Is Terminal-Bench 4.0 still meaningful?

Known problems

• It has about 10 false-positive and 6 false-negative tasks (secondary detail).^{[1][1] Terminal-Bench 4.0 review Epoch AI Primary}
• Anthropic reports Opus 5.5 at 66.4% and GPT-6 Astra at 57.9%. Artificial Analysis puts Kimi K3.1 Pro first at 67.2%.^{[3][3] Introducing Claude Opus 5.5 Anthropic Vendor[2][2] Terminal-Bench 4.0 Artificial Analysis Primary}
• Version 2.0 is replaced: its top score was 84.7% (Apr 2026), and the live site shows only 4.0.^{[4][4] Terminal-Bench 2.0 leaderboard Primary}
• UK AISI found that a rise from 1 to 10 million tokens raised Terminal-Bench 2.0 scores by about 25%.^{[5][5] More compute, more capability UK AISI Primary}
• Chinese cards often run Western agent tests on Claude Code or Codex harnesses.^{[6][6] DeepSeek-V4.1-Flash model card Primary[7][7] GLM-5.3 model card Primary} Some report their best score across harnesses.^{[8][8] Qwen3.8-Flash-Next-FP8 model card Primary}

What to look at instead

The report names no direct successor.

Where it appears in the report

Chapter 1, Benchmarks

• The harness and the host can move a score more than the model
• Figure 4. Best score on benchmarks that still resist, November 2026
• Figure 5. Thirteen Flawed verdicts in ten weeks
• Table 2. First model to cross each threshold in “Go deeper: The full saturation timeline (14 benchmarks)”
• Table 4. Epoch AI Benchmark Review verdicts and the main defects found, August to October 2026 in “Go deeper: Epoch's verdicts, benchmark by benchmark”
• Table 6. Coding and software-agent benchmarks, November 2026 in “Go deeper: Coding and software-agent benchmarks (14)”
• Figure 14. OpenAI and Anthropic share only five headline benchmarks

Chapter 2, Evaluation science

• Error bars and adaptive tests make a score trustworthy at lower cost

Chapter 3, Tools and technology

• Two harnesses carry model benchmarking, and Harbor carries agent tasks
• Eval environments and RL training environments merged into one layer

Chapter 6, China and Asia

• Chinese model cards report agentic scores and no safety evaluations
• Figure 3. Chinese Cards Report Western Agent Benchmarks, Their Own Tests and No Safety Results
• Figure 4. Changing the Harness Alone Moves DeepSeek's Score by Almost 9 Points