OpenAI Debuts GPT-5.5; API Token Prices Double

OpenAI released GPT-5.5 on April 23 and opened API access on April 24. The company describes the model as agentic: capable of planning sequences of actions, calling external tools and checking its own output to handle multi-step tasks with less human intervention.

GPT-5.5 is the first retrained base model since GPT-4.5 and was co-designed with NVIDIA’s GB200 and GB300 NVL72 rack-scale systems. OpenAI has rolled the model into ChatGPT and Codex for Plus, Pro, Business and Enterprise customers, with API access available after the product launch.

API pricing doubled versus GPT-5.4. Standard rates are $5 per million input tokens and $30 per million output tokens. OpenAI says GPT-5.5 completes many Codex tasks with fewer tokens than GPT-5.4, and estimates effective costs rise by roughly 20% after accounting for that efficiency. An independent test lab, Artificial Analysis, reported results consistent with OpenAI’s efficiency claim. A higher-tier GPT-5.5 Pro option is priced at $30 per million input tokens and $180 per million output tokens and applies extra parallel compute for harder problems.

OpenAI published benchmark results alongside the release. On Terminal-Bench 2.0, which tests command-line workflows that require planning and tool coordination, GPT-5.5 scored 82.7 percent, versus 75.1 percent for GPT-5.4 and 69.4 percent for a competing model. On SWE-Bench Pro, which evaluates solving GitHub issues, GPT-5.5 reached 58.6 percent and reportedly resolves more issues in a single pass than prior versions. An internal Expert-SWE benchmark, with tasks estimated to take a median of 20 human hours, produced a score of 73.1 percent for GPT-5.5 compared with 68.5 percent for GPT-5.4. In a one-million-token retrieval test (MRCR v2), GPT-5.5 scored 74.0 percent versus 36.6 percent for GPT-5.4.

OpenAI did not report a GPT-5.5 result on MCP Atlas, a tool-use benchmark, where another model recorded a 79.1 percent score. OpenAI reported that GPT-5.5 Pro leads on BrowseComp, the company’s web-browsing benchmark for agentic tasks, with a 90.1 percent score.

The company said GPT-5.5 matches GPT-5.4’s per-token latency in production serving despite higher capability. OpenAI reported internal adoption gains: more than 85 percent of employees now use Codex weekly across departments including engineering and marketing. One example cited a communications team that used GPT-5.5 to process six months of speaking requests and produce a scoring and risk framework to automate low-risk approvals.

At an output volume of 10 million tokens per month, GPT-5.5 standard would cost $300, while a competing model would cost $250 under listed rates, leaving roughly a 20 percent premium. OpenAI and independent testers recommend stress-testing token efficiency on real workloads before migrating. Observers will watch MCP Atlas results for users building heavy tool-use orchestration and follow whether gains on Terminal-Bench translate into production improvements for unattended terminal agents and DevOps automation.

Greg Brockman called the update “a real step forward.” Chief scientist Jakub Pachocki described the last two years of model progress as “surprisingly slow.”