Terminal-Bench 4.0 Standardized Run: Claude Opus 5.5 First at 65.15%, Kimi K3.1 Pro Third at One-Third the Cost
Vals.ai updated the Terminal-Bench 4.0 leaderboard on October 26 using the standard mini-SWE-agent harness. The refresh adds Kimi K3.1 Pro to the leaderboard and shifts the top of the rankings. Claude Opus 5.5 holds first place. The cost-per-test column surfaces several counterintuitive results.
Full Top-10 Leaderboard
| Rank | Model | Score | Cost/Test | Tokens/Test | Duration/Test |
|---|---|---|---|---|---|
| 1 | Claude Opus 5.5 | 65.15% | $13.20 | 28.27M | 1h04m |
| 2 | Claude Sonnet 5.5 | 64.14% | $16.51 | 58.53M | 1h22m |
| 3 | Kimi K3.1 Pro | 63.64% | $4.57 | 16.57M | 49m21s |
| 4 | GPT-6 Astra | 59.60% | $9.58 | 3.01M | 35m42s |
| 5 | Claude Fable 5.1 | 58.08% | $17.18 | 18.82M | 1h00m |
| 6 | Wonder Full | 58.08% | $10.53 | 31.82M | 1h12m |
| 7 | Gemini 4 Argon | 57.58% | $17.64 | 16.66M | 53m35s |
| 8 | GPT-6.1 Sol | 55.05% | $1.72 | 2.86M | 45m12s |
| 9 | Claude Opus 5 | 53.53% | $18.60 | 23.71M | 1h07m |
| 10 | GPT-6.1 Terra | 49.49% | $2.04 | 15.59M | 32m45s |
Source: vals.ai/benchmarks/terminal-bench-4, updated October 26, 2026. Standard mini-SWE-agent harness; no scaffold engineering or multi-rollout.
Three Cost Anomalies Worth Noting
Sonnet 5.5 costs more than Opus 5.5 at lower rank. Claude Sonnet 5.5 finishes second at 64.14% but costs $16.51 per test against $13.20 for Opus 5.5. The reason: Sonnet 5.5 generates 58.53M tokens per test versus Opus 5.5’s 28.27M. Sonnet’s lower per-token price does not offset the dramatically higher token volume.
Kimi K3.1 Pro is the value pick in the top three. At $4.57 per test — 65% cheaper than Opus 5.5 and 72% cheaper than Sonnet 5.5 — Kimi K3.1 Pro reaches 63.64% with 16.57M tokens per test and a 49-minute median duration. For production agent pipelines where task throughput per dollar matters more than raw benchmark position, Kimi K3.1 Pro is the frontrunner among top-3 models.
GPT-6.1 Sol is the value leader overall. At $1.72 per test, GPT-6.1 Sol reaches 55.05% — rank 8, but 73% cheaper than Kimi and 87% cheaper than Opus 5.5. Claude Opus 5 at rank 9 ($18.60 per test) costs more than any other model in the top 10 and scores less than every model ranked above it.
The Vendor Self-Report Gap on Kimi
Moonshot reported 67% for Kimi K3.1 Pro on Terminal-Bench 4.0 at launch on October 4. The vals.ai independent measurement on the standard mini-SWE-agent harness shows 63.64% — a 3.36-point gap. The discrepancy is consistent with different scaffold choices. Moonshot’s self-reported figure used a different harness configuration; independent benchmark runs use mini-SWE-agent across all models to enable apples-to-apples comparison. Both numbers are correct in context; they measure different things.
Terminal-Bench 2.0 and 2.1 results, which used a range of scaffolds and harnesses, showed gaps of similar magnitude between vendor-reported and independent scores across multiple models.
What the Leaderboard Doesn’t Capture
Terminal-Bench 4.0 tasks are frontier-difficulty: software, science, ML operations, hardware, security, and media tasks in a terminal environment. The benchmark measures raw model capability given a minimal bash-shell scaffold — it excludes multi-rollout strategies, extended thinking, and proprietary agent frameworks that most production deployments use. Real-world agent system performance varies substantially from these numbers depending on scaffold engineering.
The cost-per-task figures use real API pricing at the time of the run. Models with lower token output on TB4.0 may use more tokens on different task distributions.