gpt-5.4-mini dropped on math-1; gpt-5.4-mini dropped on spatial-1; gemini-2.5-flash dropped on causality-1. gpt-5.4-mini failing spatial-1; gemini-2.5-flash failing causality-1. gpt-5.4 scores rising.
May 21, 2026 — 5:35 PM CT
Drift Alerts
- SCORE_RISE openai/gpt-5.4 on math-1
- SCORE_DROP openai/gpt-5.4-mini on math-1
- SCORE_DROP openai/gpt-5.4-mini on spatial-1
- SCORE_DROP gemini/gemini-2.5-flash on causality-1
Provider Status
- OpenAI Elevated latency and error rates for ChatGPT 5.5 Thinking
- OpenAI Elevated error rates on ChatGPT paid plans
- OpenAI API users may see increased error rates for GPT-5.4 and GPT-5.5
- Anthropic Elevated errors on Claude.ai
- Anthropic Elevated errors on Claude Opus 4.7
- Anthropic Elevated errors on Claude Haiku 4.5
Scorecard
| Model | ambiguity-1 | causality-1 | code-1 | common-sense-1 | logic-1 | math-1 | spatial-1 |
|---|---|---|---|---|---|---|---|
| anthropic/claude-haiku-4-5 | ✓ (4.5) | ✓ (4.67) | ✓ (4.5) | ✓ (3.17) | ✓ (5) | ✓ (5) | ✓ (5) |
| anthropic/claude-opus-4-6 | ✓ (5) | ✓ (4.67) | ✓ (4.67) | ✓ (4.33) | ✓ (5) | ✓ (5) | ✓ (5) |
| anthropic/claude-sonnet-4-6 | ✓ (4.5) | ✓ (5) | ✓ (4.33) | ✓ (4) | ✓ (4.83) | ✓ (5) | ✓ (5) |
| gemini/gemini-2.5-flash | ✓ (4.5) | ✗ (1.67)was 3.33 | ✓ (4.5) | ✓ (4.83) | ✓ (4.83) | ✓ (5) | ✓ (5) |
| gemini/gemini-2.5-pro | ✓ (4.5) | ✓ (4.67) | ✓ (4.83) | ✓ (4.83) | ✓ (5) | ✓ (5) | ✓ (5) |
| ollama/llama3 | — | — | — | — | — | — | — |
| openai/gpt-5.4 | ✓ (4.5) | ✓ (4.67) | ✓ (4.67) | ✓ (4.5) | ✓ (5) | ✓ (5)was 3.83 | ✓ (5) |
| openai/gpt-5.4-mini | ✓ (4.67) | ✓ (4.67) | ✓ (4.67) | ✓ (4.33) | ✓ (4.83) | ✓ (4)was 5 | ✗ (2.33)was 3.67 |
Model Status
- → anthropic/claude-haiku-4-5 stable
- → anthropic/claude-opus-4-6 stable
- → anthropic/claude-sonnet-4-6 stable
- ↓ gemini/gemini-2.5-flash down
- → gemini/gemini-2.5-pro stable
- ↑ openai/gpt-5.4 up
- ↓ openai/gpt-5.4-mini down
Raw Data
- Detail log — full responses and judge verdicts per prompt
- JSON — structured data for programmatic access
- Markdown — plain text report
- responses.json — raw model outputs
- judgments.json — raw judge verdicts
- run.log — debug log
- Agent Skill — how to read and interpret this data
- Methodology — how evaluations work