OpenAI shipped GPT-5.5 today. Six weeks after GPT-5.4, the model codenamed “Spud” is now live for Plus, Pro, Business, and Enterprise ChatGPT subscribers, and coming to the API shortly.

This is the release that OpenAI president Greg Brockman described as “a new class of intelligence” and “a big step toward more agentic and intuitive computing.” He wasn’t underselling it.

What Makes GPT-5.5 Different

The previous generation of OpenAI models were increasingly capable, but they were still largely designed to respond to prompts. GPT-5.5 is built to act on them.

The model handles text, images, audio, and video inside a single unified system rather than routing inputs to separate specialised models. That is not a minor technical detail. It means an agent built on GPT-5.5 can process a voice call, screenshot, document, and spreadsheet simultaneously and take action across all of them without stitching together outputs from multiple tools.

The other big shift is agentic reliability. GPT-5.5 can handle multi-step workflows with less user input, including writing and debugging code, researching online, analysing data, building documents, operating software, and moving across tools until a task is complete. This is the “chief of staff” capability that enterprise AI teams have been waiting for since agents first became a serious topic of conversation.

On benchmarks that actually matter for agentic work:

• Terminal-Bench 2.0 (complex command-line workflows): 82.7%, vs Claude Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5%
• Expert-SWE (software engineering tasks): 73.1%, up from GPT-5.4’s 68.5%
• OSWorld-Verified (real desktop environment automation): 78.7%, narrowly ahead of Opus 4.7’s 78.0%

OpenAI also claims it matches GPT-5.4’s per-token latency. More capable, same speed.

The Pricing Reality

GPT-5.5 API pricing is $5 per million input tokens and $30 per million output tokens, double GPT-5.4’s $2.50 and $15 respectively. The context window is 1 million tokens.

That is a meaningful cost increase for teams running high-volume production workloads. The counterpoint is that a more capable model often requires fewer iterations to complete a task, which can offset some of the per-token cost. For truly agentic workflows, where the model is driving a process to completion rather than answering a question, the economics will depend on what you are building.

The model is available now in ChatGPT and Codex for paid subscribers. API access will follow once OpenAI completes additional cybersecurity guardrails. The company worked with nearly 200 trusted early-access partners and conducted targeted testing for advanced cybersecurity and biology capabilities before today’s launch.

What This Means for Business

If you have been watching the AI model race from the sidelines waiting for the technology to mature, this is a useful moment to update your priors.

GPT-5.5 is not a better autocomplete tool. It is an AI system designed to complete work, the kind of work that currently requires a human to open applications, move information between them, make judgement calls along the way, and finish something. Whether that is processing customer enquiries, generating financial reports, or managing a multi-step workflow that touches multiple internal tools, the model is now capable enough to handle large portions of it autonomously.

That creates three practical implications for businesses right now.

The agent opportunity is real. Teams that have been experimenting with AI agents on simpler tasks now have a foundation capable enough to handle more complex ones. If your AI pilot programme has stalled because the model was not reliable enough on multi-step processes, GPT-5.5 is worth re-evaluating.

Your data infrastructure matters more, not less. A more capable model that can act across systems is only as useful as the data it can access. Businesses without clean, structured, accessible data will not get more out of GPT-5.5 than they got out of GPT-5.4. The model’s quality raises the ceiling. Your data quality is still the floor.

The competitive window is narrowing. OpenAI, Anthropic, and Google are all releasing flagship models within weeks of each other in 2026. Every quarter that passes with competitors deploying agents and you building internal consensus is a quarter of compounding advantage lost.

The model is live. The question is what you do with it.

---

If you’re building with Claude or Codex right now, grab the free Working With Claude field guide. Thirty-two pages on the full ecosystem, Claude Code in depth, and how to roll agents out properly. Get the free guide.