OpenAI released GPT-5.5 on April 23, 2026 — codenamed “Spud” internally — rolling it out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex. API access followed on April 24. The model delivers meaningful capability improvements over GPT-5.4, particularly in agentic coding, computer use, and long-horizon reasoning, while matching GPT-5.4’s per-token latency in real-world serving. It ships with the most restrictive cybersecurity controls OpenAI has ever deployed on a publicly released model.
What’s New in GPT-5.5
OpenAI’s framing of GPT-5.5 is straightforward: it’s smarter than GPT-5.4, faster per task due to token efficiency, and designed for multi-step agentic work where the model plans, uses tools, navigates ambiguity, and continues toward a goal without constant human direction.
Key capability areas where gains are strongest:
- Agentic coding: GPT-5.5 handles complex, multi-file coding tasks with fewer corrective interventions than GPT-5.4 — delivering better results with fewer total tokens consumed in Codex
- Computer use: Stronger performance on tasks requiring the model to operate software interfaces, navigate applications, and complete workflows across multiple tools
- Knowledge work: Real but incremental gains over GPT-5.4 for research, analysis, and document creation tasks
- Scientific research: OpenAI specifically highlights “early scientific research” as a domain showing meaningful improvement — consistent with the company’s stated interest in deploying AI for drug discovery and materials science applications
Token Efficiency: The Underrated Improvement
One of the most practically significant GPT-5.5 improvements isn’t raw capability — it’s token efficiency. OpenAI states that GPT-5.5 completes the same Codex tasks using significantly fewer tokens than GPT-5.4, producing better results with less total compute. Despite being more capable and more expensive per token, the total cost of completing a given task can actually be lower with GPT-5.5 than with GPT-5.4 — because it requires fewer iterations.
This matters considerably for enterprise API users with high-volume agentic workflows. If GPT-5.5 completes a coding task in three tool calls that GPT-5.4 required five for, the higher per-token price is offset by the reduced total token count. OpenAI claims this efficiency advantage persists across most Codex use cases.
The Cybersecurity Controls
GPT-5.5 ships with what OpenAI describes as “our strongest set of safeguards to date” — specifically targeting cybersecurity misuse. The model deploys stricter classifiers that detect and refuse requests that could facilitate attacks, exploit development, or vulnerability research outside authorized contexts. OpenAI acknowledges these classifiers “some users may find annoying initially” as the company tunes them over time.
The decision to add stricter cybersecurity controls to a publicly released model reflects a direct response to the attention that Anthropic’s Claude Mythos Preview generated — and the finding by the UK’s AI Security Institute that GPT-5.5 matches Mythos-level performance on cybersecurity benchmarks. If a public model is approaching the cybersecurity capability ceiling that prompted Anthropic to restrict Mythos to 40 organizations, OpenAI’s answer is deploying aggressive filters rather than access controls.
Pricing and Access
- GPT-5.5 (standard): $5 per million input tokens, $30 per million output tokens via API
- GPT-5.5 Pro: $30 per million input tokens, $180 per million output tokens — available in ChatGPT Pro, Business, and Enterprise only
- Context window: 1.1 million tokens (922K input, 128K output)
- ChatGPT access: Plus, Pro, Business, Enterprise users
- Codex access: Plus, Pro, Business, Enterprise, Edu, and Go plans
- Free tier: Not available
NVIDIA’s Internal Deployment
One of the most notable early adoption signals: NVIDIA gave over 10,000 of its staff early access to GPT-5.5 through Codex before the public launch — not just engineers, but also employees in legal, finance, and operations. The breadth of internal deployment across non-technical functions is one of the clearest indicators yet that frontier AI models are being seriously used for general knowledge work, not just code generation.
How It Compares
Independent benchmarks tell a nuanced story. On the hardest cybersecurity evaluation challenges, GPT-5.5 scored 71.4% versus Claude Mythos’s 68.6% — statistically within margin of error. On Tom’s Guide’s head-to-head comparison across seven categories, GPT-5.5 lost to Claude Opus 4.7 in all seven. The model appears strongest in agentic, computer-use, and tool-calling contexts and somewhat weaker than Claude in pure reasoning and writing quality evaluations — consistent with the different training emphases of the two labs.
Conclusion
GPT-5.5 is OpenAI’s clearest answer to the competitive pressure from Anthropic’s Claude Opus series — a model that pushes agentic capability and token efficiency simultaneously, while shipping with the safety controls needed to deploy it responsibly at public scale. Browse our full directory to compare ChatGPT, Claude, and every AI model shaping the development landscape in 2026.
