Benchmarks

Benchmark / 6 November 2026

Terminal-Bench 2.1

Midas with DeepSeek V4.1 Flash reaches 92.2% at $0.08 per task. The previous best, Opus 5.5, scores 87.6% at $0.50. The same DeepSeek V4.1 Flash run on a standard harness scores 74.5% at $0.10.

The field

The horizontal axis is cost per task, the vertical axis is accuracy, zoomed to 60% and up. Nineteen frontier systems from vals.ai, plus the Midas run, on the same plane. Midas sits above every model and to the left of all but the cheapest, weakest runs.

The left edge is the point of the chart. Every system in the cheap vicinity scores far below Midas, and the previous best costs more than six times as much per task.

This result used the production harness with one pinned model, DeepSeek V4.1 Flash, to demonstrate Midas' ability to maximize model performance. Using Midas auto, its production setting, it would have scored even higher.

Terminal-Bench 2.1

Difficult terminal-based coding tasks

Terminal-Bench 2.1 · Oct 01, 2026 · Midas: Nov 06, 2026 · Other runs: vals.ai

AccuracyCost
Lab:

Pinch to zoom, then drag to pan

Terminal-Bench 2.1 accuracy versus cost per taskFrontier systems from vals.ai plotted by accuracy and cost per task. Midas with DeepSeek V4.1 Flash reaches 92.2 percent at 8 cents per task, above every model in the chart. The same model on its standard harness scores 74.5 percent at 10 cents. Zoom with Ctrl or Cmd plus scroll, drag to pan, or Shift-drag to select an area.EFFICIENTINEFFICIENTLOWEST COSTBEST PERFORMANCE$0.00$0.50$1.00$1.50$2.00$2.50$3.00Midas + DeepSeek V4.1 Flash92.2% · $0.08DeepSeek V4.1 Flash74.5% · $0.10Claude Opus 5.5GPT-6 AstraGPT-5.6 SolClaude Fable 5.1GPT-5.6 Luna60%70%80%90%100%

Accuracy vs cost per task (linear scale). Midas: internal run, 6 Nov 2026.

Field as of 1 October 2026

Ranked

Efficacy and cost per completed task

Sorted by efficacy

  1. 01Midas + DeepSeek V4.1 Flash92.2$0.08
  2. 02Claude Opus 5.587.6$0.50
  3. 03GPT-6 Astra87.3$1.34
  4. 04GPT-5.6 Sol85.8$1.02
  5. 05Claude Fable 5.185.0$3.07
  6. 06Claude Opus 584.6$0.89
  7. 07GPT-6 Sol83.1$0.37
  8. 08Claude Sonnet 5.583.1$0.63
  9. 09Gemini 3.8 Flash81.3$1.54
  10. 10Kimi K380.9$0.34
  11. 11Claude Fable 580.5$1.43
  12. 12GPT-5.6 Luna79.0$0.05
  13. 13Muse Spark 1.3 Max79.0$0.60
  14. 14MiMo V2.6 Flash76.4$0.02
  15. 15DeepSeek V4.1 Flash74.5$0.10
Terminal-Bench 2.1 accuracy and cost per task for the plotted systems
SystemProviderAccuracyCost per taskSource
Midas + DeepSeek V4.1 FlashMidas92.2%$0.080Midas run
Claude Opus 5.5Anthropic87.6%$0.50vals.ai
GPT-6 AstraOpenAI87.3%$1.34vals.ai
GPT-5.6 SolOpenAI85.8%$1.02vals.ai
Claude Fable 5.1Anthropic85.0%$3.07vals.ai
Claude Opus 5Anthropic84.6%$0.89vals.ai
GPT-6 SolOpenAI83.1%$0.37vals.ai
Claude Sonnet 5.5Anthropic83.1%$0.63vals.ai
Gemini 3.8 FlashGoogle81.3%$1.54vals.ai
Kimi K3Moonshot AI80.9%$0.34vals.ai
Claude Fable 5Anthropic80.5%$1.43vals.ai
GPT-5.6 LunaOpenAI79.0%$0.054vals.ai
Muse Spark 1.3 MaxMeta79.0%$0.60vals.ai
MiMo V2.6 FlashXiaomi76.4%$0.023vals.ai
DeepSeek V4.1 FlashDeepSeek74.5%$0.099vals.ai
Grok 4.7SpaceXAI73.4%$1.08vals.ai
GPT-6 LunaOpenAI73.0%$0.029vals.ai
Muse Spark 1.3Meta72.3%$0.66vals.ai
MiMo V2.6 ProXiaomi67.8%$0.073vals.ai
DeepSeek V4 Flash 0731DeepSeek67.0%$0.021vals.ai

All non-Midas numbers are from vals.ai, recorded 1 October 2026. Midas is an internal run from 6 November 2026. Accuracy and cost per task.