Anthropicreinstated July 1
Claude Fable 5 is back: the most powerful public model returns to service
Anthropic shipped Fable 5 — the first Mythos-class model, a tier above Opus — on June 9, but by June 12 the US imposed export controls: Amazon researchers had found a way around its safeguards. Controls were lifted June 30, and since July 1 the model is globally available again — including on Amazon Bedrock. Agencies have already done the e-commerce math: Fable for the hard stuff (transactional MCP assistants, parsing hundreds of reviews, listing audits), cheaper models for the routine.
- In Stripe's tests, Fable 5 “compressed months of engineering into days” — on a 50-million-line codebase
- $10/$50 per million tokens — double Opus 4.8: do the math on where the upgrade pays off
- The lesson of the 18-day pause: never chain critical pipelines to a single model
- Post-return coding benchmarks cratered — but it's the censor-router, not the model: the before/after breakdown is below ↓
Sources: Anthropic (opens in a new tab) · Fable 5 announcement (opens in a new tab) · Amazon / Bedrock (opens in a new tab) · TechCrunch (opens in a new tab) · EcommLetter (opens in a new tab) · No7 Software (opens in a new tab) · benchmarks: Artificial Analysis (opens in a new tab) · The Deep Dive / BridgeBench (opens in a new tab) · Decrypt (opens in a new tab) · The Batch (opens in a new tab) · Simon Willison (opens in a new tab) · ModemGuides (opens in a new tab)
The Claude lineup · July 2026
Haiku 4.5fast tasks
Sonnet 5$2 / $10 · routine work
Opus 4.8$5 / $25
✦ Fable 5 · Mythos-class$10 / $50 · 1M context
June timeline
June 9release
June 12paused: export controls
July 1back + Bedrock
Was Fable 5 “nerfed”? Verified numbers: before the pause vs after the return
before (June 9–12)after (since July 1)
Artificial Analysis · Intelligence Index−4.9
before64.9 · #1
after60*
BridgeBench · Debugging−60.3
before86.2
after25.9
BridgeBench · Refactoring−35.2
before73.6
after38.4
BridgeBench · Hallucination resistance−14.2
before75.9
after61.7
Where the points went: BridgeMind test, 12 debugging tasks
3 reached Fable 59 intercepted by the classifier → Opus 4.8
Arena.AI (blind human comparisons):
frontend code −27coding −18
documents +34expert writing +25creative +9
The expert verdict: same model — paranoid gatekeeper. Coding categories dropped 35–60 points, Q&A only 14, and non-coding ones actually rose: that's the signature of an input filter, not dumber weights. The new classifier reroutes “suspicious” code to Opus 4.8 (even the words “security”, “hook”, Rust, and Win32 trigger it).
Anthropic admits the cost: the workaround found by Amazon researchers is now blocked in >99% of cases, but “benign requests get flagged more often” — the false-positive rate stays unpublished. Fable 5 is included at 50% of weekly limits through July 7, credits after that.
A counterpoint from real work: after the return, Simon Willison shipped the sqlite-utils 4.0 release on Fable 5 — 34 commits, 5 bugs found, $149.25 total, of which the Opus fallback cost $0.32 (0.2%). For e-commerce content and analytics the “nerf” barely registers — it hits systems coding.
* The Artificial Analysis “after” score is for “Claude Fable 5 (with fallback)”: the classifier intercepts ~8–9% of tasks and Opus 4.8 answers them — that's what drags the score down. BridgeBench is so far the only published full re-run (LMArena and Artificial Analysis haven't released fresh re-evaluations yet); it's unclear whether intercepted tasks were counted as Fable failures. All figures were checked against primary sources by verifier agents.