GPT-5.5: Built for Agents, Optimized for Completion
OpenAI's GPT-5.5 isn't just a better chatbot — it's built from the ground up to use tools, check its own work, and finish complex tasks. Here's what the specs mean for builders and buyers.
GPT-5.5 doesn't just benchmark better — it's built from the ground up for agentic work, signaling that the major labs have shifted their optimization target from impressive demos to completing real tasks reliably. Released on April 23, OpenAI's announcement frames this as a new model class defined by understanding complex goals, using tools, and carrying tasks through to completion.
The technical specs tell a focused story. As of April 2026, API pricing sits at $5/M input and $30/M output tokens, with context windows of 400K in Codex and 1M via API — numbers @swyx documented alongside a 20% Codex inference speed improvement. The model scores 82.7% on SWE-bench (as of April 2026), and it's the first OpenAI model co-designed with Nvidia's GB200 and GB300 NVL72 hardware. That hardware partnership matters: the software-hardware optimization loop, long standard in mobile chips, is now arriving in frontier AI infrastructure.
For developers building on the API, the context expansion is immediately useful — 1M tokens means loading substantial codebases, long conversation histories, or large document sets in a single call. The SWE-bench score puts agentic coding at a level where unsupervised pull-request generation becomes a reasonable starting point, not a curiosity. OpenAI also claims GPT-5.5 matches GPT-5.4's per-token latency while outperforming it across evaluations, which means you're not trading speed for capability.
The contrast with the open-source track is worth noting. On the same day, @simonw was requesting a sort-by-size feature on HuggingFace specifically to find small DeepSeek quantized models — a reminder that a meaningful slice of the developer community is actively optimizing for local inference, not API spend. The $30/M output price is competitive for production workloads but remains a real constraint for teams running high-volume pipelines, and that tension is only growing.
For product teams not building AI infrastructure, the ChatGPT-native availability is the practical headline: GPT-5.5 and GPT-5.5 Pro roll out to Plus, Pro, Business, and Enterprise tiers, which means the capability jump lands in the tool your team already uses.
The bottom line: If you're evaluating which model handles multi-step agentic work reliably, test GPT-5.5 first — its combination of 1M-token context, 82.7% SWE-bench (as of April 2026), and hardware-matched inference speed sets the new baseline.