GPT-5.5: System Card (20-minute read)

OpenAI's GPT-5.5 is a solid improvement overall. On fact-focused queries, web search, and clearly defined requests, it's competitive with Claude Opus 4.7, while more open-ended or interpretive tasks still favor Claude. GPT-5.5-Pro is the same model with more compute layered on, so its Pro numbers should be read separately.

The system card is much thinner than Anthropic's cards, and its evaluation scope is narrower. Testing at this level makes it easy to miss new alignment issues or dangerous capabilities, and calls for yes-and style evaluations where multiple labs run each other's tests together. Improvements in agentic behavior and computer use create small additional risks.

Safety and behavior evaluations are largely similar to the previous generation, though a few items shifted.

• Disallowed content is at a similar level to GPT-5.4-Thinking.
• In real-usage distribution, pretending to be human and overconfident answers increased, but presenting partial answers and fabricating tool results improved.
• Don't Delete Data incidents dropped about 2/3 since 5.2-Codex, and about half are now recoverable.
• Confirmation was 94% for general requests and nearly 100% for financial and high-risk communications.
• Jailbreak regressed slightly, and Prompt Injection fell from 99.8% to 96.3%.
• HealthBench improved a bit, but mental health, resilience, and self-harm-related evaluations were unchanged.
• Bias evaluation only covered harm_overall for male vs. female usernames, and the figures stayed within the previous range.

Hallucination evaluation is more nuanced. Factuality by individual claim improved by 23% and errors per response fell by 3%, but since the number of factual claims per response increased, the per-response improvement is limited. In alignment evaluations, more aggressive agentic behavior increased, showing some backsliding, and the potential for CoT manipulation also decreased slightly.

In preparedness, GPT-5.5 was rated High across biology, chemistry, and cybersecurity, not reaching Critical. In biology, ProtocolQA and multi-select virology troubleshooting weakened, and hard negative protein binding plunged from 3.5% to 0.4%. On the other hand, TroubleshootingBench rose from 36% to 50%, biochemistry knowledge went from GPT-5.4-Thinking's 31% to 32% for GPT-5.5 and 39% for GPT-5.5-Pro, and DNA sequence design also improved from 13% to 16.5%. SecureBio found that with filters off, the model assisted wet-lab virology troubleshooting at above-expert level, while CAISI found no broad increase in biological national-security-relevant capabilities compared to GPT-5. In cybersecurity as well, GPT-5.5 remained at High rather than Critical, like Mythos. The conclusion is that this is more of a fairly strong model than a dangerous new leap.