OpenAI launches GPT-5.5 with desktop automation, premium tier

OpenAI announced GPT-5.5 on April 23, 2026. The company says the model can perform actions inside desktop applications-clicking menus, typing text and operating terminals-to complete multi-step professional workflows. GPT-5.5 supports a roughly 1.1 million-token context window and is available in a basic tier and a higher-priced Pro tier that costs $100 per month for individuals.

OpenAI made the model available inside ChatGPT with a mode the company calls “GPT-5.5 Thinking,” which OpenAI says delivers faster, more concise answers on difficult problems. Business customers face a per-output-token rate that OpenAI says is about double the cost of GPT-5.4, despite the model’s token-efficiency gains.

OpenAI provided internal examples and benchmark results to illustrate performance. Its financial team used GPT-5.5 to process 24,771 K-1 tax forms totaling 71,637 pages and completed the task two weeks faster than the prior year. In internal tests the model scored 88.5% on investment banking modeling tasks and 60% on the FinancialAgent v1.1 benchmark, four points higher than GPT-5.4. GPT-5.5 scored 84.9% on GDPval, 78.7% on OSWorld-Verified for autonomous operations on real machines, and 98% on the Tau2-bench Telecom dataset for complex customer-service workflows.

OpenAI reported efficiency improvements: the company says GPT-5.5 uses about 40% fewer tokens on Codex-style programming tasks and reaches correct answers with fewer conversational turns. The model is capable of reasoning across codebases larger than 500,000 tokens with high retrieval accuracy. An employee in OpenAI’s Go-to-Market group estimated that automating weekly business reports with GPT-5.5 could save roughly 5 to 10 hours of manual work per week.

OpenAI said GPT-5.5 assisted in building parts of the company’s own infrastructure. The model analyzed production traffic and generated custom load-balancing heuristics that OpenAI reports increased token generation speed by about 20%. In a developer test, GPT-5.5 returned a nearly complete 12-diff code stack to re-architect a markdown editor that required minimal human correction.

OpenAI described GPT-5.5 as optimized for self-correction and for reducing hallucinations through improved cross-referencing of large information libraries. The company noted lower latency and more structured outputs, and highlighted performance in law, data science, business and education. OpenAI acknowledged a trade-off between capability and cost, warning that large-scale agentic deployments can produce substantial bills.

Dan Shipper, founder and CEO of Every, described the model as “the first coding model that has ‘serious conceptual clarity.'” He said GPT-5.5 produced a code rewrite that matched what his engineering team approved after days of manual debugging.