Bench Pro. Whether that seat is real depends on which leaderboard you read.
What happened to Claude Fable 5?
Claude Fable 5 was offline from June 12 to July 1, 2026 because of a US export-control order, not a technical failure. Anthropic launched the model on June 9 alongside Claude Mythos 5, its less-restricted sibling for vetted partners. Three days later, the government applied export controls after learning of a report in which Amazon researchers bypassed Fable 5's safeguards and got it to identify software vulnerabilities. Anthropic's own testing later found that every model it checked, including GPT-5.5 and its older Claude models, could reproduce the same demonstration. The government lifted the controls on June 30, and Anthropic restored Fable 5 across Claude.ai, the Claude Platform, Claude Code, and Cowork the next day, with a new safety classifier that blocks the reported technique in over 99% of cases and routes blocked requests to Opus 4.8. Mythos 5 remains limited to approved partners. (Anthropic's full account)Nineteen days is a long time in this market. While Fable 5 sat dark, OpenAI announced GPT-5.6 as a gated preview and Claude Opus 4.8 held the top active score. The suspension turned "which model is best at coding" from a settled question into a contested one, and the restoration does not settle it back.
Does Fable 5 actually lead SWE-Bench Pro?
On Anthropic's own scaffolding, yes. On a neutral harness, nobody knows, because Fable 5 has never been scored on one. Three different numbers currently claim the top of SWE-Bench Pro, and all three are technically real:- 59.1%: GPT-5.4 (xHigh), the best score on Scale AI's standardized public leaderboard, where every model runs identical scaffolding.
- 69.2%: Claude Opus 4.8, the best active model on the llm-stats vendor aggregate, where each lab reports its own tuned-harness results.
- 80%: Claude Fable 5, the all-time high on that same vendor aggregate, produced with Anthropic's own tooling.
Which Fable 5 numbers are independently confirmed?
Two figures survive contact with third parties. The independent leaderboard vals.ai confirms Fable 5 at 95.0% on SWE-Bench Verified, a separate and easier benchmark. Artificial Analysis independently measures 1,932 on GDPval-AA. Those are the defensible numbers today.The 80.3% SWE-Bench Pro figure comes from Anthropic's system card and has no independent confirmation. Independent evaluators explicitly flag it as vendor-reported, and Epoch AI's neutral evaluation of Fable 5 was still pending as of mid-June. Until it publishes, every Fable 5 SWE-Bench Pro comparison carries an asterisk.
