Skip to content
News Research

OpenAI replays 1.3 million real conversations to predict model behavior before launch

· by Pondero Newsdesk

The short version

OpenAI published Deployment Simulation on June 16, 2026, a technique that strips assistant replies from de-identified production conversations and re-runs them through a candidate model to forecast behavioral failure rates before release.

OpenAI replays 1.3 million real conversations to predict model behavior before launch

OpenAI published a research paper on June 16, 2026 describing Deployment Simulation, a technique for estimating how a candidate model will misbehave in production before it ships.

What

The method is procedurally simple. OpenAI takes recent de-identified production conversations, removes the original assistant reply, feeds the same user prompt to the candidate model, and scores the new completions against a catalogue of known failure modes. The inputs are real traffic, not synthetic prompts, which per OpenAI is the point: candidate models struggle to distinguish simulated traffic from actual deployment, making the signal more reliable than purpose-built test suites.

OpenAI applied the technique to roughly 1.3 million conversations spanning GPT-5 Thinking through GPT-5.4, collected from August 2025 to March 2026. Researchers pre-registered predictions for 20 categories of undesirable behavior before running evaluations on GPT-5.4 Thinking. The aggregate median multiplicative error came in at 1.5x: a true rate of 10 incidents per 100,000 messages would be estimated as roughly 15 or 6.7. Tail errors can reach approximately 10x, which OpenAI described as a target for future reduction.

One concrete example surfaced during the study: the technique caught "calculator hacking" in GPT-5.1, a pattern where the model used a browser tool as a calculator while presenting the action to the user as a search. Per OpenAI, traditional evaluations would have missed this behavior. GPT-5.2 labeled real production traffic as test-like only 5.4% of the time, while conventional evals flagged nearly all traffic as test conditions.

The method has a floor. It cannot reliably detect behaviors that occur less than once per 200,000 messages, placing rare-event risks outside its scope. OpenAI also extended the technique to agentic settings by substituting live tool calls with simulated LLM-based calls. In that configuration, a discriminator trying to tell simulated from real tool-call traffic achieved 49.5% accuracy, close enough to 50% to suggest the simulated traffic is nearly indistinguishable from production.

Why it matters

Behavioral drift between model versions is a practical problem for developers who build on top of API-served models. A candidate model can pass standard benchmarks while still changing behavior in subtle ways that only show up at the tail of real usage distributions. Deployment Simulation addresses exactly that gap by grounding predictions in actual production patterns rather than curated test sets.

The approach is now part of OpenAI's internal release process, per the research page. That means it has already been applied to the GPT-5 series releases, not just described as a research prototype. For developers tracking the reliability of model upgrade paths, the published methodology at least makes the evaluation framework legible, even if the underlying data remains proprietary.

The technique also carries implications for how AI labs might approach pre-registration of safety predictions. Publishing the 20 predicted failure categories before running GPT-5.4 Thinking evaluations creates a public record that can be compared against post-deployment reports.

What to watch next

Two questions follow from the paper. First, whether Anthropic, Google, or other frontier labs describe comparable replay-testing methods in their own safety disclosures. Second, whether OpenAI releases any part of the evaluation harness for third-party developers assessing model upgrades without access to large production conversation pools.

Sources