GPT-5.5 Instant became the new default ChatGPT model on May 5, 2026, replacing the previous default across Free, Plus, and Pro subscription tiers. For the 900 million weekly active users that OpenAI reported in February 2026, this means that the model they encounter every time they open ChatGPT changed quietly, without a public announcement or launch event, as part of OpenAI’s increasingly rapid default model update cycle. The change matters not just for benchmark numbers — though those are significant — but for a feature that fundamentally alters what ChatGPT can be used for: memory that draws from past conversations, uploaded files, and connected Gmail accounts to personalise responses. GPT-5.5 Instant scored 81.2 on the AIME 2025 mathematics benchmark, compared to 65.4 for its predecessor — a meaningful improvement on a benchmark that tests the kind of multi-step mathematical reasoning required for advanced problem solving in science, engineering, and finance. On MMMU-Pro multimodal reasoning, it scored 76 versus the previous model’s 69.2. These are not incremental improvements — they represent a genuine capability jump for what is now the most widely deployed AI model in the world.
The Memory Feature — Why This Is the Most Important Change
The benchmark improvements in GPT-5.5 Instant are meaningful but predictable. What distinguishes this model update from previous default ChatGPT replacements is the memory architecture. GPT-5.5 Instant can use search to pull from past conversations, files uploaded to ChatGPT, and Gmail to personalise its answers. Memory sources are now visible across all models, giving users the ability to see exactly what ChatGPT remembers about them, delete specific memories, and correct inaccuracies in what the system has retained. This is live for Plus and Pro subscribers on web, with mobile and free-tier rollout planned for coming weeks.
The practical implications of persistent, search-grounded memory are substantial. Previous ChatGPT memory features were limited to a static list of facts the model had been explicitly told to remember. GPT-5.5 Instant’s memory can now search across the full history of a user’s interactions — surfacing relevant past context automatically when it is likely to improve the current response. A user who asked ChatGPT about a project three months ago can now get a response to a new question that takes that prior context into account without re-explaining it. A user who has uploaded their company’s quarterly reports can get analysis that draws on those documents across subsequent sessions. A Gmail-connected user can get email drafting, scheduling, and communication assistance that is aware of their actual communication style and relationships rather than responding to a generic prompt.
“Memory sources are now visible across all models, giving users the ability to delete or correct what ChatGPT remembers. This feature is live for Plus and Pro on web, with mobile and free-tier rollout planned for coming weeks.” — Build Fast with AI, reporting on GPT-5.5 Instant default switch, May 2026
GPT-5.5 Instant vs Previous Default Model — Key Benchmarks
| Benchmark | GPT-5.5 Instant | Previous Default | Improvement | What It Measures |
| AIME 2025 Mathematics | 81.2 | 65.4 | +15.8 points | Multi-step math reasoning — AMC/AIME competition level |
| MMMU-Pro Multimodal | 76.0 | 69.2 | +6.8 points | Understanding images, charts, and diagrams with text |
| AI Intelligence Index | Above 60 (full GPT-5.5) | Below 60 (previous) | Threshold crossed | Composite reasoning across tasks |
| Memory integration | Full — search across history, files, Gmail | Limited static memory | Qualitative leap | Personalisation across sessions |
| Memory transparency | Visible and editable by user | Opaque | Full transparency | User control over retained data |
How GPT-5.5 Instant Differs From the Full GPT-5.5
GPT-5.5 Instant is the efficient, fast version of OpenAI’s GPT-5.5 family — designed for speed and cost-effectiveness at the scale of 900 million weekly users, rather than maximum capability on complex tasks. The full GPT-5.5, released on April 23, 2026, is the frontier version of the model and is what OpenAI describes as its most capable for advanced reasoning, complex analysis, and extended multi-step tasks. The full GPT-5.5 scored above 60 on the AI Intelligence Index; GPT-5.5 Instant is competitive but optimised for the latency and cost profile that makes it viable as a default for all users.
In practice, Plus and Pro subscribers can switch between GPT-5.5 Instant and the full GPT-5.5 depending on task complexity. For the majority of everyday ChatGPT tasks — writing, summarising, answering questions, light coding assistance, and general research — GPT-5.5 Instant is adequate and significantly faster. For tasks requiring the deepest reasoning — complex mathematical proofs, extensive code debugging across large files, nuanced long-form analysis — the full GPT-5.5 delivers meaningfully better results. Free users get GPT-5.5 Instant with usage limits; Plus and Pro users get GPT-5.5 Instant as the default with the ability to switch to full GPT-5.5 for high-complexity tasks. The tiering is designed to maximise user experience at each subscription level while preserving the premium value of the higher-cost tiers.
