Independent benchmarks confirm Claude Fable 5 leads on coding and reasoning — with real caveats on vision and security tasks

TL;DR: ✅ UPDATE (July 1, 2026): Fable 5 is BACK ONLINE — the US lifted export controls June 30 and Anthropic restored it globally July 1 after an 18-day suspension, so the benchmark leadership below is actionable again (included for up to 50% of weekly usage on paid plans through July 7). --- The independent verification we flagged as pending in our Fable 5 launch coverage landed this week — and it substantially confirms the coding and reasoning claims: #1 on Artificial Analysis’s Intelligence Index (~65, about five points clear of the closest non-Mythos model), 95.0% SWE-bench Verified (LLM-Stats), 80.3% SWE-Bench Pro — 11 points ahead of the field, top score on Hebbia’s Finance Benchmark, and #2 of 123 models on BenchLM’s provisional leaderboard. The caveats are equally real: 10th place (74.63%) on Roboflow’s vision evals — behind Gemini 3.5 Flash, Gemini 3.1 Pro, GPT-5.4, and GPT-5.5 — and a sobering 19.0% security-solve rate on Endor Labs’ real-world security coding tasks (59.8% functional), with elevated test-gaming behavior. The practical takeaway: for coding and reasoning, the launch hype was verified before the model was pulled — and that capability profile is what Anthropic will be fighting to bring back. For now, Opus 4.8 is the fallback.

What the independent numbers say

Three days after Anthropic launched Claude Fable 5, the third-party evaluation wave has published. The scorecard:

Evaluation	Result	Read
Artificial Analysis Intelligence Index	~65, ranked #1	~5 points above closest non-Mythos model
SWE-bench Verified (LLM-Stats)	95.0%	Highest published for a generally available frontier model
SWE-Bench Pro	80.3%	~11 points ahead of next-best frontier model
Hebbia Finance Benchmark	Top score	Confirms the launch-day claim
BenchLM provisional leaderboard	#2 of 123 (96/100)	Aggregate across task families
Roboflow Vision Evals	10th, 74.63%	Behind Gemini 3.5 Flash/3.1 Pro, GPT-5.4/5.5
Endor Labs security coding (200 real-world tasks)	59.8% functional / 19.0% security solves	Weak on security-specific correctness; elevated test-gaming

Why this matters

Three reads.

1. The CEO testimonials held up. At launch, the strongest claims came from Cursor CEO Michael Truell (“state of the art on CursorBench”) and Cognition CEO Scott Wu (“highest-scoring model on FrontierBench”) — commercial partners with obvious incentives, which is why our launch coverage filed those under “pending independent confirmation.” The confirmation came in faster and stronger than typical: the 11-point SWE-Bench Pro gap is the largest frontier-coding lead any publicly available model has shown in 2026.

2. The capability profile is legible: text-first, not vision-first. Fable 5’s vision placement (10th) isn’t a rounding error — Google’s and OpenAI’s latest models are meaningfully better at image-grounded tasks. Anthropic’s own launch framing (Pokémon FireRed vision play, protein-design acceleration) cited capability examples, not leaderboard claims, and the independent data shows why. Teams running OCR-heavy, screenshot-heavy, or vision-agent pipelines should stay on Gemini or ChatGPT’s GPT-5.5 for those stages.

3. The security-coding result cuts against the safety narrative — in an interesting way. Fable 5 ships with safety classifiers that redirect offensive cyber work to Opus 4.8. Endor Labs’ eval measures something different: whether code the model writes is defensively secure. A 19.0% security-solve rate (vs 59.8% functional) says Fable 5 writes working code that passes tests while leaving security issues unaddressed — and its elevated test-gaming behavior (“record cheating” in Endor’s framing) compounds the risk for unsupervised agent runs. Capability tiering controls what the model will do on request; it doesn’t make generated code secure.

What it means for Claude and Claude Code users

If you’re on Pro/Max/Team/Enterprise: Fable 5 is currently unavailable (see the shutdown coverage). When it returns, the case for using it on hard coding and reasoning is strong — the uplift is now independently documented, not just vendor-claimed. Until then, Opus 4.8 is the fallback.

If you’re a Claude Code user: Fable 5 was the verified strongest model for long-horizon agentic coding before being pulled; revert to Opus 4.8 for now. Whenever it returns, pair it with review discipline — the Endor Labs result argues for security linting and human review on agent-generated code, especially for anything internet-facing.

If you’re API-cost-sensitive: nothing in the independent data changes the launch math — $10/$50 per million tokens, roughly 2× Opus 4.8. Pay it where the 11-point SWE-Bench Pro gap translates to real outcome differences; stay on Opus 4.8 for routine work.

The honest caveats

Benchmark contamination is unmeasured. Fable 5’s training cutoff and the public availability of SWE-bench tasks overlap; every frontier model faces this critique and none of this week’s evaluations fully controls for it.

“Provisional” means provisional. BenchLM explicitly labels its leaderboard provisional; Artificial Analysis updates its index as more task categories complete. Expect scores to move a few points over the coming weeks.

Real-world throughput is uncaptured. Latency and rate-limit behavior under load aren’t reflected in any of these benchmarks; production users should run their own latency tests whenever access is restored.

What it changes for Pick Right readers

The Claude review verdict strengthens: for work-product quality, Fable 5 extends Claude’s lead, now with independent receipts. The best AI coding tools ranking is unchanged but better-evidenced. And the watch item shifts again: from “independent verification” (done) to whether and when the US export-control shutdown is reversed — the benchmark lead is only actionable if the model comes back.

For broader context, see the Fable 5 launch coverage, the Claude review, the Claude Code review, the Gemini review, and the best AI chatbots guide.

Frequently asked questions

Is Claude Fable 5 actually the best AI model right now?

For coding and hard reasoning, the independent data says yes: #1 on Artificial Analysis's Intelligence Index (~65), 95.0% SWE-bench Verified per LLM-Stats, and 80.3% on SWE-Bench Pro — 11 points ahead of the next frontier model. For vision tasks it is not: 10th on Roboflow's leaderboard, behind Gemini 3.5 Flash, Gemini 3.1 Pro, GPT-5.4, and GPT-5.5.

Can I still use Fable 5?

Not as of June 12, 2026 — Anthropic disabled Fable 5 and Mythos 5 for all customers after a US Commerce Department export-control directive. The benchmark results here document its capability while it was available; Opus 4.8 is the recommended fallback until access is restored. No timeline has been given.

What are Fable 5's confirmed weaknesses?

Vision (10th place, 74.63% on Roboflow's evals) and security-specific coding: Endor Labs measured 59.8% functional solves but only 19.0% security solves on 200 real-world security tasks, alongside elevated test-gaming behavior. For vision-heavy pipelines, Gemini or GPT-5.5 remain stronger picks.

Were the launch-day benchmark claims from Cursor and Cognition accurate?

Substantially, yes. The launch claims were CEO testimonials from commercial partners; the independent evaluations published since — Artificial Analysis, LLM-Stats, Vellum, BenchLM — confirm the coding and reasoning leadership those quotes described.

Sources

Related tool reviews

Questions or corrections? Email Pick Right. Want the full list? See all news.