Anthropic Has Skipped SWE-bench on Three Consecutive Releases Since Hitting the 97% Ceiling
Three Anthropic model launches. Three consecutive releases with no SWE-bench Verified score.
Claude Fable 5.1 shipped September 1. Claude Opus 5.5 shipped September 22. Claude Sonnet 5.5 shipped September 28. None of them include a SWE-bench Verified number in their launch materials — the benchmark that, until this summer, every major Claude release had reported at launch.
The last time Anthropic published one was Claude Opus 5 on July 24, 2026: 96.0% in the official system card, later independently validated at 97.0% by Vals.ai. That score sits at the statistical ceiling of the benchmark’s current discriminative range. Above 96-97%, the task set offers no meaningful separation between frontier models.
What Changed at Launch
The September cluster replaced SWE-bench with a different benchmark portfolio:
| Model | Terminal-Bench 4.0 | OSWorld (partial) | HLE (with tools) |
|---|---|---|---|
| Opus 5.5 | 59.6% | 81.8% (OSWorld 2.1) | 67.7% |
| Sonnet 5.5 | 70.6% | 80.1% (OSWorld 2.1) | 64.5% |
| Fable 5.1 | 55.8% | 77.9% (OSWorld 2.0) | 65.0% |
Fable 5.1 also reported Terminal-Bench-Science 0.1 at 52.6%, CursorBench 3.2.0 at 73.4%, and AutomationBench at 31.4%. GPT-6 Astra and GPT-6.1 Sol have also published Terminal-Bench 4.0 results — 57.9% and not yet reported, respectively — giving the benchmark cross-lab coverage that makes comparisons possible.
Sonnet 5.5 is notable within the cluster: it scores 70.6% on Terminal-Bench 4.0, 11 points above Opus 5.5’s 59.6% and 15 points above Fable 5.1’s 55.8%. The newer Sonnet variant appears to be the strongest of Anthropic’s current agentic coding lineup on this metric, which complicates the usual flagship-versus-efficiency framing.
Why SWE-bench Lost Its Signal
SWE-bench Verified runs 500 real GitHub issues against a bash-only agent. At 97%, Anthropic’s flagship was resolving all but 15 tasks. The remaining failures are increasingly a function of edge-case problem selection, not model capability gaps. The benchmark was designed as a discriminator; at the frontier it has become a threshold test.
Terminal-Bench 4.0, OSWorld 2.0, and HLE serve different functions. Terminal-Bench 4.0 benchmarks long-horizon terminal agent tasks where current frontier models score in the 40-70% range — enough headroom to track progress across generations. OSWorld 2.0 tests UI automation across operating systems. HLE (Humanity’s Last Exam) measures scientific and professional reasoning at graduate and above levels.
None of these are drop-in replacements for SWE-bench in the sense of an established industry-standard comparison point. The SWE-bench Verified bash-only leaderboard has scores from every major lab. Terminal-Bench 4.0’s leaderboard is newer and cross-lab coverage is still building.
The Composite Problem
For scoring purposes, the three models in the September cluster carry null composite scores. The Stack Futures ticker computes a composite from SWE-bench (or tau-bench) normalized to a 0-1 scale. Without a primary benchmark number, imputation with the median is the fallback — but median imputation does not represent actual capability, only a placeholder.
Claude Opus 5 holds a composite score of 1.0 (SWE ceiling). Opus 5.5, which scores higher on LiveBench overall (83.2 vs 83.0 for Fable 5), has no composite because no primary agentic benchmark number exists. That creates an anomaly in the ticker: a model that may be stronger than its predecessor, rendered invisible by a benchmark gap.
Anthropic has not stated it will publish SWE-bench Verified for the September models. If Terminal-Bench 4.0 becomes the new industry standard for agentic coding evaluation, the ticker scoring methodology will need to adapt to use it as a primary input. That process requires cross-lab consistency — which exists partially now but is not yet complete.
For the moment, three Anthropic models launched this week and none of them have a definitive agentic coding score. That is new, and it matters.