GPT-5.5 "Spud"

What it is

GPT-5.5, known inside OpenAI by the codename “Spud”, was released on 23 April 2026, with the API following a day later on 24 April. By the standards of this line it was a substantial step, and although newer models have since arrived, gpt-5.5 remains widely deployed and is worth understanding on its own terms. It shipped in three variants, Instant, Thinking and Pro, covering the usual split between a fast default, a slower model that reasons before answering, and a higher-effort tier for the hardest problems. In practice that ladder lets a developer trade latency for depth: Instant for quick, cheap turns, Thinking when a problem needs the model to work through intermediate steps before replying, and Pro when accuracy on a genuinely hard task is worth the extra time and cost.

The default in ChatGPT

The clearest sign of how central gpt-5.5 became is that in May 2026 GPT-5.5 Instant replaced GPT-5.3 Instant as the default model in ChatGPT. That is the position that decides what most people actually use, since the overwhelming majority of ChatGPT sessions run on whatever the default happens to be rather than on a model the user has deliberately chosen. Promoting gpt-5.5 to that slot put it in front of an enormous audience almost overnight.

Specifications and benchmarks

On paper the specifications were ambitious. OpenAI listed gpt-5.5 with a context window of 1,050,000 tokens, large enough to hold a substantial book or a sizeable code base in a single prompt, and set API pricing at $5 per million input tokens and $30 per million output tokens. The gap between the two, with output priced six times higher than input, is typical of the frontier and reflects the greater compute cost of generating text than of reading it; for anyone budgeting a deployment it means the length of the model’s answers, and not only the size of the prompt, drives the bill. Reported benchmark figures placed it at 82.7% on Terminal-Bench 2.0, a test of an agent working in a command-line environment, and at 51.7% on Tiers 1 to 3 of FrontierMath with 35.4% on the harder Tier 4, a research-level mathematics benchmark. As ever, these are the numbers reported at launch, and how a model performs on real work can differ from any single scoreboard.

The goblins and gremlins footnote

The most-quoted footnote to gpt-5.5 has nothing to do with benchmarks. OpenAI disclosed that the model had a recurring habit of mentioning goblins and gremlins, seemingly at random, and traced the tic to reward signals left over from a retired internal personality nicknamed “Nerdy”. The company filtered the offending data and added a developer-prompt fix in Codex to suppress the behaviour. It is a small episode, but a revealing one, because it shows how choices made during training, including personalities a lab experiments with and later drops, can leave odd residues in a shipped model that only surface once millions of people are using it.

Where it sits and what to watch

For readers trying to place gpt-5.5 in context, the useful point is that it represents the mature form of a rapid release cadence rather than a clean-sheet design. The three-variant structure, the very large context window and the agentic and mathematical benchmarks all reflect where the frontier had moved by early 2026: towards models expected to act as coding agents and to reason over long inputs, rather than merely to chat. That is also why gpt-5.5 stayed in service even as successors appeared, since a model that is good enough, well priced and already wired into countless products is not quickly displaced.

Where to look next is the line’s continued cadence. OpenAI has iterated quickly, and gpt-5.5 sits as a recent predecessor rather than the final word, still carrying a large share of live traffic while newer models take the headlines. For the wider picture of how these systems compare, and how OpenAI’s releases fit alongside rival families, see our large language models hub and the broader AI section.