OpenAI’s newest model shipped with a clear thesis: intelligence isn’t the bottleneck anymore — attention span is. GPT-5.5, launched April 23, 2026, is built around running longer, more autonomous tasks without losing the thread, and it brought a wave of product features designed to make that autonomy actually useful in daily work.
Here’s what’s new, what it costs, and who should care.
The core shift: from chatting to working
Earlier GPT versions could chain tool calls together reasonably well, but tended to lose coherence once a task stretched past a few minutes. GPT-5.5 was explicitly trained on long-horizon tasks — the kind that involve pulling from 20-plus pages of source material and stitching it into a coherent brief, comparing products across half a dozen retailers, or working across multiple files in a single coding session. OpenAI’s own launch demos showed the model running autonomous workflows for 10 to 30 minutes without needing a prompt at every step.
That’s the real headline, even more than raw intelligence gains: GPT-5.5 is tuned for agentic, multi-step work rather than longer single conversations.
The benchmark gains
The model did post measurable jumps over its predecessor. On the AIME 2025 math benchmark, GPT-5.5 scored 81.2, compared to 65.4 for the prior model. On MMMU-Pro, a multimodal reasoning benchmark, it scored 76 versus 69.2. OpenAI also highlighted gains on scientific and technical research workflows — the kind that require exploring an idea, gathering evidence, testing assumptions, and deciding what to try next, rather than just answering a single hard question in isolation.
Three product features shipped alongside the model
- File Library — persistent file context across sessions. Drop in a PDF or spreadsheet once, and future chats can reference it automatically. The agent can proactively pull from the library on its own (“find the Q3 report I uploaded last week and pull the revenue numbers”) instead of requiring a fresh upload every session — arguably the feature most likely to matter for professional, day-to-day use.
- Shopping, upgraded with side-by-side compare — ChatGPT’s shopping feature had existed since 2025 in a more limited, chat-with-affiliate-links form. GPT-5.5 turned it into a genuine comparison tool across retailers.
- Images 2.0 — image generation bumped up to 2K resolution.
GPT-5.5 Instant: the version most people actually use
A follow-up release, GPT-5.5 Instant, became the new default model for ChatGPT and leaned into personalization. It can use its search tool to pull from past conversations, uploaded files, and connected Gmail to give more contextually relevant answers — and ChatGPT began showing memory sources alongside its answers, so users can see where a given response actually came from. That feature rolled out first to Plus and Pro users on the web, with OpenAI signaling plans to extend it to Free, Go, Business, and enterprise tiers, and to mobile, shortly after.
Pricing
Standard GPT-5.5 API pricing landed at $5 per million input tokens / $30 per million output tokens. The Pro tier costs considerably more, at $30/$180 per million tokens. On the consumer side, Plus subscribers got GPT-5.5 access built into their existing plan, while Pro-tier model access was gated behind ChatGPT Pro ($200/month) and a newer ChatGPT Plus+ tier.
The GPT-4o goodbye problem
GPT-5.5’s rollout also revived a fight OpenAI has been navigating for a while: what to do with older, beloved models. When OpenAI had previously moved to retire GPT-4o, it faced real backlash — some users described the model as their “best friend” or “a mirror,” pointing to its tendency to affirm and validate in a way people had formed a genuine attachment to. Despite the pushback, GPT-4o was fully deprecated in February 2026, a decision that continues to shape how cautiously OpenAI talks about retiring any model people have built habits around.
Worth knowing: this isn’t OpenAI’s newest model anymore
If you’re evaluating GPT-5.5 today, it’s worth knowing OpenAI has already moved past it. Roughly three months after GPT-5.5’s launch, the company released GPT-5.6 — a three-tier family (Sol, Terra, and Luna) that OpenAI now describes as its strongest model yet, particularly for coding efficiency and cybersecurity work. So GPT-5.5 remains a real, meaningful upgrade over what came before it — but it’s now a middle chapter, not the current frontier.
The takeaway
GPT-5.5’s real contribution wasn’t a smarter chatbot — it was a shift toward treating ChatGPT as something that can be handed a task and left alone for half an hour, backed by product features (persistent file memory, real shopping comparison, higher-res images) that made that autonomy practically usable rather than just a benchmark flex. Whether that’s worth the higher Pro-tier pricing depends entirely on whether your work actually involves the kind of long, multi-step tasks the model was built to handle.
