OpenAI's GPT-5.5 System Card Details Safety Push and Pro Variant

OpenAI has published the system card for GPT‑5.5, a new model the company describes as designed for "complex, real-world work" — writing code, researching online, analyzing information, creating documents and spreadsheets, and moving across tools to get things done.

The document, updated on April 24, 2026, to include additional information about safeguards for deploying GPT‑5.5 and GPT‑5.5 Pro through the API, lays out both the model's intended scope and the predeployment safety process OpenAI ran before release. The stakes are straightforward: the model is built to act with less human guidance across multiple tools, which raises the bar for the safeguards wrapped around it.

What OpenAI says GPT‑5.5 does

The system card draws a direct comparison with earlier models. According to OpenAI, GPT‑5.5 "understands the task earlier, asks for less guidance, uses tools more effectively, checks it work and keeps going until it's done."

Each phrase in that sentence describes a shift toward autonomy. Understanding a task earlier means fewer clarification rounds between the user and the model. Asking for less guidance means the model fills in gaps on its own. Using tools more effectively and checking its own work — and continuing "until it's done" — describe a system intended to carry multi-step jobs through to completion rather than stopping partway and handing the work back.

That positioning matters because it targets exactly the workloads where AI systems are moving from assistants that answer questions to agents that execute them: codebases, research pipelines, document production, spreadsheet analysis, and workflows that span several software tools. Those are also the deployments where errors compound, since a mistake made in step one propagates through everything downstream. OpenAI's claim that the model checks its own work is its answer to that concern.

The evaluation process

Before release, OpenAI says it subjected GPT‑5.5 to its "full suite of predeployment safety evaluations" and its Preparedness Framework, the company's internal process for assessing catastrophic risks from frontier models. The evaluations included targeted red-teaming for advanced cybersecurity and biology capabilities — two risk categories that feature prominently in current AI policy debates, because they map directly onto concerns about models lowering the barrier to offensive cyber operations or biological threat creation.

The company also collected feedback on real use cases from nearly 200 early-access partners before release. That number is one of the few concrete figures in the card, and it signals the scale of external validation OpenAI ran: this was not a lab-only assessment but a process that drew on external organizations working with the model in production-like conditions ahead of general availability.

OpenAI summarizes the outcome in one line: "We are releasing GPT‑5.5 with our strongest set of safeguards to date, designed to reduce misuse while preserving legitimate, beneficial uses of advanced capabilities."

That framing — reduce misuse while preserving legitimate use — echoes the central tension in current safeguard design. Restrictive safety layers can blunt a model's usefulness for defensive security research, biology, and other sensitive but legal domains. OpenAI's stated goal is to keep those beneficial uses intact while narrowing the paths to abuse.

GPT‑5.5 Pro and test-time compute

The system card also introduces GPT‑5.5 Pro, and the relationship between the two variants is technical rather than architectural. In OpenAI's words, GPT‑5.5 Pro "is the same underlying model using a setting that makes use of parallel test time compute."

In practice, that means Pro spends more compute at inference time — running work in parallel and selecting among results — to improve output quality on hard problems, without being a separately trained model.

This shared foundation shapes how OpenAI handles safety analysis. "We generally treat GPT‑5.5's safety results as strong proxies for GPT‑5.5 Pro," the company states. The reasoning is explicit: since the underlying model is the same, its capabilities and risk profile should largely carry over.

But not entirely. OpenAI carves out an exception: "we separately evaluate GPT‑5.5 Pro in certain cases because we judge that the setting could materially impact the relevant risks or appropriate safeguards posture." More parallel test-time compute can, in principle, let a model solve problems it would otherwise fail — which means the inference setting itself can change the risk picture even when the weights are identical. OpenAI's decision to evaluate Pro separately in those cases acknowledges that inference-time scaling is a safety variable, not just a performance knob.

How to read the evaluation results

One methodological note in the card deserves attention. "Except where noted, the results in system cards describe evaluations we ran in an offline setting," OpenAI writes.

That means the safety results reflect controlled, offline testing rather than measurements taken from live deployment with real users and real tools. OpenAI flags this openly, and readers comparing numbers across system cards should keep it in mind: offline evaluations are the industry's standard predeployment instrument, but the gap between offline scores and in-the-wild behavior is a known limitation that deployment monitoring is expected to cover.

The April 24 update

The card carries an update stamp: "This card was updated on April 24, 2026, to include additional information about safeguards for the deployment of GPT‑5.5 and GPT‑5.5 Pro in the API."

The detail signals that the API deployment path — where developers build GPT‑5.5 and GPT‑5.5 Pro into their own products — warranted additional safeguard documentation beyond what covered the original release. For enterprise and developer audiences, API-specific safeguard information is often the most operationally relevant part of a system card, since it defines what protections apply when the model runs inside third-party systems rather than OpenAI's own interfaces.

Why the system card format matters

System cards have become a standard artifact of frontier AI releases. They serve three audiences at once: regulators and policymakers tracking whether developers are evaluating the risks they care about; enterprise buyers who need documentation of safety posture before procurement; and researchers who use the evaluations to compare practices across labs.

For GPT‑5.5, the card's most consequential disclosures are procedural rather than numerical in this section: the model went through the full Preparedness Framework, was red-teamed specifically for advanced cyber and biology capabilities, and was exercised by nearly 200 early-access partners before release. Those steps — not marketing claims about capability — are what external observers can actually hold the company to.

The card also continues a trend worth watching: safety analysis that tracks inference-time scaling. As labs extract more capability from fixed models through test-time compute, the safety story can no longer attach only to training runs. OpenAI treating GPT‑5.5 Pro's inference setting as a factor that "could materially impact the relevant risks" is an early institutional acknowledgment of that shift.

What comes next

The immediate so-what for developers: both GPT‑5.5 and GPT‑5.5 Pro now have documented, API-specific safeguard information as of the April 24, 2026 update, and organizations building on either variant should read that section before assuming the two are interchangeable from a compliance standpoint. The longer-term question is whether OpenAI's proxy approach — treating one variant's safety results as standing in for the other, with case-by-case exceptions — becomes the standard template as inference-time scaling multiplies the number of deployment configurations built on a single set of weights.