OpenAI Just Released GPT-5.5 — Here Is What Actually Changed

GPT-5.5 Is Live — Here Is What You Need to Know

OpenAI released GPT-5.5 on April 23, 2026, rolling it out to Plus, Pro, Business, and Enterprise users in ChatGPT and Codex the same day. GPT-5.5 Pro — the higher-accuracy variant — is rolling out to Pro, Business, and Enterprise users simultaneously. API access is not live at launch; OpenAI is working through additional cybersecurity safeguards before releasing to API developers. Pricing for the API is confirmed: $5 per million input tokens and $30 per million output tokens for GPT-5.5, and $30 per million input and $180 per million output for GPT-5.5 Pro.

The release came six weeks after GPT-5.4 — the fastest turnaround between major model releases OpenAI has shipped. OpenAI President Greg Brockman called it "a new class of intelligence" and "a big step towards more agentic and intuitive computing." The internal codename is "Spud."

What GPT-5.5 Is Designed to Do

OpenAI is positioning GPT-5.5 as a model for work that unfolds across tools and over time — not a faster chatbot. The official launch copy focuses on four categories: agentic coding, computer use, knowledge work, and early scientific research. The underlying claim is that GPT-5.5 can receive a messy, multi-part task and handle the planning, tool use, error-checking, and follow-through without the user managing every intermediate step.

Greg Brockman described it as "way more intuitive to use — it can look at an unclear problem and figure out what needs to happen next." He used a concrete example: a math professor used GPT-5.5 and Codex together to build an algebraic geometry app from a single prompt in 11 minutes. Mark Chen, OpenAI's Chief Research Officer, highlighted gains in scientific and technical research workflows, including potential applications in drug discovery.

One benchmark stands out from a product standpoint: 78.7% on OSWorld-Verified — the test for whether AI agents can operate real computer environments the way a human does, navigating interfaces, using applications, and completing tasks in real software contexts. This is directly relevant to Codex's computer use feature, which launched April 16, 2026. OpenAI is no longer treating computer use as a demo — it is core to the product surface.

Benchmarks vs. Previous Models and Competitors

OpenAI published comparison data at launch showing GPT-5.5 consistently ahead of GPT-5.4, Gemini 3.1 Pro, and Claude Opus 4.5 across its benchmark suite. The specific numbers OpenAI highlighted:

Third-party benchmark scores across SWE-Bench, AIME, and GPQA Diamond were not published at launch time. OpenAI's own comparison table showed GPT-5.5 leading GPT-5.4 on the benchmarks it selected, but independent verification from Artificial Analysis and BenchLM is pending as of April 24, 2026. Check back — this table will update when third-party scores are confirmed.

Token Efficiency: The Key Claim

OpenAI's most commercially significant claim is token efficiency: GPT-5.5 delivers better results with fewer tokens than GPT-5.4 for most Codex users. The model is priced higher per token ($5/M input vs. GPT-5.4's $2.50/M input in API), but OpenAI argues the effective cost per completed task is lower because fewer tokens are needed to finish the same work. Brockman described it as "a faster, sharper thinker for fewer tokens." This claim has not yet been independently verified by developer benchmarking communities — treat it as directionally likely but unconfirmed until third-party evals ship.

Who Gets Access and When

API access is delayed specifically because of cybersecurity safeguards. OpenAI flagged cybersecurity as a Preparedness Framework category with GPT-5.2 and has continued building on those mitigations. GPT-5.5 ships with "stricter classifiers for potential cyber risk" — Brockman acknowledged some users may initially find these classifiers over-aggressive as OpenAI tunes them over time.

API Pricing vs. Competitors

The Batch and Flex tiers (half standard rate) are the practical API entry point for most teams running automated pipelines — $2.50/M input puts GPT-5.5 price-competitive with Claude Opus 4.6 on input tokens, though output is double. Priority processing at 2.5x standard rate is for latency-sensitive production workloads. The 1M token context window applies across all API tiers.

The Competitive Context

GPT-5.5 did not launch into a quiet market. Anthropic released Claude Opus 4.7 on April 16, 2026 — one week earlier — with its own emphasis on long-running software work and multi-step autonomy. OpenAI launched workspace agents in ChatGPT on April 22, then GPT-5.5 on April 23. The sequencing looks deliberate: stake the agentic computing territory with workspace agents, then immediately back it with a new frontier model.

Kimi K2.6 from Moonshot AI (released April 20, 2026) sits at $0.60/M input and leads SWE-Bench Pro at 58.6% — a benchmark OpenAI has not yet published a GPT-5.5 score for. The open-weight, low-cost competitor is exactly the pressure that makes OpenAI's token-efficiency argument important: if GPT-5.5 genuinely needs fewer tokens to complete the same tasks, the $5/M input price is less punishing than it looks on paper.

According to Fortune, OpenAI now has 4 million active Codex users, 9 million paying business users, 900 million weekly active users across ChatGPT, and over 50 million subscribers. Those numbers matter because GPT-5.5 rolls out to all of them on day one — no waiting list, no beta access. For a model release, that is an unusually clean launch.

What This Means for Codex Users Specifically

GPT-5.5 is the new default model powering Codex. OpenAI says it has "carefully tuned the experience so GPT-5.5 delivers better results with fewer tokens than GPT-5.4 for most users" in Codex specifically. Usage limits remain generous across subscription tiers — the implication is that the token efficiency gain offsets the higher per-token cost enough that most users stay within their plan limits at the same or better task output quality.

Workspace agents — launched April 22, one day before GPT-5.5 — are Codex-powered cloud agents that run long workflows, use connected apps, remember context across sessions, and continue working when the user is away. GPT-5.5 powers these agents from launch day. If you are on a paid ChatGPT plan, both features are available now without any additional setup.

What to Watch Next

Three things worth tracking in the coming weeks: first, independent benchmark verification from Artificial Analysis and BenchLM — OpenAI's own comparison data is not a neutral source. Second, the API launch date and whether any changes to the announced pricing come with it. Third, developer feedback on the cybersecurity classifiers — Brockman warned they may initially flag legitimate use cases, and how quickly OpenAI tunes them will determine the model's practical utility for security-adjacent development work.