GPT-5.5, codenamed 'Spud,' is OpenAI's latest frontier model, released April 23, 2026. It is a fully retrained (not fine-tuned) agentic model that excels at computer use, agentic coding, and scientific reasoning. It scored 82.7% on Terminal-Bench 2.0, 84.9% on GDPval, and is OpenAI's strongest agentic coding model to date.

When did GPT-5.5 launch?

GPT-5.5 launched on April 23, 2026 (Thursday). It rolled out simultaneously to ChatGPT Plus, Pro, Business, and Enterprise, to the Codex CLI, Codex IDE extension, Codex cloud environment, and to the API. OpenAI had finished pretraining GPT-5.5 on March 24, 2026.

What's the difference between GPT-5.5 and GPT-5.4?

GPT-5.5 is a fully retrained model built specifically for agentic work. Compared to GPT-5.4, it is dramatically better at computer use (browsing, clicking, filling forms), multi-step autonomous tasks, and agentic coding. OpenAI reports that senior engineers found GPT-5.5 'noticeably stronger than GPT-5.4 and Claude Opus 4.7 at reasoning and autonomy.' GPT-5.4 remains available as a fallback.

How much does GPT-5.5 cost?

GPT-5.5 launched at approximately $1.50 input / $12 output per million tokens — similar to GPT-5.4 and roughly 10x cheaper than Claude Opus 4.7. It is currently available to ChatGPT Plus ($20/month), Pro ($200/month), and Business/Enterprise subscribers. API access is rolled out to all paying developers.

Quick Answer

What is GPT-5.5? OpenAI's New Agentic Model Explained

Published: April 24, 2026

What is GPT-5.5?

GPT-5.5, codenamed “Spud,” is OpenAI’s latest frontier model, launched April 23, 2026. It is a fully retrained agentic model — not a fine-tune of GPT-5.4 — built specifically for autonomous computer use, agentic coding, and multi-step scientific reasoning.

Last verified: April 24, 2026

The headline numbers

Benchmark	GPT-5.5 score	What it measures
Terminal-Bench 2.0	82.7%	Autonomous agent terminal tasks
GDPval	84.9%	Economically valuable knowledge work
SWE-bench Pro	58.6%	Industry-realistic coding
SWE-bench Verified	78.2%	Real GitHub issue resolution
τ²-Bench Telecom	79.1%	Multi-turn tool use
SciCode	State of the art	Scientific programming

OpenAI calls GPT-5.5 its “strongest agentic coding model to date.” The Terminal-Bench 2.0 score of 82.7% narrowly beats the Anthropic Claude Mythos Preview (82.0%) and easily surpasses Claude Opus 4.7 (69.4%).

What “agentic” means here

OpenAI president Greg Brockman framed GPT-5.5 as “a new class of intelligence” and “a big step towards more agentic and intuitive computing.” In concrete terms, the model can:

Browse the web and click through pages natively
Test flows by interacting with live web apps
Capture screenshots and iterate on what it sees
Plan, execute, and refine work with minimal human intervention
Keep working through multi-step tasks without requiring a human to supervise every move
Catch issues in advance and predict testing or review needs without being prompted

The key capability upgrade is computer use: GPT-5.5 is now OpenAI’s first model where autonomous browser and desktop control is native rather than an extension.

Where to get it

GPT-5.5 is available everywhere OpenAI ships on day one:

Surface	Availability
ChatGPT Plus ($20/mo)	✅ April 23, 2026
ChatGPT Pro ($200/mo)	✅ April 23, 2026
ChatGPT Business/Enterprise	✅ April 23, 2026
OpenAI API	✅ April 23, 2026
Codex CLI	✅ `gpt-5.5` is default
Codex IDE extension	✅
Codex Cloud	✅

If gpt-5.5 isn’t available in your account yet, OpenAI’s docs instruct developers to fall back to gpt-5.4.

NVIDIA confirmed that GPT-5.5 runs on NVIDIA infrastructure and that NVIDIA itself uses it for internal agentic workflows — a zero-data-retention deployment with read-only production access via Codex Skills.

What’s the “Spud” codename?

Axios reported that OpenAI’s internal codename for GPT-5.5 is “Spud.” OpenAI has a history of quirky internal codenames (strawberry for o1, orion for what became 4.5), but the public product name is simply GPT-5.5 — not Spud, which is strictly an internal artifact.

How GPT-5.5 compares to April 2026 competitors

Model	Released	SWE-bench Verified	Terminal-Bench 2.0	Input $/1M
GPT-5.5	Apr 23	78.2%	82.7%	$1.50
Claude Opus 4.7	Apr 16	87.6%	69.4%	$15.00
Claude Sonnet 4.6	Feb 2026	74.1%	62.8%	$3.00
Gemini 3.1 Pro	Apr 2026	70.8%	58%	$1.25
Gemini 3.1 Ultra	Apr 2026	73.4%	66%	$2.50

GPT-5.5 sits in the sweet spot: second-best coder, best agent, cheapest frontier.

Who is GPT-5.5 for?

Agent builders: The default choice for new production agent work in April 2026.
Codex users: Already the default model in Codex CLI, IDE, and cloud.
Computer-use workloads: Native, no extra plugin layer.
Budget-conscious teams: Costs roughly 1/10 of Claude Opus 4.7.
Enterprise customers: Zero-data-retention deployments available via the API’s data controls.

Who should not use GPT-5.5?

Long-context workloads (>400K tokens). Opus 4.7 (1M) and Gemini 3.1 Pro (2M) are better.
Hard, real-world coding refactors. Opus 4.7 still wins SWE-bench Verified by 9.4 points.
Anyone locked into a specific coding agent. Cursor and Claude Code default to Anthropic models; switching requires explicit config.

The speed of the 2026 model cycle

OpenAI shipped GPT-5.5 exactly one week after Anthropic shipped Claude Opus 4.7 on April 16, 2026. Fortune called this cadence “AI model launches starting to look like software updates.” Expect another major frontier release within the next 2–4 weeks.

For now, GPT-5.5 is the state of the art on agentic benchmarks — and the cheapest way to build production AI agents in April 2026.

Last verified: April 24, 2026. Sources: OpenAI (openai.com/index/introducing-gpt-5-5), NVIDIA Blog, VentureBeat, Fortune, Axios, Interesting Engineering, gHacks, MarkTechPost, OpenAI Codex docs (developers.openai.com/codex/models).