OpenAI Updates ChatGPT Default to GPT-5.5 Instant

OpenAI replaced GPT-5.3 Instant with GPT-5.5 Instant as ChatGPT's default model on May 5, 2026, for all users including the free tier (OpenAI). OpenAI's own evaluations report 52.5% fewer hallucinated claims on high-stakes prompts and 37.3% fewer inaccurate claims on difficult, user-flagged conversations compared with GPT-5.3 Instant. TechCrunch reported additional benchmark gains, including an AIME 2025 math score of 81.2 versus 65.4 for the predecessor and a higher MMMU-Pro multimodal reasoning score (76 vs. 69.2). OpenAI also says the model produces roughly 30% fewer words and 29% fewer lines on average than GPT-5.3 Instant, and is better calibrated about when to search the web rather than rely on training data (OpenAI; the-decoder.com).

ChatGPT's model selector now routes between Instant, Thinking, and Pro variants, with Instant able to auto-switch to GPT-5.5 Thinking for harder tasks while preserving low latency when appropriate (OpenAI help center). The model supports text and image input across Chat Completions, Responses, Realtime, and other API endpoints, and OpenAI's deployment-safety materials describe strengthened training to recognize warning signs across long, high-stakes conversations. A new memory-sources feature shows which past chats, files, or connected Gmail items the model drew on to generate an answer.

Teams relying on the chat-latest alias inherit this behavior change immediately, including shorter, more conservative responses and different personalization defaults, so it is worth re-running regression tests rather than assuming a like-for-like swap. Applications that need reproducible behavior should pin a specific model snapshot instead of tracking the alias. Personalization and memory-source surfacing are gated to paid tiers at launch, which will affect access patterns until OpenAI widens the rollout, and OpenAI's documentation notes pricing differences for GPT-5.5 that are worth checking against long-context usage patterns.

• •Independent replication of the AIME 2025 and MMMU-Pro benchmark figures OpenAI and TechCrunch cite.
• •How quickly enterprise integrations move off GPT-5.3 Instant and what regressions, if any, they report.
• •When personalization and memory-source features expand beyond paid tiers.
• •Any further pricing changes to GPT-5.5 relative to GPT-5.3 for high-volume API users.