OpenAI Releases GPT-5.5, Touting Gains in Coding, Knowledge Work and Science

OpenAI has begun rolling out GPT-5.5, the latest flagship in its GPT-5.x series, to paying ChatGPT and Codex users. The company says GPT‑5.5 understands what you're trying to do faster and can carry more of the work itself, emphasizing improvements in agentic coding, computer-based knowledge work and scientific research, while maintaining per-token latency on par with its predecessor GPT-5.4 [4]. API access is expected “very soon” at new, higher price tiers [4].

Benchmarks and claims

OpenAI’s announcement publishes a suite of vendor-run benchmarks showing GPT-5.5 ahead of GPT-5.4 across the board. On Terminal-Bench 2.0, which tests complex command-line workflows, GPT-5.5 scores 82.7% versus 75.1% for GPT-5.4. On SWE-Bench Pro (real-world GitHub issue resolution) it reaches 58.6%. The company’s internal Expert-SWE evaluation, covering long-horizon coding tasks with a median human time of 20 hours, shows 73.1% versus 68.5% [4].

For general knowledge work, GPT-5.5 scores 84.9% on GDPval (44 occupations), 78.7% on OSWorld-Verified (operating real computer environments) and 98.0% on Tau2-bench Telecom (customer-service workflows) without prompt tuning [4]. In scientific research, OpenAI highlights leading performance on GeneBench (multi-stage genetics data analysis) and BixBench (bioinformatics), and notes an internal version helped discover a new proof about Ramsey numbers in combinatorics, later verified in Lean [4].

The company also cites Artificial Analysis’s Coding Index, a weighted average of 10 evaluation tests run by an external party, where GPT-5.5 delivers state-of-the-art intelligence at half the cost of competitive frontier coding models [4]. OpenAI says GPT-5.5 uses “significantly fewer tokens” than GPT-5.4 to complete the same Codex tasks, making it more efficient as well as more capable [4].

Agentic coding and computer use

The announcement centers on “agentic coding”: the model’s ability to plan, iterate, use tools, check its work and carry changes across a codebase. Dan Shipper, founder and CEO of Every, described GPT-5.5 in Hebrew as the first coding model he had used that demonstrates truly clear conceptual understanding [4]. Pietro Schirano, CEO of MagicPath, reported that GPT-5.5 merged a branch with hundreds of client-side and code-structure changes into a main branch that had also shifted significantly, completing the job in a single pass in roughly 20 minutes [4]. Justin Boitano, Senior Vice President of Enterprise AI at NVIDIA, said in Hebrew that GPT-5.5 is notably smarter and more persistent than GPT-5.4, with stronger coding performance and more reliable tool use, and that it continues working on tasks for significantly longer without stopping too early — which is especially important for the complex and lengthy work users pass to Cursor [4]. OpenAI reports that more than 85% of its own staff now use Codex weekly across functions from engineering to finance [4].

GPT-5.5 also drives Codex’s computer-use skills — seeing screen contents, clicking, typing and navigating interfaces. The company cites internal workflows: reviewing 24,771 K-1 tax forms (71,637 pages) two weeks faster than the prior year, and automating weekly business reports to save 5–10 hours per week [4].

Scientific research

OpenAI positions GPT-5.5 as a “co-scientist.” Beyond benchmark scores, it cites Daria Onutmaz, professor of immunology at the Jackson Laboratory, who said a gene-expression analysis covering 62 samples and nearly 28,000 genes that would have taken months was completed in a single session [4]. Bartosz Naskręcki, senior lecturer in mathematics at Adam Mickiewicz University in Poznań, used GPT-5.5 in Codex to build an algebraic-geometry visualization app from a single prompt in 11 minutes [4]. Brandon White, co-founder and CEO of Axiom Bio, said in Hebrew that using GPT-5.5 in their harness to analyze vast biochemical databases and predict drug effects on humans significantly improved accuracy on their hardest drug-discovery evaluations [4].

Cybersecurity: High capability, not Critical

Under OpenAI’s Preparedness Framework, GPT-5.5 is rated High (not Critical) in both cybersecurity and biological/chemical domains [7]. The company says it does not meet the Critical threshold — defined as the ability to “identify and develop functional zero-day exploits of all severity levels in many hardened real-world critical systems without human intervention” [7].

Vendor and third-party evaluations show measurable gains. On Capture-the-Flag professional challenges, CVE-Bench (web vulnerability exploitation) and Cyber Range (end-to-end network operations), GPT-5.5 outperforms GPT-5.4. In Cyber Range, it passes scenarios including EDR Evasion, Firewall Evasion and Leaked Token that GPT-5.4 failed [7]. The internal VulnLMP evaluation found GPT-5.5 could sustain multi-day vulnerability research campaigns and produce proof-of-concept inputs, but did not independently produce a functional full-chain exploit against hardened targets; the bottleneck was exploit-development judgment, not breadth of search [7].

External evaluators Irregular, CAISI and UK AISI confirmed stronger offensive cyberperformance than prior models. UK AISI judged GPT-5.5 the strongest model overall on narrow cyber tasks (pass@5 90.5% ± 12.9% vs. GPT-5.4’s 71.4% ± 19.8%) and observed one successful run on a 32-step corporate-network attack simulation estimated at 20 expert hours [7]. UK AISI cautioned that the range omits real-world defensive tooling and the model failed a separate critical-infrastructure simulation [7].

OpenAI is expanding its Trusted Access for Cyber program, offering verified defenders fewer restrictions on GPT-5.5 for defensive work, while deploying stricter classifiers for potential cyber-risk requests that the company says may be annoying initially as they are tuned over time [4].

Safety and system-card findings

The GPT-5.5 system card notes that chain-of-thought controllability is lower than for GPT-5.4 Thinking and GPT-5.2 Thinking: at 50k-character CoTs, GPT-5.5 successfully controls only 0.2% versus 0.3% and 0.5% respectively, which OpenAI says increases monitorability confidence [7]. First-person fairness evaluation (harm_overall 0.0112) is on par with GPT-5.1 and within confidence intervals of GPT-5.2 and 5.4 [7].

Biological evaluations show only modest improvement over GPT-5.4-thinking (1.35% on biochemistry knowledge), well below OpenAI’s 30% investigation threshold [7]. Protein-binding and DNA-design tasks remain below expert baselines [7]. SecureBio’s pre-deployment assessment found strong high-level reasoning but noted uncertainty about jailbreak robustness [7].

Availability and pricing

GPT-5.5 is rolling out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex. GPT-5.5 Pro is rolling out to Pro, Business, and Enterprise users in ChatGPT. In ChatGPT, GPT-5.5 Thinking is available to Plus, Pro, Business, and Enterprise users; the announcement does not list GPT-5.5 Thinking as a Codex offering. Codex users get GPT-5.5 (base model) with a 400K context window and a Fast mode (1.5× speed for 2.5× cost) [4].

API pricing is set at $5 per 1M input tokens and $30 per 1M output tokens for gpt-5.5 (1M context window), with Batch/Flex at half rate and Priority at 2.5×. GPT-5.5-pro is priced at $30/$180 per 1M input/output [4]. OpenAI argues the higher prices are offset by greater token efficiency.

Gaps and unknowns

Independent benchmark replication and large-scale real-world reliability data are not yet available. The API launch timeline remains "very soon." Third-party release histories show GPT-5.5 as the current flagship as of June 2026, following GPT-5.1 through 5.4 releases through early 2026 [3].