GPT-5.5 Makes OpenAI's Case for Execution at Scale

OpenAI's GPT-5.5 launch is not being sold as a prettier chatbot. It is being sold as infrastructure for work.

OpenAI's Pitch: A Model That Keeps Going

The company's headline claim is blunt: GPT-5.5 is "a new class of intelligence for real work." In practice, that means OpenAI wants developers and enterprises to think about the model less as a conversation engine and more as an execution layer for coding, computer use, research, and document-heavy workflows. Greg Brockman put it even more clearly on OpenAI's press call, describing the release as a step toward "more agentic and intuitive computing," and "a faster, sharper thinker for fewer tokens compared to 5.4," according to TechCrunch.

That framing matters because GPT-5.5 arrives in a market where raw model quality is no longer enough. Frontier labs now have to answer a harder question: what is the model actually for? OpenAI's answer is that GPT-5.5 should be the model you hand a messy, multi-step task and trust to keep going.

On OpenAI's own numbers, the case is credible. GPT-5.5 posts 82.7% on Terminal-Bench 2.0, up from 75.1% for GPT-5.4. It reaches 35.4% on FrontierMath Tier 4, versus 27.1% for 5.4. And the most eye-catching result is in long context: Graphwalks BFS at 1 million tokens jumps to 45.4%, from 9.4% on the previous model. In Codex, OpenAI says GPT-5.5 runs with a 400K context window, while the API version will support 1 million tokens. For anyone building coding agents, research systems, or document-heavy workflows, those are not cosmetic gains.

They also line up with what OpenAI emphasized across the launch materials and briefing. Chief scientist Jakub Pachocki said the improvements are significant and that "the last two years have been surprisingly slow," again via . Chief research officer Mark Chen said the model shows "meaningful gains on scientific and technical research workflows" and could help expert scientists make progress. Michael Truell, CEO of Cursor, said in OpenAI's release that GPT-5.5 is "noticeably smarter and more persistent" with stronger coding performance and more reliable tool use.

The Gains Are Real, but So Are the Gaps

But this launch is not a clean sweep.

For one thing, GPT-5.5 does not win everywhere. OpenAI's own benchmark table shows it still trails Anthropic on SWE-Bench Pro and MCP Atlas, and trails Gemini 3.1 Pro on BrowseComp. Even the GDPval gain, which is supposed to reflect real occupational work, is modest: 84.9% versus 83.0% for GPT-5.4. That is an improvement, but not the sort of jump that automatically rewrites the market.

The second issue is reliability. Independent benchmarking from Artificial Analysis says GPT-5.5 retakes the top spot on its Intelligence Index with a score of 60. That is a real marker of frontier status. But the same group also shows a weaker factuality profile on AA-Omniscience, where GPT-5.5's hallucination rate is reported at 86%, versus 36% for Claude Opus 4.7. In other words, OpenAI may have built a stronger execution model, but not one that reliably knows when to shut up.

That tradeoff is not academic. The more a model is positioned as an actor inside coding, research, and computer-use loops, the more costly overconfidence becomes. A chatbot can be wrong and annoying. An agentic system can be wrong, persistent, and expensive.

The Real Test Is Economics and Reliability

Then there is the price. OpenAI's API pricing page lists GPT-5.5 at $5 per million input tokens and $30 per million output tokens, exactly double GPT-5.4's $2.50 and $15. OpenAI argues that efficiency softens the blow, saying GPT-5.5 uses about 40% fewer tokens than GPT-5.4 for the same Codex tasks. That may be true in tightly optimized coding workflows. It is less certain in broader enterprise usage, where messy prompts, retries, and orchestration overhead often eat the savings.

This is why GPT-5.5 looks less like a pure capability leap and more like a strategic repositioning of what OpenAI thinks the flagship model should be. The company is trying to move the conversation away from chatbot vibes and toward labor substitution in high-value workflows. Coding. Operations. Research. Technical analysis. Software that does not just answer, but advances the task state.

That makes sense. It is also where competitive pressure is highest. Anthropic has been pushing hard on coding and tool use. Google has been leaning into long-context research and workflow integrations. If OpenAI wants to own the premium end of the market, it needs a model that feels like a better coworker, not just a more polished assistant.

The good news for OpenAI is that GPT-5.5 appears to move in that direction. The bad news is that the launch still reads as a bundle of tradeoffs: stronger coding, better long context, higher cost, incomplete benchmark dominance, and lingering hallucination issues. Brockman said GPT-5.5 "feels like it's setting the foundation for how we're going to use computers." Maybe. But foundations matter because other things get built on top of them, and right now this one still looks uneven in places that matter.

For builders, the takeaway is straightforward. GPT-5.5 is probably real progress if your stack depends on persistence, tool use, and codebase-scale reasoning. It is not obviously a universal upgrade if your constraint is reliability per dollar. OpenAI has shipped a more execution-oriented flagship model. Now it has to prove that the economics and factuality are good enough for the jobs it wants that model to do.