What OpenAI system cards contain and why this one matters
OpenAI publishes a system card before each major model deployment. These documents report benchmark scores, safety risk ratings, red-teaming results, and refusal calibration decisions. The GPT-5.5 system card follows the same format established by prior releases like the GPT-4o system card.
WARNING
Source verification: The GPT-5.5 system card URL (openai.com/index/gpt-5-5-system-card) is on an official OpenAI domain, but this article was written close to the document’s publication date. If the URL does not resolve when you click it, the document may have been moved, renamed, or temporarily removed. This guide describes the structure and categories present in OpenAI system cards based on their established format — it does not quote specific scores or ratings. Go to the source document directly for all numerical data.
This article is not a summary of the system card’s findings. It is a guide to reading the system card efficiently — which sections to prioritize based on your role, what to extract, and how to act on what you find.
Where to find the numbers that matter
OpenAI system cards typically report evaluation results across coding, reasoning, math, knowledge, and multilingual benchmarks. The GPT-4o system card, for example, included scores on HumanEval, GPQA, MATH, and MMLU alongside prior-model baselines. The GPT-5.5 system card follows this pattern.
Here is what to look for and why:
| What to Find in the System Card | Where to Look | What to Do With It |
|---|---|---|
| Coding benchmark scores (e.g., HumanEval, SWE-bench) | Benchmark comparison tables, usually early in the document | Compare against your internal GPT-4o eval results on the same benchmarks — if the system card’s improvement is <5% on your target task category, the upgrade may not justify migration cost |
| Reasoning/math scores (e.g., GPQA, MATH) | Same comparison tables | Check whether multi-step reasoning improved enough to affect your chain-of-thought or planning prompts |
| Hallucination and factuality metrics | Factual accuracy or calibration section | Extract the hallucination rate delta from the prior model — this determines whether you can reduce output verification steps |
| Multilingual performance | Language-specific evaluation subsection | Only relevant if your application handles non-English input; skip otherwise |
NOTE
System card benchmarks are run under controlled prompting conditions — typically zero-shot or standardized few-shot. Your production setup (custom system prompts, tool-use schemas, multi-turn context) will produce different results. Treat system card scores as a ceiling estimate, not a deployment guarantee. Always re-run your own eval suite before migrating.
How to read the Preparedness Framework ratings
OpenAI’s Preparedness Framework assigns risk levels — low, medium, high, or critical — across threat domains. OpenAI’s stated deployment policy requires all categories to remain at or below “medium” for a model to ship. The GPT-4o system card rated all categories at medium or below; the GPT-5.5 system card applies the same framework.
The four risk domains to check:
- Cybersecurity: Does the model provide meaningful uplift for offensive cyber operations beyond publicly available tools?
- Biological/chemical threats: Does the model offer actionable synthesis guidance beyond what standard search returns?
- Persuasion: Can the model generate manipulative content at scale that exceeds baseline human capability?
- Model autonomy: Can the model operate independently, acquire resources, or resist shutdown?
When you open the system card, go to the Preparedness Framework section and note the specific rating for each domain. If any category moved up from GPT-4o’s rating, that shift signals increased capability in a sensitive area — and may affect your compliance posture or acceptable-use policies.
The agentic evaluation section: what to extract for agent builders
For teams building AI coding agents or automated pipelines, the agentic capability evaluation is the highest-priority section. Based on the structure of prior OpenAI system cards, this section typically covers:
- Multi-step task execution without human checkpoints
- Tool use and function calling behavior with external APIs or code execution
- Autonomous resource acquisition — whether the model attempts to exceed granted permissions
- Instruction following under ambiguity — behavior when directives conflict or are underspecified
The specific failure modes documented in this section are the most actionable content in the entire system card. Each failure mode maps to a guardrail requirement in your agent architecture.
TIP
Open the agentic evaluation section and make a numbered list of every failure mode OpenAI’s red team identified. For each one, check whether your current agent pipeline has a corresponding checkpoint, fallback, or constraint. Any gap is a pre-deployment fix. OpenAI’s internal testers already found these weak points — use their findings instead of discovering the same failures in production.
Refusal calibration: the silent breakage risk
Refusal boundary changes are the most common cause of unexpected failures during model upgrades. The system card’s refusal calibration section documents which request categories GPT-5.5 declines, and how those boundaries shifted from the prior model.
Three things to extract:
- Security-related code generation boundaries — under what conditions does the model refuse to generate exploit code, penetration testing scripts, or vulnerability analysis?
- Hardcoded vs. context-dependent refusals — which content categories are always blocked, and which depend on framing or system prompt context?
- Refusal rate delta from GPT-4o — did over-refusal decrease (fewer false blocks on legitimate requests) or did under-refusal decrease (more blocks on harmful requests)?
If you have existing prompts that work with GPT-4o and touch security, medical, legal, or other sensitive domains, cross-reference them against the refusal changes before your first GPT-5.5 API call.
Who should read which section first
- API developers migrating from GPT-4o: Start with refusal calibration. Identify every category where the boundary moved. Test your existing prompts against the new model before deploying.
- Agent architecture teams: Start with the agentic evaluation. Catalog failure modes. Map each to your pipeline’s checkpoint design.
- Security-focused prompt engineers: Start with the red-teaming residual risks. These describe what OpenAI knows remains exploitable after mitigation — directly relevant if your application handles code security or access control.
- Engineering leads evaluating the upgrade: Start with the benchmark comparison tables. Compare system card scores on your most relevant benchmarks against your internal GPT-4o results.