Less than two months between GPT-5.4 and GPT-5.5. Even OpenAI president Greg Brockman admitted during Thursday’s press briefing that “there are probably enough model releases that it’s getting hard to distinguish one from another.”

He’s right. But GPT-5.5 deserves your attention anyway — not because of what it promises, but because of what it reveals about where this industry is heading at full speed.

The Pitch: Less Hand-Holding, More Doing

OpenAI’s headline claim is deceptively simple: GPT-5.5 can do more with less guidance. Hand it a messy, multi-part problem and it figures out the plan on its own — choosing tools, checking its work, iterating toward a solution.

“What is really special about this model is how much more it can do with less guidance,” Brockman said. “It can look at an unclear problem and figure out just what needs to happen next.”

This is the agentic AI vision every lab has been chasing. Not an AI that answers questions, but one that does things. And based on early reports, GPT-5.5 actually delivers — at least in specific domains.

Where It Flexes Hardest

Coding is the main event. GPT-5.5 powers an upgraded Codex running on NVIDIA’s GB200 NVL72 systems. One early tester handed it a “sloppily vibecoded codebase” and asked for a clean, professional rewrite. The model handled it. For developers drowning in technical debt, that’s not a demo — it’s a lifeline.

Computer use gets a real upgrade. Better at navigating software, creating documents, building spreadsheets — the digital busywork that devours knowledge workers’ days. This feeds directly into OpenAI’s “super app” vision: ChatGPT, Codex, and an AI browser unified into one enterprise platform.

Scientific research sees meaningful gains. Chief Research Officer Mark Chen highlighted breakthroughs in drug discovery and mathematics workflows. Not consumer-facing yet, but it signals where the long-term value actually lives.

The Benchmark Brawl

Here’s where it gets interesting. GPT-5.5 narrowly beats Anthropic’s Claude Mythos Preview on Terminal-Bench 2.0 — one of the most-watched coding benchmarks in the industry. That matters because Mythos has been the talk of Wall Street since its April announcement.

Broader benchmarks show GPT-5.5 consistently outscoring GPT-5.4, Google’s Gemini 3.1 Pro, and Claude Opus 4.5. But as Simon Willison noted, “the jagged frontier continues to hold” — excellent at some things, surprisingly stumbling on others in ways nobody can predict.

The head-to-head with Claude Opus 4.7 (launched just a week earlier) is essentially a coin flip. Both offer 1M-token context windows. Both lean on thinking-style reasoning. Both are positioned as best-in-class for agentic coding. The competition has never been tighter.

Double the Price. Is It Worth It?

API pricing lands at $5 per million input tokens and $30 per million output tokens — twice GPT-5.4’s per-token cost.

OpenAI’s counter: the model is significantly more “token efficient,” using fewer tokens to complete the same task. Brockman called it “a faster, sharper thinker for fewer tokens compared to 5.4.”

Does that math work? Depends entirely on your use case. Simple queries? You’re paying more. Complex, multi-step tasks where GPT-5.4 burned through tokens on false starts? The efficiency gains might actually save money. Real-world data over the coming weeks will settle the debate.

There’s also a new Codex Fast mode — 1.5x faster token generation at a 2.5x price premium. Speed costs, literally.

The Cybersecurity Problem Nobody Can Dodge

The most sobering detail: OpenAI classified GPT-5.5 as “High” cybersecurity risk — meaning it can “amplify existing pathways to severe harm.” It stops short of “Critical” (unprecedented new attack vectors), but that’s a thin comfort.

This isn’t theoretical. Anthropic already limited Mythos rollout over its ability to identify software vulnerabilities. Unauthorized access reports followed. The cybersecurity capabilities of frontier models have become the hottest — and most dangerous — conversation in AI.

OpenAI’s answer is a concept called “Trusted Access for Cyber”: stricter classifiers that gate vulnerability-identification capabilities for general users. The model can find and patch advanced security flaws, but that power sits behind additional safeguards.

“GPT-5.5 underwent extensive third-party safeguard testing and red teaming for cyber and bio risks,” said VP of Research Mia Glaese. Translation: these models are powerful enough that safety isn’t a checkbox anymore. It’s a core product feature.

What Pachocki Said That Should Worry Everyone

Chief Scientist Jakub Pachocki dropped this during the briefing:

“We see pretty significant improvements in the short term, extremely significant improvements in the medium term. In fact, I would say the last two years have been surprisingly slow.”

Surprisingly slow. Models in November, December, March, April — and the guy building them thinks this pace is leisurely. Let that sink in.

The Super App Gambit

Zoom out and the strategy is obvious. OpenAI wants to be the enterprise AI operating system — ChatGPT for interaction, Codex for development, an AI browser for everything else. GPT-5.5’s cross-domain capabilities are the foundation.

It’s a direct shot at Microsoft’s Copilot ecosystem and Google’s Workspace AI. And it lands the same week DeepSeek V4 drops at a fraction of the price, making the competitive landscape as chaotic as it’s ever been.

The Bottom Line

If you’re a ChatGPT Plus subscriber, GPT-5.5 is available now. The practical difference: better handling of complex, open-ended requests without spelling everything out. Ask it to “organize my project” instead of feeding it 15 sub-tasks.

If you’re a developer, the Codex improvements are worth testing immediately. Messy codebase refactoring alone could justify the upgrade.

If you’re a business leader, the ROI calculus just got harder. More capable, yes — but at double the per-token price. Your mileage depends on task complexity.

And if you’re watching from the sidelines? The gap between “AI assistant” and “AI coworker” is closing faster than anyone expected. GPT-5.5 isn’t there yet. But you can see it from here.

OpenAI’s bet is clear: move fast, charge more, and trust that capability wins. The market will decide if they’re right.