Published: May 6, 2026
OpenAI began rolling out GPT-5.5 Instant on May 5, 2026, replacing GPT-5.3 Instant as the default model behind ChatGPT for all users, including those on the free tier. The company claims a 52.5% reduction in hallucinated claims on high-stakes prompts covering medicine, law, and finance, alongside measurable gains in math, science, and multimodal reasoning benchmarks. GPT-5.5 Instant is the lighter, general-purpose sibling of GPT-5.5 Thinking and GPT-5.5 Pro, both of which were announced on April 23, 2026.
How does GPT-5.5 Instant compare to GPT-5.3 Instant on benchmarks?
OpenAI reports consistent improvements across several public benchmarks. On AIME 2025, a competition-level math test, GPT-5.5 Instant scores 81.2 compared to 65.4 for GPT-5.3 Instant. On GPQA, a PhD-level science benchmark, the new model lands in the mid-80% range. The multimodal MMMU-Pro test yields 76 for GPT-5.5 Instant versus 69.2 for its predecessor.
Health-related benchmarks also show gains. HealthBench rises to 51.4 from 49.6, while HealthBench Professional, which tests clinical reasoning, jumps to 38.4 from 32.9. OpenAI skipped the GPT-5.4 version number entirely; GPT-5.3 Instant had launched only about two months earlier, in March 2026.
What changed in hallucination and accuracy rates?
The headline safety figure is a 52.5% reduction in hallucinated claims on high-stakes prompts, specifically those involving medicine, law, and finance. OpenAI also measured a 37.3% reduction in inaccurate claims across conversations that users had previously flagged for errors. These numbers come from internal evaluations, and independent testing will take time to confirm.
The model also produces shorter responses on average: roughly 30% fewer words and 29% fewer lines than GPT-5.3 Instant. OpenAI says the model is less likely to ask unnecessary follow-up questions and less prone to inserting what the company calls “gratuitous emojis” into its replies.
What are the new personalization and search features?
GPT-5.5 Instant introduces deeper personalization. The model can pull context from a user’s past conversations, saved files, and connected Gmail accounts. When it uses stored information, it now shows the memory source so users can verify what data informed the response. For users who prefer not to have their history referenced, OpenAI has added a “temporary chat” mode that opts out of personalization entirely.
Web search triggering has also been updated. The model is reportedly better at deciding when to search the web for current information rather than relying on its training data. Image and photo analysis has been improved as well, along with STEM-related answers more broadly.
Who gets access and what does the API cost?
GPT-5.5 Instant is rolling out globally starting May 5, 2026 to all ChatGPT users, both free and paid. Paid subscribers who prefer the previous model can still select GPT-5.3 Instant for a three-month transition period.
For developers, the GPT-5.5 API was announced on April 24, 2026, priced at $5 per million input tokens and $30 per million output tokens. The API supports a context window of one million tokens. This positions GPT-5.5 Instant as the intended daily driver for both consumer and developer use cases, sitting below the more capable GPT-5.5 Thinking and GPT-5.5 Pro models in the product lineup.
For full details, see OpenAI’s official announcement at https://openai.com/index/gpt-5-5-instant/.
