OpenAI Releases GPT-5.5 and a Unified Desktop Agent

On April 23, 2026, OpenAI released GPT-5.5, a new frontier model optimized for autonomous multi-step tasks. The rollout includes a unified desktop application that merges ChatGPT, Codex, and Atlas into a single workspace. This architectural shift enables the model to maintain context across web browsing, writing code, and automating general software operations.

Model Variants and Pricing

OpenAI introduced two primary model tiers alongside a specialized processing mode called GPT-5.5 Thinking. The standard GPT-5.5 model is available immediately to ChatGPT Plus, Pro, Business, and Enterprise users. The higher-performance GPT-5.5 Pro utilizes parallel test-time compute for complex data science and legal research, restricted to the Pro ($200 per month), Business, and Enterprise tiers.

API access is scheduled for a future update. The general API supports a 1 million token context window.

Model	Input Price (per 1M tokens)	Output Price (per 1M tokens)
GPT-5.5	$5.00	$30.00
GPT-5.5 Pro	$30.00	$180.00

Unified Desktop Application

The launch of the desktop client centralizes OpenAI’s previous standalone tools. By combining ChatGPT with the Codex coding agent and the Atlas browser agent, developers can execute cross-application workflows without manually moving data between windows. Within the Codex environment, the model operates with a restricted 400,000 token context window.

Benchmark Results

OpenAI reports that GPT-5.5 achieved state-of-the-art results on 14 technical benchmarks. The model consistently outperformed Anthropic’s Claude Opus 4.7 and Google Gemini 3.1 Pro in agentic evaluations.

Benchmark	GPT-5.5 Score	Claude Opus 4.7 Score
Terminal-Bench 2.0	82.7%	69.4%
OSWorld-Verified	78.7%	78.0%
FrontierMath Tier 4	39.6% (Pro variant)	22.9%

The standard model also scored 73.1% on Expert-SWE, an internal coding benchmark measuring tasks that average 20 hours of human work. On GDPval, which tests economic tasks across 44 fields, the model set a new record at 84.9%.

Infrastructure and Training

Internally codenamed Spud, the model was trained and serves on NVIDIA GB200 and GB300 NVL72 infrastructure. The architecture is more token-efficient than GPT-5.4, delivering equivalent AI coding workflows at half the compute cost of competing models. OpenAI utilized GPT-5.5 during development to write new load-balancing heuristics for its own serving infrastructure, resulting in a 20 percent increase in token generation speeds.

Security testing evaluated biological and cyber threats. Under OpenAI’s Preparedness Framework, the cybersecurity capability rating is classified as High.

If you build systems that rely on cross-application context, the integration of Codex and Atlas within a single desktop environment changes how you structure multi-step automation. You should evaluate the GPT-5.5 API pricing structure against your current retrieval setups to determine if the 1 million token context window offsets the higher output token costs.