Not because the version number ticked up. That happens every few months now, and most of us have stopped getting excited about incremental bumps. No, this model genuinely changes how you interact with AI on real engineering and research tasks. It plans.So picks up tools on its own.So checks its own work. And here's the thing that actually matters: it keeps going until the job is done, instead of handing you some half-baked answer and basically saying "good luck with the rest."
OpenAI's calling it "a new class of intelligence for real work." Marketing speak, obviously. But the benchmarks and early tester reactions suggest there's real meat behind the slogan. The model went live April 23, 2026, for Plus, Pro, Business, and Enterprise users across ChatGPT and Codex. API access is on its way.
So what's actually going on here?
The Quick Version
GPT-5.5 is OpenAI's new flagship, built specifically for agentic coding, research, and gnarly multi-step problems. It's available right now in ChatGPT and Codex if you're a paid subscriber.
The coding benchmarks are wild. 82.7% on Terminal-Bench 2.0. 58.6% on SWE-Bench Pro. And 73.1% on Expert-SWE, which tests tasks estimated to take a human engineer around 20 hours. Twenty hours!
It burns through fewer tokens than GPT-5.4 while keeping the same per-token latency. So yeah, faster and cheaper in practice.
On the science side, GeneBench jumped to 25.0% from 19.0%, and BixBench hit 80.5% for bioinformatics work.
API pricing lands at $5 per million input tokens and $30 per million output tokens. The Pro tier runs $30/$180. Safety testing was extensive, with OpenAI rating bio and cyber capabilities as "High" risk and running the model past roughly 200 early-access partners before flipping the switch.
What Is This Thing, Really?
GPT-5.5 follows GPT-5.4 in the lineup, but calling it a simple iteration would be selling it short. Way short. This model was designed from scratch for agentic work. It doesn't just sit there answering prompts.Plus maps out multi-step workflows, grabs tools when it needs them, double-checks whether its output actually works, and loops back when something breaks.
From [OpenAI's announcement], the model "understands what you're trying to do faster and can carry more of the work itself." In practice, that means you can throw a messy, barely-defined task at it and walk away. It'll figure out how to break things down without you babysitting every step.
The internal codename was "Spud." I'm not making that up. Axios confirmed it on launch day.
One nerdy detail that caught my eye: they built this thing using NVIDIA GB200 and GB300 NVL72 systems. That's an absurd amount of compute, and you can genuinely feel it in the output quality.
The Coding Benchmarks Tell a Story
This is where it gets real for anyone who writes code professionally.
| Benchmark | GPT-5.5 | What It Measures |
|---|---|---|
| Terminal-Bench 2.0 | 82.7% | Complex command-line workflows |
| SWE-Bench Pro | 58.6% | Real GitHub issue resolution |
| Expert-SWE | 73.1% | Long-horizon coding, ~20hr human tasks |
Claude Opus scored 69.4% on Terminal-Bench 2.0. That 13-point gap? Not a rounding error. That's a canyon.
Numbers only tell part of the story though. What early testers have been saying is honestly more compelling. Dan Shipper, CEO of Every, called GPT-5.5 "the first coding model I've used that has serious conceptual clarity." He threw a real debugging scenario at it. Something had eaten up days of his team's time. GPT-5.4 choked on it. GPT-5.5 spit out essentially the same fix his best engineer eventually wrote on their own.
Then there's the NVIDIA engineer with early access who said: "Losing access to GPT-5.5 feels like I've had a limb amputated."
Nobody says about a minor upgrade.
Where the model really flexes in day-to-day engineering work:
- Holding context across huge codebases without losing the thread. You know how older models just... Forget what they were doing three files ago? That.
- Reasoning through ambiguous failures where the stack trace is basically useless
- Propagating changes across related files after a refactor so nothing gets left out of sync
- Debugging and testing with dramatically less back-and-forth ping-pong
The Cursor team noted GPT-5.5 "stays on task for significantly longer without stopping early, which matters most for the complex, long-running work our users delegate." Honestly one observation tells me more than any benchmark table could.
It's Not Just About Code
GPT-5.5 pulled 84.9% on GDPval, which evaluates knowledge work across 44 different occupations. On OSWorld-Verified, measuring how well the model actually operates inside real computer environments, it hit 78.7%.
The science improvements are something else entirely.
GeneBench went from 19.0% to 25.0%. BixBench landed at 80.5% for bioinformatics. FrontierMath Tier 1 through 3 improved across the board.
Researchers have been using it to chew through massive gene datasets and surface insights that would normally take months of manual work. Not "a bit faster." A completely different category of capability.
The model also handles document generation, spreadsheet creation, and presentations better than its predecessor. Combine it with Codex's computer use features and GPT-5.5 can navigate interfaces, click through applications, coordinate across tools. It's starting to feel less like a chatbot and more like a coworker who just happens to live in your browser. If you're curious about how the broader [agentic coding tool landscape] is shifting, context makes all of this click even harder.
What's It Going to Cost You?
GPT-5.5 comes included with ChatGPT Plus at $20/month, Pro at $200/month, and Business and Enterprise plans. No extra charge. Just shows up in your model selector.
