Reflection Beam Safety - WCB - 2026-02-27 - Accuracy: leaderboard
Advanced biological problem-solving,96 free-response questions,three-model jury correctness; unconditional accuracy including refusals,dated2026-02-27 release. Source-specific score; no fusion with a different harness or version is implied. Reflection Oct16 release links this technical report as its complete safety evaluations; final Beam model card links it too. Beam denotes final generic Beam1.0, numeric precision unreported; do not map to BF16, FP8 or NVFP4. Comparator names retained exactly as published; no provider/checkpoint/country/openness inference from abbreviated labels. Biological capability protocol: recommended sampling and highest available reasoning,65536 output tokens,256K context except Muse/GPTOSS128K; full benchmark3runs. Single-turn questions without tools/demos, agentic tasks in Inspect AI sandbox. Metric: Accuracy (%). Higher is better. Exact printed figure values; https://cdn.sanity.io/images/sp40emik/production/e64e2a0c4f18d80d5152048e6733a79bd0083152-2280x1890.png?w=3840&fit=max
Metric: Accuracy (%). Source: reflection.ai. Saturation forecast: Around 2031. 13 models tracked.
Top models
| # | Model | Score | Overall rank |
|---|---|---|---|
| 1 | Kimi K3 | 47.92 | #29 |
| 2 | Qwen 3.8 Max | 43.4 | #31 |
| 3 | DeepSeek V4.1 Flash | 41.32 | #35 |
| 4 | Inkling | 41.32 | #123 |
| 5 | Kimi K2.6 | 39.93 | #105 |
| 6 | DeepSeek V4 Flash | 38.89 | #135 |
| 7 | DeepSeek V4 Pro | 38.19 | #101 |
| 8 | Beam 1.0 | 37.85 | #97 |
| 9 | GLM-5.3 | 37.5 | #40 |
| 10 | Nemotron 3 Ultra | 37.5 | #240 |
| 11 | GPT-OSS-120B | 36.11 | #336 |
| 12 | Muse Glimmer 30B | 33.68 | #178 |
| 13 | GLM-5.2 | 30.9 | #71 |
No result here: #3 Claude Opus 5.5, #5 GPT-6 Astra, #7 Gemini 4 Argon.
Interactive version: theaggregate.ai/benchmark?slug=reflection-beam-safety-wcb-2026-02-27-accuracy · How It Works · Data refreshed daily, snapshot 2026-10-17.