OpenAI drops GPT-5.5 with major agentic coding leaps

Coding with an AI agent has usually meant hovering your finger over the cancel button, waiting for it to misread the codebase and hallucinate five new bugs. OpenAI wants to change that dynamic.

The company just launched GPT-5.5, pitching it as its smartest and most intuitive model yet for complex agentic workflows, research, and operating software. Following its initial rollout in ChatGPT and Codex, OpenAI made both GPT-5.5 and GPT-5.5 Pro available in the API on April 24.

More intelligence without the latency hit

Bigger frontier models usually come with painful trade-offs in speed and cost. OpenAI claims GPT-5.5 breaks that pattern by matching GPT-5.4's per-token serving latency while using significantly fewer tokens and fewer retries to get Codex tasks across the finish line.

According to data from Artificial Analysis's Coding Index, the model delivers state-of-the-art coding performance at roughly half the cost of competing frontier options. On Terminal-Bench 2.0, which tests multi-step command-line environments, it scored 82.7%. On SWE-Bench Pro, measuring real-world GitHub bug fixing, it hit 58.6%, completing more end-to-end tasks in a single pass.

Feedback from early testers was unusually punchy. Dan Shipper, CEO of Every, noted that GPT-5.5 was the first model he used with genuine conceptual clarity, successfully reproducing a clean architectural rewrite his senior human engineer had chosen after days of debugging. MagicPath CEO Pietro Schirano said it resolved a massive multi-hundred-diff branch merge in a single 20-minute run. One early access engineer at NVIDIA joked that losing access felt like having a limb amputated.

Beyond pure software engineering, OpenAI is pushing the model as a universal computer operator. It scored 84.9% on the 44-occupation GDPval benchmark, 78.7% on OSWorld-Verified for autonomous desktop usage, and 98.0% out of the box on Tau2-bench Telecom. Internal OpenAI teams are leaning hard on it: the finance group reportedly processed 24,771 K-1 tax forms spanning over 71,000 pages, cutting two weeks off their timeline.

Why it matters

For engineers and knowledge workers, GPT-5.5 looks like an immediate upgrade worth testing on production workloads. The real headline here isn't just raw benchmark flex; it's token efficiency. In multi-step agentic workflows, repeated retries and rambling contexts are what blow up cloud budgets. A model that understands the whole architectural graph and solves issues in fewer passes directly slashes operating expenses.

Across the industry, this release ramps up an intense war of attrition with Anthropic. With Anthropic having recently dropped Claude Opus 4.7 and teasing its Mythos model, OpenAI needed a fast answer before either company tests public markets later this year, as noted by The Verge. Testing showed GPT-5.5 pulling ahead of Claude Opus 4.7 in multi-step autonomy and self-directed test generation.

We think OpenAI's strategy of optimizing for fewer tokens and lower error loops is exactly the right battleground. However, letting AI take the wheel on local machines is still fraught with risk. An autonomous desktop score of 78.7% means roughly one in five actions still breaks down, and when models begin executing high-stakes enterprise tasks like tax filings or live system refactors, the gap between semi-autonomous speed and unattended disaster remains uncomfortably narrow.