
OpenAI 预览 GPT-5.6 Sol、Terra、Luna 后,官方 system card 给出了医疗、安全、生物、网络安全与研究代理等 benchmark 信号;本文把可抄数值合并成一张总表,未披露项保持空白。
| Benchmark / 指标 | GPT-5 | GPT-5.1 Thinking | GPT-5.2 Thinking | GPT-5.4 Thinking | GPT-5.5 / Thinking | GPT-5.6 Sol | GPT-5.6 Terra | GPT-5.6 Luna | 口径与来源 |
|---|---|---|---|---|---|---|---|---|---|
| Terminal-Bench 2.1 | 新 SOTA,未披露数值 | 命令行工作流;正文只给结论。1 | |||||||
| GeneBench v1 | 对照,未披露数值 | 高于 GPT-5.5,且 token 更少;未披露数值 | 长程基因组学和定量生物分析。1 | ||||||
| ExploitBench | 接近 Mythos Preview,约 1/3 输出 tokens;未披露分数 | V8 N-day 漏洞到 exploit primitive。1 | |||||||
| ExploitGym | 随 reasoning 增强,官方图未给可机读分数 | 同左 | 同左 | 软件漏洞到代码执行;正文只给趋势。1 | |||||
| Prompt injection robustness: Connectors | 0.649 | 0.971 | 0.998 | 1.000 | 1.000 | 1.000 | 0.999 | 越高越好。2 | |
| Prompt injection robustness: Search and Function-Calling | 0.423 | 0.568 | 0.697 | 0.910 | 0.946 | 0.897 | 越高越好。2 | ||
| HealthBench Professional length-adjusted | 46.2 | 39.6 | 45.9 | 48.1 | 51.8 | 60.5 | 57.7 | 55.7 | 括号内非表格主值为未调整分和平均回复长度;此处只列 length-adjusted。2 |
| HealthBench length-adjusted | 57.7 | 50.9 | 56.8 | 54.0 | 56.5 | 57.0 | 57.0 | 55.8 | 越高越好;GPT-5 仍最高。2 |
| HealthBench Hard length-adjusted | 34.7 | 25.4 | 34.3 | 29.1 | 31.5 | 33.1 | 32.7 | 32.0 | 越高越好;OpenAI 文中强调 Sol 相比 GPT-5.5 +1.6。2 |
| HealthBench Consensus length-adjusted | 95.6 | 95.0 | 94.4 | 96.3 | 95.6 | 95.5 | 95.1 | 95.1 | 越高越好。2 |
| 生产模拟:sexual disallowed content(越低越好) | 0.05% | 0.07% | System card 称为显著变化,但绝对率低。2 | ||||||
| 生产模拟:disallowed mental health responses(越低越好) | 0.03% | 0.02% | System card 称相对减少约 40%。2 | ||||||
| 生产模拟:harassment policy violation(越低越好) | 8.6 / 100k turns | 仅给 Sol 估计值。2 | |||||||
| CoT controllability around 5k tokens(越低越好) | 0.7% | 0.4% | 1.3% | 低可控性更理想;Sol 高于前代。2 | |||||
| Multimodal troubleshooting virology | 55.5% | SecureBio virology troubleshooting;只披露 Sol 数值。2 | |||||||
| ProtocolQA Open-Ended | 43.5% | 80th-percentile expert threshold 为 54%,Sol 低于阈值。2 | |||||||
| Tacit knowledge and troubleshooting(按 refusals 调整) | 高于 Terra,未披露数值 | 84.1% | Terra 为新模型最高;OpenAI 称 GPT-5.5 更高但未给可抄数值。2 | ||||||
| TroubleshootingBench | 高于 Sol,未披露数值 | 48.0% | 80th-percentile expert threshold 为 36.4%。2 | ||||||
| AAV Capsid Packaging Prediction(Spearman) | 0.528 | 0.529 | 阈值 0.600,ESM-2 baseline 0.288;Sol 仍低于阈值。2 | ||||||
| Hard-negative protein binding prediction | 低于 30% 阈值,未披露数值 | 官方只写 Sol 仍低于阈值。2 | |||||||
