GPT-5.5 Instant becomes ChatGPT's default model — fewer hallucinations, smarter web routing

What changed

OpenAI confirmed on May 5 that GPT-5.5 Instant replaces GPT-5.3 Instant as ChatGPT’s default model. The rollout sequence:

• Plus and Pro users on web — immediate
• Mobile clients — over the following days
• Free tier, Go Business, and Enterprise tiers — “in the coming weeks”
• API access — available as the same model the chat product uses; previous-generation GPT-5.3 remains callable as chat-latest for three months, then retires

No price changes for any ChatGPT tier. The substitution is in-place: the same $20/month Plus subscription you had on May 4 routes to GPT-5.5 Instant on May 5 instead of GPT-5.3 Instant.

The benchmark deltas

OpenAI’s announcement and TechCrunch’s coverage cite three specific numerical improvements over GPT-5.3 Instant:

• AIME 2025 math: 65.4 → 81.2 (+15.8 points)
• MMMU-Pro multimodal reasoning: 69.2 → 76 (+6.8 points)
• Hallucinated claims on high-stakes prompts (medicine, law, finance): approximately 52% fewer per OpenAI’s own evaluation set

The hallucination delta is the most consequential for everyday users. ChatGPT’s defaults have long been “fast but sometimes confidently wrong” relative to Claude, which biases more cautiously and hedges more explicitly. A 52% reduction on the specific category of prompts where being wrong has the most real-world cost — medical, legal, financial — closes a real gap. It does not close all of the gap. Independent benchmark replications and side-by-side use tests will take weeks to filter through. OpenAI’s own numbers are the right starting point, not the ending one.

What didn’t change

A few things to keep in proportion:

Writing quality vs. Claude is the same shape. GPT-5.5 Instant is a faster, more accurate default — but the underlying voice characteristics that make professional writers reach for Claude Opus 4.7 are not addressed by an Instant-tier update. For long-form writing you’ll publish, the Claude advantage on prose voice persists.

No image, audio, or video generation news. This is a model swap inside the existing ChatGPT product surface. Sora is still shut down; DALL-E 3 retired April 2026; the image-generation story for OpenAI users remains DALL-E 4 inside ChatGPT.

Personalization upgrades land alongside the model. OpenAI is also rolling out enhanced personalization from past chats, uploaded files, and connected Gmail to Plus and Pro users. This is a memory/RAG improvement bundled with the model upgrade — useful, separate from the model itself.

GPT-5.5-Cyber is a different product. Two days after Instant rolled out, OpenAI announced a limited-preview cybersecurity-specialized variant for vetted security teams. Not consumer-facing; not available through standard ChatGPT subscriptions.

What this means for ChatGPT users

For free-tier users: Wait. The Free rollout lands “in the coming weeks.” Free users will see the model swap silently when it ships. No action required.

For Plus and Pro users: The default is already updated. Notable behavioral shifts to watch for: tighter, more direct answers (per OpenAI’s framing), better routing decisions on when to search the web, fewer confidently-wrong claims on factual queries. The personalization layer is the bigger UX change for power users.

For API customers: The “chat-latest” alias on GPT-5.3 deprecates in three months. Update integrations that hard-coded GPT-5.3 to GPT-5.5 Instant during that window. Cost economics stay roughly comparable; check the API pricing page for current per-token rates.

For Claude users considering a switch: This update does not move the writing-quality needle. If you switched to Claude for output quality on documents that ship, the reasons that justified that switch remain. ChatGPT’s breadth, ecosystem (Custom GPTs, image generation, voice), and now-stronger default model are still the right reasons to keep an OpenAI subscription as a complement, not a replacement.

The pattern across May 2026 launches

GPT-5.5 Instant lands in a busy stretch. In the same week:

• Anthropic doubled Claude Code rate limits and secured all of SpaceX’s Colossus 1 capacity — adding 300+ MW and 220,000+ GPUs
• Anthropic announced 10 financial-services agent templates with Microsoft 365 deep integration and named JPMorgan, Goldman Sachs, and Citi as deployed customers
• Google DeepMind published the AlphaEvolve impact paper and an “AI co-mathematician” research preview

The pattern: every major lab is shipping smaller, more focused product updates rather than headline frontier-model launches. The era of “GPT-5 stuns the world” is being replaced by quieter, more enterprise-shaped releases that compound advantage in specific domains. GPT-5.5 Instant fits that pattern — useful, real improvement, not a category-redefining moment.

The pragmatic read

GPT-5.5 Instant is a good update that lands the way most 2026 model updates do: real, measurable improvement on specific benchmarks; minimal cost impact for end users; and a noticeable reduction in the worst behaviors of the previous default. The 52%-fewer-hallucinations claim is the most reader-meaningful number in the announcement, and if it holds in independent testing, it’s the right reason to think slightly better of ChatGPT’s reliability on factual queries going forward.

What it isn’t: a reset of the ChatGPT-vs-Claude-vs-Gemini competitive picture. Each lab continues to win on different axes — Claude on writing quality and coding; Gemini on research and Workspace integration; ChatGPT on breadth, ecosystem, and now a meaningfully more reliable default. The take for users paying $20/month somewhere: the question isn’t “which is best at everything” — it’s “which is best at the specific work I do most.” May 5’s update doesn’t change that calculation. It just makes ChatGPT a slightly better answer for users who were already there.

For deeper background, see the ChatGPT review and the head-to-head Claude vs ChatGPT comparison.

Questions or corrections? Email Pick Right. Want the full list? See all news.