“GPT-5.5 Instant can use search to pull from past conversations, files, and Gmail to personalize answers. Memory sources are now visible across all models.” — Build Fast with AI, AI news May 18, 2026
The Competitive Context — GPT-5.5 Instant vs Claude and Gemini
GPT-5.5 Instant’s deployment as the default ChatGPT model arrives in a competitive context where Anthropic’s Claude Opus 4.7 leads on coding (SWE-bench Pro at 64.3%) and creative writing quality, while Gemini 2.5 Pro leads on scientific reasoning. The AIME 2025 score of 81.2 for GPT-5.5 Instant positions it competitively with Gemini 2.5 Pro on mathematical reasoning — one of the domains where Google has claimed the clearest advantage. For everyday ChatGPT users, the benchmark comparison is less important than the practical experience improvement, which rests primarily on the memory feature rather than the raw capability numbers.
The memory feature is where GPT-5.5 Instant’s competitive positioning is most distinctive. Claude has introduced memory features that allow Claude to retain context and preferences across conversations, but the implementation differs. GPT-5.5 Instant’s ability to search across the full history of user interactions, connected files, and Gmail provides a broader data integration than Claude’s current memory implementation. Gemini’s memory architecture is integrated with Google account data through Google Workspace, giving it similar breadth for Gmail and Drive but concentrated within Google’s ecosystem. GPT-5.5 Instant’s approach — connecting to Gmail regardless of whether the user is a Google Workspace customer — is more universally applicable for ChatGPT’s 900 million user base, which spans all email providers and workflow ecosystems.
| Feature | GPT-5.5 Instant (ChatGPT) | Claude (Anthropic) | Gemini 2.5 (Google) |
| Default model tier | Default for Free, Plus, Pro | Claude Sonnet 4.6 (free default) | Gemini 2.5 Flash (free default) |
| Memory across sessions | Yes — search across history, files, Gmail | Yes — retained preferences, limited history search | Yes — Google account integration |
| Memory transparency | Visible and editable by user | Manageable via settings | Google Activity controls |
| Gmail integration | Yes — all Gmail accounts | Via Google Workspace connectors | Native Google ecosystem |
| AIME 2025 math score | 81.2 (Instant) | Not separately benchmarked at Instant level | Comparable — Gemini leads scientific benchmarks |
| Multimodal (MMMU-Pro) | 76.0 | 3.75MP vision — strong | Natively multimodal — strong |
| Weekly active users | 900 million | Hundreds of millions (enterprise dominant) | Billions (via Search and Android) |
The Privacy and Governance Questions Around Persistent Memory
GPT-5.5 Instant’s expanded memory capabilities — particularly the Gmail integration and cross-session history search — raise governance questions that OpenAI has addressed in part through the memory transparency features but has not fully resolved. The ability to see, edit, and delete memories is a meaningful user control. It addresses the most obvious concern: that ChatGPT might retain sensitive information indefinitely without the user’s awareness. But the deeper question is not about the information ChatGPT retains — it is about the information it accesses. When GPT-5.5 Instant searches across a user’s Gmail history to personalise a response, it is accessing the contents of private email communications and synthesising them into model outputs. The scope of that access, the data handling practices that govern it, and the circumstances under which that accessed data could be used for model training or other purposes are questions that OpenAI has addressed in its privacy documentation but that will require continued scrutiny as the memory feature rolls out to the full user base.
In our hands-on review of ChatGPT’s memory settings as of May 2026, the memory management interface is more transparent than previous implementations. Users can click on the memory indicator, see exactly what the system has retained, delete specific entries, and disable memory entirely. The Gmail integration requires explicit connection authorisation and can be disconnected. These controls are genuine and functional. What they do not change is the fundamental architecture: a model with access to years of email history and months of conversation history is a qualitatively different tool than a model that responds to each session independently. The user experience improvement is real; the governance maturity of institutions and individuals in managing what AI systems know about them is the factor that will determine whether that improvement is net positive.
Key Takeaways
• OpenAI made GPT-5.5 Instant the default ChatGPT model on May 5, 2026, replacing the previous default across Free, Plus, and Pro tiers — the 900 million weekly active users of ChatGPT encountered the new model without a public launch announcement.
• GPT-5.5 Instant scored 81.2 on AIME 2025 mathematics (+15.8 vs predecessor) and 76.0 on MMMU-Pro multimodal reasoning (+6.8 vs predecessor) — meaningful improvements on benchmarks that test the reasoning quality most relevant to professional and academic use cases.
• The most significant new capability is persistent memory that searches across past conversations, uploaded files, and connected Gmail accounts to personalise every response. Memory sources are visible and editable by users — live for Plus and Pro on web, rolling out to mobile and free tiers.
• GPT-5.5 Instant is the efficient version of the GPT-5.5 family; the full GPT-5.5 (released April 23) is the frontier capability version. Plus and Pro subscribers can switch between them depending on task complexity — everyday tasks use Instant, complex analysis uses the full model.
