Claude Fable 5 Tops SWE-bench Verified at 95% as the 2026 Coding-Agent Race Tightens

The number everyone quoted this month was 95%. The number that actually matters is twelve.

That’s how many models now cluster inside the 80-to-95% band on SWE Bench Verified, the test that defined coding-agent progress for two years. When the whole frontier passes, the scoreboard stops being a scoreboard.

The Benchmark Hit Its Ceiling, Not a Winner

The benchmark broke faster than the models improved.

Anthropic shipped Claude Fable 5 on June 9, 2026. Third-party leaderboard trackers vals.ai and Steel.dev put it at 95.0% — the top score for any publicly-launched model. One caveat the headlines skipped: Anthropic never published that number itself. The 95% is a leaderboard result, not a vendor claim.

And Fable 5 isn’t even the raw leader. Its restricted-access sibling Mythos 5 — the same model with safety classifiers lifted for authorized users — edges it at 95.5% on Steel.dev (Anthropic). Mythos 5 isn’t a product anyone can deploy, so the honest framing is narrow: Fable 5 leads the models actually available to launch.

The distance between first place and twelfth is now thinner than measurement error between trackers.

That’s not a leaderboard. That’s a tie.

Three Findings That Killed the Scoreboard

The saturation story isn’t one number. It’s three, and they point the same direction.

First, the clustering. About twelve models now land between 80% and 95% (AgentMarketCap). Claude Opus 4.8 sits at 88.6%, Gemini 3.1 Pro at 80.6% on a single attempt per Google DeepMind, and GPT-5.2 at 80.0% on Steel.dev — already outpaced by OpenAI’s newer GPT-5.5 at 82.6% on vals.ai. A spread that tight can’t rank anything.

Second, the contamination. An audit found solution leakage in roughly a third of successful Verified patches, with file-path recall running strikingly high (DigitalApplied). Models weren’t always solving the problem. Sometimes they were recalling it.

Third, the walkout. OpenAI’s Frontier Evals team stopped reporting Verified entirely after concluding that more than 60% of sampled problem tasks were unsolvable as written (DigitalApplied).

When the people running a benchmark stop trusting it, the benchmark is done.

Breaking change:

• Claude Fable 5 and Mythos 5: Suspended June 12, 2026, under a US government export-control directive. Both remain offline as of June 26. Anthropic recommends Claude Opus 4.8 as the interim option; return expected around July 1, 2026 (Anthropic).

Who Reads the Signal

Private evaluation infrastructure wins here.

Organizations that built task-specific evaluation harnesses before the leaderboard broke now hold an asymmetric advantage. They measure what matters to their use cases. Everyone else measures a decreasing amount of signal.

SWE-bench Pro is the designated successor. Organizations contributing to its design have a head start.

The evaluation stack benefits broadly. LLM as a Judge infrastructure, Rubric Design, rigorous Cohen's Kappa reliability checks, and annotation tooling like Label Studio all gain importance as fixed-benchmark credibility collapses. The ELO Rating for LLMs approach — ranking models head-to-head on task-specific performance — captures signal Verified no longer provides.

The advantage goes to whoever stops trusting the public leaderboard first.

Who Gets Left Behind

Teams running procurement through benchmark scores.

If your process for choosing a coding model runs through “which one has the best SWE-bench Verified number,” you’re using a signal that OpenAI already decided isn’t fit for its own frontier evaluations. That’s not a minor caveat. That’s a structural mismatch between the question you’re asking and the answer you’re reading.

Vendors marketing Verified scores as differentiators are in the same position. The numbers are real. The differentiation isn’t.

SWE-bench Pro isn’t ready to fill the gap. A May 2026 independent audit found roughly 32% verifier-error rate in Pro — the successor is itself contested. The Human Evaluation for AI baseline that makes benchmarks trustworthy hasn’t been rebuilt.

You’re either building evaluation capacity now or waiting for the next leaderboard collapse.

What Happens Next

Base case (most likely): Vendors shift marketing away from Verified toward private harnesses. SWE-bench Pro takes 12–18 months to establish credibility. Signal to watch: The next flagship model announcement cites SWE-bench Pro or a private harness score instead of Verified. Timeline: Second half of 2026.

Bull case: A contamination-resistant successor earns multi-vendor buy-in. Cross-model comparison recovers. Signal: An independent benchmark with a low audited verifier-error rate gets adopted by multiple labs simultaneously. Timeline: Mid-2027.

Bear case: SWE-bench Pro saturates before establishing authority. Leakage follows the field to the new benchmark. Evaluation credibility in coding doesn’t recover for two years. Signal: Pro scores cluster above 80% within six months of broad adoption. Timeline: Q4 2026 through 2027.

Frequently Asked Questions

Q: Which AI model has the highest SWE-bench Verified score in 2026? A: Among publicly-launched models, Claude Fable 5 holds 95.0% (vals.ai, Steel.dev). Claude Mythos 5 technically scores 95.5% (Steel.dev) but is restricted to Project Glasswing users and currently suspended under a US export-control directive as of June 2026.

Q: How did Claude Opus, Gemini 3.1 Pro, and GPT-5.2 score on SWE-bench in 2026? A: Claude Opus 4.8 scored 88.6% (Steel.dev); Gemini 3.1 Pro scored 80.6% in a single-attempt measurement per Google DeepMind; GPT-5.2 scored 80.0% (Steel.dev). OpenAI’s newer GPT-5.5, launched April 2026, improved to 82.6% on vals.ai.

Q: Is SWE-bench still a meaningful benchmark in 2026 or has it become saturated? A: Saturated. A dozen models cluster between 80–95%. OpenAI dropped it from Frontier Evals after an audit found most sampled tasks were unsolvable as written, and roughly a third of successful patches showed solution leakage. The field is moving toward SWE-bench Pro.

The Bottom Line

SWE-bench Verified reached 95% and stopped being a useful signal in the same month. The successor isn’t ready. The teams that will win the next evaluation cycle are already building private harnesses — not waiting for a leaderboard to tell them which model to trust.

Stay ahead, Dan.