OpenAI Announces GPT-5.5 for Coding, Research and Data Analysis

By James Calloway

OpenAI has released GPT-5.5, a frontier model that scores a state-of-the-art 82.7% on Terminal-Bench 2.0 — a benchmark for complex command-line workflows — while matching the per-token latency of its predecessor, GPT-5.4. The company announced the rollout to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex, with GPT-5.5 Pro limited to Pro, Business, and Enterprise tiers. As of an April 24, 2026 update, both models are also available in the API.

The release stakes are considerable. OpenAI frames GPT-5.5 as "the next step toward a new way of getting work done on a computer," positioning it in the contested agentic AI market where coding ability, computer use, and autonomous task completion decide enterprise contracts.

What do the benchmarks actually show?

The headline numbers are strong on their face:

• Terminal-Bench 2.0: 82.7%, described as state-of-the-art accuracy on planning-heavy command-line work.
• SWE-Bench Pro: 58.6% on real-world GitHub issue resolution, solving more tasks end-to-end in a single pass than previous models.
• GDPval: 84.9% across knowledge work in 44 occupations.
• OSWorld-Verified: 78.7% for autonomous operation of real computer environments.
• Tau2-bench Telecom: 98.0% on complex customer-service workflows — notably, without prompt tuning.
• Additional knowledge-work scores: 60.0% on FinanceAgent, 88.5% on internal investment-banking modeling, and 54.1% on OfficeQA Pro.

OpenAI says GPT-5.5 beats GPT-5.4 on all three coding evals while using fewer tokens. On Artificial Analysis's Coding Index, the company claims state-of-the-art intelligence at half the cost of competitive frontier coding models.

A caveat the company itself discloses: some labs have noted evidence of memorization on certain evals, and all GPT results were run at "xhigh" reasoning effort in a research environment that "may provide slightly different output from production ChatGPT." Benchmark gains do not always survive contact with production traffic.

The takeaway in plain words: the scores are genuinely high, but they come from OpenAI's own testing configuration, and at least one eval carries a memorization flag.

How fast is it, and what does it cost?

Larger models usually serve slower. OpenAI says GPT-5.5 breaks that pattern, matching GPT-5.4 per-token latency in real-world serving. It also completes the same Codex tasks with significantly fewer tokens.

API pricing is where the trade-off shows:

• gpt-5.5: $5 per 1M input tokens, $30 per 1M output tokens, with a 1M context window. Batch and Flex at half rate; Priority at 2.5x.
• gpt-5.5-pro: $30 per 1M input tokens, $180 per 1M output tokens.
• Codex Fast mode: tokens generated 1.5x faster for 2.5x the cost.
• Codex access includes a 400K context window across Plus, Pro, Business, Enterprise, Edu, and Go plans.

OpenAI concedes the model is priced higher than GPT-5.4 but argues token efficiency compensates. In Codex, the company says it tuned the experience so most users get better results with fewer tokens.

The takeaway: cheaper per task, perhaps, but more expensive per token — efficiency claims rest on OpenAI's own workload measurements.

What did early testers say?

The tester quotes are enthusiastic, and should be read as such — they come from partners OpenAI selected for early access.

Dan Shipper, Founder and CEO of Every, called GPT-5.5 "the first coding model I've used that has serious conceptual clarity." His test is at least concrete: after an engineer rewrote part of a post-launch system, Shipper asked GPT-5.5 to reproduce that rewrite from the broken state. GPT-5.4 could not. GPT-5.5 could.

Pietro Schirano, CEO of MagicPath, said "It genuinely feels like I'm working with a higher intelligence, and there's almost a sense of respect," after the model merged a branch with hundreds of frontend changes into a substantially changed main branch in roughly 20 minutes.

An NVIDIA engineer with early access went further: "Losing access to GPT-5.5 feels like I've had a limb amputated."

Senior engineers reported the model outperforming GPT-5.4 and Claude Opus 4.7 on reasoning and autonomy, including one case where a request to re-architect a comment system in a collaborative markdown editor returned a nearly complete 12-diff stack.

Can it do real science?

OpenAI pushes a research-co-scientist narrative, and supplies two external examples.

Derya Unutmaz, an immunology professor at the Jackson Laboratory for Genomic Medicine, used GPT-5.5 Pro to analyze a gene-expression dataset with 62 samples and nearly 28,000 genes, producing a report he said would have taken his team months. Bartosz Naskręcki, assistant professor of mathematics at Adam Mickiewicz University in Poznań, built an algebraic-geometry app from a single prompt in 11 minutes.

The most striking claim: an internal GPT-5.5 version with a custom harness discovered a new proof about off-diagonal Ramsey numbers — central objects in combinatorics, the study of how discrete structures fit together — later verified in Lean, a formal proof checker. Results in this area are rare, and OpenAI calls it a concrete example of the model producing a useful mathematical argument rather than code or explanation. The phrase "internal version with a custom harness" deserves scrutiny; this is not the shipping product.

On GeneBench, a new genetics and quantitative-biology eval requiring reasoning over ambiguous data and hidden confounders, GPT-5.5 shows clear improvement over GPT-5.4. On BixBench, a bioinformatics benchmark, it achieved leading performance among models with published scores. One unnamed drug-discovery partner reported "significant accuracy gains on our hardest drug discovery evals."

The takeaway: the scientific results are real but partially self-reported, run on internal or custom configurations, and await independent replication.

What about safety and cybersecurity?

This is the most consequential part of the announcement, and the least quantified. OpenAI rates GPT-5.5's biological/chemical and cybersecurity capabilities as High under its Preparedness Framework — below Critical, but a step up from GPT-5.4.

Before release, OpenAI ran its full safety suite, worked with internal and external red-teamers, added targeted testing for advanced cybersecurity and biology, and gathered feedback from nearly 200 trusted early-access partners.

The cyber policy has three parts:

• Stricter safeguards: tighter controls on higher-risk activity and sensitive cyber requests, plus protections against repeated misuse. OpenAI admits the stricter classifiers "some users may find annoying initially, as we tune them over time."
• Trusted Access for Cyber: verified users can apply at chatgpt.com/cyber for fewer restrictions on defensive security work; organizations defending critical infrastructure can apply for cyber-permissive models like GPT-5.4-Cyber under strict security requirements.
• Government partnerships: exploring how advanced AI supports officials protecting taxpayer data, the power grid, and water supplies.

OpenAI's stated position: powerful models will be broadly distributed regardless, so the best path is putting them in defenders' hands. Critics will note this framing conveniently aligns expanded access with commercial growth. The system card, updated April 24, 2026, details the safeguards.

How did OpenAI hit its performance targets?

The infrastructure story is unusual. GPT-5.5 was co-designed for, trained with, and served on NVIDIA GB200 and GB300 NVL72 systems — and OpenAI says the model itself improved the stack that serves it. Codex analyzed weeks of production traffic and wrote custom partitioning heuristics that increased token generation speeds by over 20%.

Internally, OpenAI reports more than 85% of its staff uses Codex weekly. The Finance team reviewed 24,771 K-1 tax forms totaling 71,637 pages, accelerating the task by two weeks versus the prior year. A Go-to-Market employee automated weekly reports, saving 5-10 hours a week.

What happens next?

OpenAI calls GPT-5.5 "an incremental but important step" toward AI that can solve problems like cybersecurity — a notably modest framing for a release this size. The bet is that agentic capability, token efficiency, and trusted access will compound across coding, knowledge work, and scientific research. The open questions are the ones the announcement does not answer: independent benchmark verification, how annoying the new cyber classifiers prove in practice, and whether the efficiency gains hold beyond OpenAI's own workloads.