• The Gmail memory integration extends personalisation beyond Google’s ecosystem, allowing ChatGPT to access and synthesise email history regardless of whether a user is a Google Workspace customer — a broader applicability than Gemini’s native Google account integration.
• Memory transparency controls — visible memories, individual deletion, full disable option, revocable Gmail authorisation — represent meaningful user governance, but the governance maturity of users and institutions in managing AI access to multi-year email and conversation histories is the outstanding question as the feature scales.
Conclusion
Making GPT-5.5 Instant the default ChatGPT model is one of the most consequential product decisions OpenAI has made in 2026 — not because of the benchmark numbers, but because of the memory architecture. The shift from a model that responds to each session independently to a model that searches across years of email history and months of conversations to personalise its output is a fundamental change in what AI assistance means. For the majority of 900 million ChatGPT users, the difference will be experienced as a chatbot that is suddenly much more useful in ways that are hard to attribute to any single feature — because the personalisation effect compounds across every response. The governance questions are real and will grow as the feature rolls out to the full user base. But the user experience improvement is also real, and in a competitive market where Claude and Gemini are offering comparable or superior capabilities in specific domains, a model that knows you better than its competitors is a durable advantage that is harder to benchmark than a AIME score or a coding evaluation. OpenAI is betting that personalisation — not raw capability — is where the consumer AI market will be won. The May 5 default switch is the opening move of that strategy.
Frequently Asked Questions
What is GPT-5.5 Instant?
GPT-5.5 Instant is the efficient, fast version of OpenAI’s GPT-5.5 model family, optimised for speed and cost at scale. It became the default ChatGPT model on May 5, 2026, replacing the previous default across Free, Plus, and Pro subscription tiers. It scored 81.2 on AIME 2025 math benchmarks and introduces persistent memory that searches across past conversations, files, and Gmail.
How is GPT-5.5 Instant different from the full GPT-5.5?
The full GPT-5.5 (released April 23, 2026) is the frontier capability version — optimised for maximum performance on complex, demanding tasks including advanced reasoning, extensive coding, and deep analysis. GPT-5.5 Instant is optimised for speed and cost at the scale of 900 million users, delivering significantly better-than-predecessor capability for everyday tasks at lower latency. Plus and Pro subscribers can switch to the full GPT-5.5 for high-complexity tasks.
What does the memory feature actually do?
GPT-5.5 Instant’s memory feature allows ChatGPT to search across your past conversation history, files you have uploaded to ChatGPT, and your connected Gmail account to personalise its responses. For example, it can reference a project you discussed three months ago, build on documents you previously uploaded, or draft emails in your communication style. Memory sources are visible and editable — you can see what ChatGPT remembers, delete specific memories, and disconnect Gmail access at any time.
Is ChatGPT’s memory privacy-safe?
OpenAI provides user controls including visibility into all retained memories, the ability to delete individual memories or all memories, and revocable Gmail connection authorisation. These are genuine controls. The broader governance question — the implications of a model with access to years of email history and months of conversations — is one that OpenAI’s privacy documentation addresses but that users and institutions should assess for their specific use cases, particularly in professional or regulated contexts where data sensitivity is high.
How does GPT-5.5 Instant compare to Claude and Gemini?
GPT-5.5 Instant is competitive on mathematical reasoning (81.2 AIME score) and multimodal capability (76.0 MMMU-Pro). Claude Opus 4.7 leads on coding benchmarks (SWE-bench Pro 64.3%) and creative writing quality. Gemini 2.5 Pro leads on scientific reasoning benchmarks. GPT-5.5 Instant’s key differentiator is the Gmail memory integration that is not limited to Google’s ecosystem — applicable across ChatGPT’s 900 million users regardless of their email provider.
References
Build Fast with AI. (2026, May 17). AI news today — May 18, 2026: 13 biggest stories. https://www.buildfastwithai.com/blogs/ai-news-today-may-18-2026
OpenAI. (2026, May 5). GPT-5.5 Instant default model update. OpenAI News. https://openai.com/news/
OpenAI. (2026, April 23). GPT-5.5 full model release. OpenAI News. https://openai.com/news/
OpenAI. (2026). ChatGPT memory settings and privacy documentation. https://help.openai.com/en/articles/chatgpt-memory
OpenAI. (2026). ChatGPT subscription tiers: Free, Plus, Pro, Enterprise. https://openai.com/chatgpt/pricing
Paramkusam, S. (2026, May 17). AI news today May 18, 2026: GPT-5.5 Instant default, memory features. Build Fast with AI.
TechInsight. (2026, May 2026). Latest AI updates May 2026: Global AI news and trends. https://techdg.in/latest-ai-updates-may-2026-global-ai-news-trends/
