The most underreported fact about GPT-5.5 is not a benchmark number. It is that this is the first fully retrained base model built since GPT-4.5, and that distinction matters more than most headlines are giving it credit for. Released on April 23, 2026, GPT-5.5 is not just another incremental update or a fine-tuned variant of something older.

It is a complete ground-up rebuild designed with one central goal: agentic computing, or AI that can plan, execute, and complete complex multi-step work on your behalf without needing a hand to hold at every step. President and co-founder Greg Brockman described it as “a new class of intelligence” and “a big step towards more agentic and intuitive computing,” noting that the model is especially good at working with less guidance.

What Makes GPT-5.5 Different From Everything Before It

When I first started digging into GPT-5.5, I expected it to feel like most AI releases: impressive on paper but modest in practice. What changed my mind was the nature of the retraining itself. GPT-5.5 is the first fully retrained base model since GPT-4.5, designed to complete complex, multi-step computer tasks with minimal human direction. Previous models in the GPT-5.x family was a fine-tuned or improved version built on the same base. GPT-5.5 is different because it was rebuilt from the ground up, meaning its understanding of context, reasoning, and task execution comes from a completely fresh foundation.

The model is specifically engineered for computer use at a professional level. It can write and debug code, browse the web, fill out spreadsheets, operate software, and move across tools until a task is finished. More importantly, it can do all of this with minimal human direction. Instead of carefully managing every step, you can give GPT-5.5 a messy, multi-part task and trust it to plan, use tools, check its work, navigate through ambiguity, and keep going. The gains are concentrated in four specific areas: agentic coding, computer use, knowledge work, and early scientific research. These are areas where progress depends on reasoning across context and taking action over time.

GPT-5.5 Benchmark Numbers That Actually Matter

The benchmark story here is worth slowing down on, because the specific numbers tell a different story than the headlines suggest. On Terminal-Bench 2.0, which tests complex command-line workflows requiring planning and iterative tool use, GPT-5.5 scores 82.7%. For comparison, Anthropic’s Claude Opus 4.7 lands at 69.4% and Google’s Gemini 3.1 Pro sits at 68.5%. That is not a marginal lead. It is the kind of gap that shows up in real daily work, not just controlled testing environments.

On GDPval, which tests knowledge work performance across 44 real occupations, including finance, legal research, and product management, GPT-5.5 matches or beats industry professionals 84.9% of the time. On OSWorld-Verified, a benchmark measuring whether a model can autonomously operate real computer environments, it reaches 78.7%. What most articles missed is that GPT-5.5 achieves this performance while matching its predecessor’s per-token latency in real-world serving. That is genuinely unusual. Bigger, more capable models are almost always slower to run under the same hardware. GPT-5.5 manages to be meaningfully smarter without slowing down, which is something that rarely happens at this scale.

On SWE-Bench Pro, which grades real-world GitHub issue resolution, GPT-5.5 reaches 58.6%. Claude Opus 4.7 scores higher at 64.3%, though OpenAI has noted that Anthropic reported signs of memorization on a subset of those problems, which may affect the comparison. On the internal Expert-SWE benchmark for long-horizon coding tasks with a median human completion time of 20 hours, GPT-5.5 scores 73.1%, up from 68.5% for GPT-5.4. For tasks that a human engineer would need around 20 hours to complete, that is a meaningful step forward in autonomous end-to-end engineering capability.

The Superapp Connection Nobody Is Talking About

This is the part of the story that most people are sleeping on. Greg Brockman used the GPT-5.5 launch to explicitly tie this model to a broader ambition: the creation of a superapp, a single unified platform combining ChatGPT, Codex, and an AI browser into one tool capable of handling your entire workday. This is not a new idea, but connecting a model release to that vision so directly is a signal that GPT-5.5 is meaningfully close to making it feel real.

Personally, I think this superapp angle is the most important business development hiding behind what everyone else is covering as a coding benchmark story. If OpenAI succeeds in building a unified platform that can autonomously navigate your calendar, write your code, draft your documents, and handle your research without switching between apps, the competitive implications for Google, Microsoft, and Anthropic would be hard to overstate.

The pace of releases is also worth noting: the gap between GPT-5.4 and GPT-5.5 was just seven weeks, and chief scientist Jakub Pachocki said the team expects “pretty significant improvements in the short term, extremely significant improvements in the medium term.” Industry insiders hint that the superapp rollout timeline could accelerate meaningfully in the second half of 2026, though no official announcement has been made.

Scientific Research and the Drug Discovery Angle

After looking into this more closely, the scientific research story buried inside this launch is genuinely worth more attention than it is getting. Chief research officer Mark Chen confirmed that GPT-5.5 shows meaningful gains on scientific and technical research workflows, and specifically flagged its potential to help expert scientists make faster progress. He also pointed to drug discovery as an area where the model could have a real-world impact, a topic that has drawn rapidly growing industry interest in recent years.

I didn’t expect this angle when I started researching, and that is exactly why it matters. One of the more striking examples is that an internal variant of GPT-5.5, built with a customized research harness, helped produce a new mathematical proof relating to off-diagonal Ramsey numbers in combinatorics, which was subsequently verified in the Lean proof assistant.

Ramsey numbers are notoriously difficult to compute, with exact values known for only a handful of small cases. GPT-5.5 also shows clear improvement on GeneBench, a new evaluation for multi-stage scientific data analysis in genetics and quantitative biology, and on BixBench, built around real-world bioinformatics analysis. Sources suggest that OpenAI is in early conversations with pharmaceutical research groups about deploying GPT-5.5 in accelerated compound analysis workflows, though no partnerships have been officially confirmed.

Pricing and Availability: Who Gets GPT-5.5?

GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users in both ChatGPT and Codex. The model was tested with approximately 200 early-access partners before public release and ships with what OpenAI describes as its strongest safety measures to date. If you are on a free plan, this model is not coming your way, at least not yet. GPT-5.5 Pro, the higher-accuracy variant designed for harder tasks, is available to Pro, Business, and Enterprise users only.

For developers, API pricing is notably higher than GPT-5.4. GPT-5.5 will cost $5 per million input tokens and $30 per million output tokens, compared to $2.50 and $15, respectively, for GPT-5.4, which is a doubling of the per-token rate. GPT-5.5 Pro is priced even higher at $30 per million input tokens and $180 per million output tokens. OpenAI argues the token efficiency gains offset the increase, since GPT-5.5 completes the same tasks with fewer tokens overall, making real-world costs more comparable than the raw numbers suggest. Many believe that as competition with Anthropic and Google intensifies through the rest of 2026, OpenAI will face increasing pressure to bring these rates down before the year is out.

GPT-5.5 is the clearest signal yet that the AI industry has moved past the era of chatbot upgrades and into something more fundamental: models that can carry the actual work, not just assist with it. With a fully retrained foundation, a dominant lead on agentic benchmarks, and a direct line to the superapp vision, GPT-5.5 represents the most serious attempt yet at owning the way people work with AI, start to finish.