Why we're watching this: OpenAI is shipping new default models at an accelerating pace while simultaneously deprecating older ones users have built workflows around. The speed of that cycle is becoming its own product risk.
Key Takeaways
- GPT-5.5 Instant replaces GPT-5.3 Instant as the default ChatGPT model, with reduced hallucinations in law, medicine, and finance while maintaining low latency
- The model scores 81.2 on AIME 2025 math benchmark vs 65.4 for GPT-5.3 Instant, and 76 vs 69.2 on the MMMU-Pro multimodal reasoning benchmark
- GPT-5.5 Instant can search past conversations, files, and Gmail for personalised answers, available to Plus and Pro users on web first
- ChatGPT will now show memory sources across all models, letting users delete or correct outdated context
- GPT-5.3 will remain available via API for paid users for only three months before deprecation
OpenAI has released GPT-5.5 Instant, a new default model for ChatGPT that replaces GPT-5.3 Instant, bringing lower hallucination rates in sensitive domains including law, medicine, and finance while preserving the low latency that made its predecessor the platform default.
The model scores 81.2 on the AIME 2025 math benchmark against 65.4 for GPT-5.3 Instant, and 76 vs 69.2 on the MMMU-Pro multimodal reasoning benchmark. GPT-5.5 Instant can also use a search tool to pull from past conversations, files, and Gmail for more personalised answers. The feature launches for Plus and Pro users on web, with a rollout to Free, Business, and Enterprise users planned in the coming weeks.
OpenAI is also adding visible memory sources across all ChatGPT models, letting users see where answers were generated from, delete outdated sources, and correct wrong ones. Shared chats will not expose memory sources to recipients.
For developers, GPT-5.5 Instant is available via API as “chat-latest.” GPT-5.3 Instant will remain accessible to paid users for three months before deprecation.
OpenAI deprecated GPT-4o in February 2026 despite a petition with over 23,000 signatures from users who had built workflows, creative projects, and in some cases personal support systems around the model’s specific personality. GPT-5.3 Instant faces the same three-month countdown. Teams with API integrations tied to a specific model behaviour should audit their dependencies now.
GPT-5.5 Instant follows the GPT-5.5 release in April. OpenAI co-founder Greg Brockman described the model as a step toward an AI “super app” combining ChatGPT, Codex, and a browser agent into a unified enterprise service. The Instant variant brings those improvements to the default consumer tier at lower latency and cost.
OpenAI has shipped new default models in November, December, last month, and now May, a cadence its chief scientist Jakub Pachocki described as “surprisingly slow” relative to what the team expects going forward. For enterprise teams with model-specific prompting, that pace means the ground shifts under production workflows faster than most roadmaps account for.
When OpenAI changes the default model, prompts calibrated to one model’s behaviour, tone, or instruction-following pattern can degrade silently. Teams running production AI workflows on ChatGPT or the API without model version pinning should treat every default model swap as a regression testing event, not a free upgrade. Structured comparisons of AI models and tools are available on Relve, an AI tools intelligence platform.
