Terminal-Bench v2.1: Claude Fable 5.1 Leads at 91.4%
The verified refresh of Terminal-Bench fixes 28 of 89 tasks and introduces continuous validation. Claude Fable 5.1 takes the top spot at 91.4%, with GPT-5.6 Sol and Claude Opus 5 close behind.
Results dashboard
Overall score
Results table
| Agent | Score | Time | Cost | Pass |
|---|---|---|---|---|
| Claude Fable 5.1 (Adaptive Reasoning, Max Effort) | 91 | — | — | Pass |
| GPT-5.6 Sol (xhigh) | 89 | — | — | Pass |
| Claude Opus 5 (Adaptive Reasoning, Max Effort) | 89 | — | — | Pass |
| Grok 4.6 (high) | 88 | — | — | Pass |
| Qwen3.8-Flash-Next | 86 | — | — | Pass |
| Gemini 3.7 Flash (high) | 85 | — | — | Pass |
| Kimi K3 (max) | 85 | — | — | Pass |
| GLM-5.3 (max) | 83 | — | — | Pass |
What changed
Terminal-Bench v2.1 is a verified refresh of Terminal-Bench v2.0, developed by the Laude Institute, Stanford University researchers, and the open-source Terminal-Bench community. The benchmark keeps the same 89 curated tasks across software engineering, system administration, data processing, model training, and security — but fixes 28 of the 89 tasks and introduces continuous validation for agentic benchmarks.
The 28 fixed tasks fell into three categories:
- External dependency drift (9 tasks): Docker images pinned for reproducibility had external dependencies that changed over time. Fixed by pinning or removing internet access where it was not essential.
- Insufficient resource budgets (8 tasks): CPU, memory, or time budgets were too tight for valid solutions — including oracle solutions — to finish consistently. Budgets expanded to accommodate correct approaches.
- Instruction-test mismatches (11 tasks): Instructions asked for one thing but tests expected another (e.g., instructions said PostgreSQL but tests expected Spark SQL). Rewritten for consistency.
After these fixes, no task is unsolved in Terminal-Bench v2.1.
Top results (artificial analysis, september 2, 2026)
The independently scored leaderboard by Artificial Analysis runs Terminal-Bench v2.1 with the Terminus 2 agent harness in an e2b sandbox, reporting pass@1 averaged over 3 repeats per task.
| Rank | Model | Score | Price/1M |
|---|---|---|---|
| 1 | Claude Fable 5.1 (Max Effort) | 91.4% | $20 |
| 2 | GPT-5.6 Sol (xhigh) | 89.5% | $8.00 |
| 3 | Claude Opus 5 (Max Effort) | 89.1% | $10 |
| 4 | Grok 4.6 (high) | 88.4% | $3.00 |
| 5 | Qwen3.8-Flash-Next | 86.1% | $0.23 |
| 6 | Gemini 3.7 Flash (high) | 85.8% | $1.50 |
| 7 | Kimi K3 (max) | 85.0% | $6.00 |
| 8 | GLM-5.3 (max) | 83.9% | $2.15 |
Provider-run snapshot (benchlm, september 1, 2026)
A separate provider-run snapshot published on BenchLM shows GLM-5.3 leading at 88.2%, followed by DeepSeek V4 Pro 0813 at 87.9% and Qwen3.8 Max at 86.6%. These results use different agent harnesses and effort settings and are displayed separately from the Artificial Analysis independent run.
What this means
- Claude Fable 5.1 is the new coding agent frontier. Released September 1, 2026, it immediately took the top spot on the most rigorous agentic coding benchmark. The 75% cache-read price cut makes long coding sessions economically viable.
- The open-weight gap is narrowing. Qwen3.8-Flash-Next at 86.1% for $0.23/MTok is within 5 points of the closed-weight leader at 1/87th the price. GLM-5.3 at 83.9% (or 88.2% in the provider run) is also within striking distance.
- Continuous validation is now a requirement. The 28 task fixes in v2.1 show that benchmarks without continuous validation produce misleading results. Any team running agent evaluations should adopt the same pattern.