For the API crowd:
GPT-5.5. $5 / 1M input tokens | $30 / 1M output tokens
GPT-5.5 Pro: $30 / 1M input tokens | $180 / 1M output tokens
Context window: 1 million tokensPeople on Reddit are already complaining about frontier model pricing creeping up relentlessly. And yeah, that's fair. But here's the flip side: OpenAI says GPT-5.5 uses fewer tokens to finish identical tasks compared to 5.4. The Artificial Analysis Coding Index reportedly shows it delivering "state-of-the-art intelligence at half the cost of competitive frontier coding models."
Whether that math pencils out for you depends on what you're doing. Complex engineering tasks that save hours of human debugging? Five bucks per million input tokens is basically nothing. For casual stuff, the Plus subscription covers everything anyway.
Safety: They Actually Took This Seriously
OpenAI ran their full safety evaluation suite before shipping, and they went deeper than usual. Internal and external red-teaming. Targeted testing around advanced cybersecurity and biology capabilities. Feedback loops with around 200 early-access partners. Full evaluation against their [Preparedness Framework].
Bio and cyber capabilities got a "High" rating. Not the top tier of "Critical," but it signals the model could amplify existing risks if someone skilled enough got creative with it. OpenAI built a "Trusted Access for Cyber" program giving verified security professionals specialized access to work with these capabilities responsibly.
Here's something I genuinely care about: GPT-5.5 is way better at not destroying your stuff when operating as an agent. It can tell the difference between its own work and changes you made, and it protects your files by default. Anyone who's had a model cheerfully nuke something they spent an hour crafting knows exactly why that matters. The [system card] digs into the evaluation methodology if you want the full picture.
At the model level, OpenAI estimates roughly 0.056% of conversation turns might trip policy-violation flags for harassment, and that's before other safety layers even engage. Low number. They acknowledge the measurement isn't perfect, which is refreshingly honest.
GPT-5.5 vs. The Competition
Claude Opus 4.7 dropped the same week because apparently nobody in AI can launch anything without company. On Terminal-Bench 2.0, GPT-5.5's 82.7% beats Claude Opus's 69.4% pretty decisively. But competitive landscapes are always messier than a single benchmark. Different models shine in different spots, and what works best depends entirely on your specific workflow.
What really separates GPT-5.5 is the agentic loop. Planning, tool use, self-verification, iteration. If your work involves handing an AI a meaty problem and trusting it to run autonomously for 20 minutes, resolving merge conflicts or re-architecting a module or grinding through a research workflow, that's where GPT-5.5 pulls ahead of the pack.
For context on where OpenAI is heading strategically and why they're pushing this hard, our coverage of [OpenAI's $12.2 billion fundraise] fills in some important blanks.
Getting the Most Out of It
If you're on Plus or Pro, GPT-5.5 should already be sitting in your model dropdown. Pick it and go.
A few things I've noticed help:
Throw messy, real-world problems at it. Don't sanitize your inputs. GPT-5.5 handles ambiguity surprisingly well and you don't need to spoon-feed every detail.
Let it plan. Fight the urge to micromanage. Seriously. The planning ability is the single biggest upgrade over 5.4 and if you keep interrupting, you're undermining the whole point.
Use Codex for engineering work. The agentic workflow there is where GPT-5.5 truly comes alive. Implementation, refactoring, debugging, and validation all flowing together continuously.
Replay tasks that older models botched. Fastest way to feel the gap.
Trust the self-correction loop. When the model checks its own work and circles back, let it run. That's where a huge chunk of the quality improvement actually lives.
Bottom Line
GPT-5.5 is a real step forward. Especially for developers and researchers who need AI holds context, plans across multiple steps, and actually finishes the work with genuine autonomy. The benchmarks are strong. Early feedback is consistently enthusiastic, borderline giddy in some cases. And the efficiency gains mean you might end up spending less despite getting substantially better results.
Perfect? Come on. Of course not. Pricing concerns are legitimate, especially at the API tier. No benchmark in the world fully captures how a model behaves in your codebase with your weird edge cases. But if you write code for a living or lean on AI for research workflows, GPT-5.5 deserves a serious test drive.
Throw your hardest problem at it. See what comes back.
And if you want to explore the broader ecosystem of tools built around models like this, check out our roundup of the [top agentic coding tools in 2026].
Sources
- OpenAI. "Introducing GPT-5.5." [openai.com/index/introducing-gpt-5-5]
- OpenAI. "GPT-5.5 System Card — Deployment Safety Hub." deploymentsafety.openai.com/gpt-5-5
- Yahoo Finance / Investing.com. "OpenAI releases GPT-5.5 with improved coding and research capabilities." uk.finance.yahoo.com
- Axios. "OpenAI releases 'Spud' GPT-5.5 model." axios.com
- SQ Magazine. "GPT 5.5 Released With Faster Coding and Research Skills." sqmagazine.co.uk
- VentureBeat. "OpenAI's GPT-5.5 is here, and it's no potato." venturebeat.com
- The Verge. "OpenAI says its new GPT-5.5 model is more efficient and better at coding." theverge.com
- CNBC. "OpenAI announces GPT-5.5, its latest artificial intelligence model." cnbc.com
- The New Stack. "OpenAI launches GPT-5.5, calling it 'a new class of intelligence.'" thenewstack.io