312generated behavioral tests
80.4%official best score
79.2%Codex best score
3Codex result rows
Codex Results by Model
| # | Model | Run | Agent | Mode | Score | Eval | Tests | Est. cost | Calls | Wall | Evidence |
|---|---|---|---|---|---|---|---|---|---|---|---|
| 1 | GPT 5.5 (xhigh) | 20260517T094700Z |
Codex /goal | Mini-SWE-compatible no internet | 79.2% | ok | 247/312 | $4.30 | 88 | 0.26h | manifest · eval json · eval summary · usage audit |
| 2 | GPT 5.5 (xhigh) | 20260518T232006Z |
Codex /goal | Paper prompt + /goal no internet | 69.2% | ok | 216/312 | n/a | 55 | 0.17h | not exported |
| 3 | GPT 5.5 (xhigh) | 20260519T215046Z |
Codex /goal | Paper prompt + Goal contract no internet | 43.6% | ok | 136/312 | n/a | 81 | 0.22h | 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
79.2% from 247/312 ProgramBench-scored tests. Raw public eval statuses: failure: 86, passed: 266.
Agent trace summary unavailable.
Likely miss class: behavioral mismatch.
2172e38aeb0eeval.tests.test_autotune.TestAutotune.test_autotune_invalid_metric (failure)2172e38aeb0eeval.tests.test_error_handling.TestArgumentValidation.test_quantize_with_unsupported_model_type (failure)2172e38aeb0eeval.tests.test_training.TestSupervisedTraining.test_supervised_invalid_loss_shows_error (failure)2172e38aeb0eeval.tests.test_quantize_dump.TestQuantize.test_quantize_creates_ftz_file (failure)2172e38aeb0eeval.tests.test_training.TestSupervisedTraining.test_supervised_help_flag (failure)2172e38aeb0eeval.tests.test_training.TestTrainingArguments.test_ova_loss_function (failure)
GPT 5.5 (xhigh) · 20260518T232006Z
Public eval evidence has not been exported for this row yet.
GPT 5.5 (xhigh) · 20260519T215046Z
Public eval evidence has not been exported for this row yet.
Official ProgramBench Results by Model
| # | Model | Provider | Score | Cost | Calls |
|---|---|---|---|---|---|
| 1 | GPT 5.5 (xhigh) | OpenAI | 80.4% | $8.25 | 80 |
| 2 | Claude Opus 4.6 | Anthropic | 75.6% | $15.58 | 242 |
| 3 | Claude Sonnet 4.6 | Anthropic | 68.6% | $13.78 | 294 |
| 4 | GPT 5.5 (high) | OpenAI | 62.8% | $2.49 | 27 |
| 5 | GPT 5.5 | OpenAI | 61.5% | $1.02 | 14 |
| 6 | Claude Opus 4.7 | Anthropic | 38.8% | $1.59 | 57 |
| 7 | GPT 5.4 | OpenAI | 37.5% | $0.24 | 9 |
| 8 | Claude Opus 4.7 (xhigh) | Anthropic | 26.9% | $2.35 | 60 |
| 9 | Gemini 3 Flash | 15.4% | $0.22 | 47 | |
| 10 | Claude Haiku 4.5 | Anthropic | 14.7% | $1.02 | 106 |
| 11 | GPT 5 mini | OpenAI | 8.3% | $0.01 | 8 |
| 12 | Gemini 3.1 Pro | 5.4% | $1.54 | 63 | |
| 13 | GPT 5.4 mini | OpenAI | 0.3% | $0.02 | 11 |