ProgramBench task detail

wintermute-cell__ngrrram.8ea13c3

Task-level results for this Codex /goal scaffold. ProgramBench baseline context is cached from the official task page; Codex scored tests are after active-branch and ignored-test filtering.

303generated behavioral tests
97.0%official best score
97.0%Codex best score
3Codex result rows

Codex Results by Model

#ModelRunAgentModeScoreEvalTestsEst. costCallsWallEvidence
1 GPT 5.5 (xhigh) 20260517T094700Z Codex /goal Mini-SWE-compatible no internet 97.0% ok 294/303 $4.21 84 0.20h manifest · eval json · eval summary · usage audit
2 GPT 5.5 (xhigh) 20260518T232006Z Codex /goal Paper prompt + /goal no internet 71.9% ok 218/303 $1.98 37 0.15h manifest · eval json · eval summary · agent summary · usage audit
3 GPT 5.5 (xhigh) 20260519T215046Z Codex /goal Paper prompt + Goal contract no internet 71.9% ok 218/303 $1.40 31 0.11h not exported

Why Scores Differ

Official ProgramBench rows are the public mini-SWE-agent submissions. GoalBench rows are separate Codex /goal submissions. Same model label does not mean the same agent, prompt, tool loop, or generated implementation. A GoalBench row is resolved only when the ProgramBench-scored pass rate is exactly 100%.

GPT 5.5 (xhigh) · 20260517T094700Z

97.0% from 294/303 ProgramBench-scored tests. Raw public eval statuses: failure: 11, passed: 321.

Agent trace summary unavailable.

Likely miss class: behavioral mismatch, terminal rendering mismatch.

  • 5eb07a00de06 tests.test_keyboard_layouts.test_azerty_layout_rendering (failure)
  • 5eb07a00de06 tests.test_keyboard_layouts.test_qwerty_to_qwertz_emulation_workflow (failure)
  • 5eb07a00de06 tests.test_keyboard_layouts.test_qwerty_to_azerty_emulation_workflow (failure)
  • 5eb07a00de06 tests.test_keyboard_layouts.test_qwertz_distinctive_keys (failure)
  • 5eb07a00de06 tests.test_keyboard_layouts.test_qwertz_layout_rendering (failure)
  • 5eb07a00de06 tests.test_ngrams.test_empty_file_triggers_panic (failure)

manifest · eval json · eval summary · usage audit

GPT 5.5 (xhigh) · 20260518T232006Z

71.9% from 218/303 ProgramBench-scored tests. Raw public eval statuses: error: 1, failure: 85, passed: 243, skipped: 3.

Agent trace: 59 shell command(s), 42 executable-observation command(s), 3 build command(s), 1 blocked attempt(s). Raw Codex logs are not published.

Likely miss class: behavioral mismatch, interactive terminal behavior mismatch, terminal rendering mismatch.

  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_run_with_tetragrams_shows_content (failure)
  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_run_with_trigrams_shows_content (failure)
  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_run_with_bigrams_shows_content (failure)
  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_run_with_words_shows_content (failure)
  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_default_run_shows_lesson (failure)
  • d3e0b8bd2908 eval.tests.test_tui_interaction.test_run_with_custom_wpm (failure)

manifest · eval json · eval summary · agent summary · usage audit

GPT 5.5 (xhigh) · 20260519T215046Z

Public eval evidence has not been exported for this row yet.

Official ProgramBench Results by Model

#ModelProviderScoreCostCalls
1 GPT 5.5 (xhigh) OpenAI 97.0% $8.86 68
2 GPT 5.5 OpenAI 96.7% $1.51 25
3 GPT 5.5 (high) OpenAI 85.8% $2.88 32
4 Claude Opus 4.7 Anthropic 84.5% $22.00 318
5 Claude Opus 4.7 (xhigh) Anthropic 73.6% $31.63 326
6 Gemini 3 Flash Google 62.0% $0.33 72
7 Claude Sonnet 4.6 Anthropic 58.1% $39.89 576
8 Claude Opus 4.6 Anthropic 47.5% $11.82 194
9 Claude Haiku 4.5 Anthropic 43.9% $0.60 124
10 GPT 5.4 OpenAI 31.0% $0.28 10
11 GPT 5.4 mini OpenAI 31.0% $0.02 7
12 GPT 5 mini OpenAI 27.7% $0.01 9
13 Gemini 3.1 Pro Google 13.5% $4.25 164