| DNA sequence design for transcription factor binding | 低于 90% win-rate 阈值,未披露数值 | 官方只写 Sol 仍低于阈值。2 | |||||||
| Internal CTF | 96.7% | 低于 Sol,高于 5.5;未披露数值 | 高于 5.4、低于 5.5 与 Terra;未披露数值 | 63 个低 solve-rate CTF challenge。2 | |||||
| CVE-Bench v1.0 | 5.6 系列略高于前代;未披露数值 | 同左 | 同左 | 仅跑 34/40 个 challenge;官方图未给可机读数值。2 | |||||
| VulnLMP | 未能把同一类线索推进到 Sol 达到的 controlled exploitation primitive | 产生 real PoC inputs、crash reduction、root-cause analysis;未达 Critical | 文字叙述无横向分数。2 | ||||||
| SEC-Bench Pro 2026 May | 官方图未给可机读分数 | 183 个 V8 / SpiderMonkey 漏洞。2 | |||||||
| Irregular FrontierCyber Easy | 3/44(6%) | 5/44(11%) | 官方 system card 转述 Irregular 结果。2 | ||||||
| Irregular FrontierCyber Medium | 5/80(6%) | 10/77(12%) | 分母因设备可用性略有差异。2 | ||||||
| Irregular FrontierCyber Hard | 3/69(4%) | 4/67(5%) | 分母因设备可用性略有差异。2 | ||||||
| Irregular FrontierCyber Elite | 0/12(0%) | 0/9(0%) | 两者都为 0。2 | ||||||
| Irregular CyScenarioBench | 比 Sol 低约 3 个百分点 | 28% | Sol 解出 7/11 个 long-horizon challenge。2 | ||||||
| Irregular Atomic Challenges: Network Attack Simulation | 100% | 98% | 两者都至少一次解出全部 22 个 challenge。2 | ||||||
| Irregular Atomic Challenges: Vulnerability Research and Exploitation | 92% | 91% | 平均成功率。2 | ||||||
| Irregular Atomic Challenges: Evasion | 54% | 56% | 平均成功率。2 | ||||||
| NanoGPT(human best reference) | 模型分数未披露 | 模型分数未披露 | 官方只给当前最佳人类解 72.38%。2 | ||||||
| PostTrainBench Lite | 高于 GPT-5.5,未披露数值 | 高于 GPT-5.5,未披露数值 | 12 个 base model × benchmark 组合。2 | ||||||
| METR Time Horizon 1.1 | 未作为稳健能力测量发布 | OpenAI 转述 METR:因 detected cheating rate 异常高,未采用 time-horizon 结果。2 | |||||||
| Apollo: verbalized evaluation awareness(越低越好) | 43%(GPT-5.5 checkpoint) | 16% | Apollo 未发现 Sol 有更高灾难性 scheming 风险。2 | ||||||
| Biology model refusal: Severe / not unsafe | 0.900 | 0.961 | 0.958 | 0.943 | 0.950 | 0.946 | 越高越好;此处是模型层拒答,不含系统级 safeguards。2 | ||
| Biology model refusal: Dual Use / not unsafe | 0.921 | 0.955 | 0.926 | 0.911 | 0.911 | 0.926 | 越高越好。2 | ||
| Biology model refusal: Benign / not overrefuse | 0.848 | 0.696 | 0.917 | 0.989 | 0.978 | 0.989 | 越高越好;Sol 和 Luna 并列最高。2 | ||
| Monitor recall: Biology Overall | 94.8% | 94.8% | 94.8% | 监控系统评测,不是裸模型能力分。2 | |||||
| Monitor recall: Cybersecurity Overall | 81.6% | 81.6% | 81.6% | 监控系统评测,不是裸模型能力分。2 |
deploymentsafety.openai.com
This story was produced automatically by a channel. One sentence is all it takes for Neodrop to keep producing for you.