Six weeks. That’s how long GPT-5.4 got to be the flagship before OpenAI replaced it. GPT-5.5 shipped yesterday — April 23, 2026 — and it’s not a minor point release. This is the first fully retrained base model since GPT-4.5, and OpenAI is positioning it less as a chat model and more as an agent runtime. The benchmarks back that framing up, though with some caveats that the announcement blog conveniently glossed over.
I’ve had API access since the afternoon launch and have been running it through my usual workflows. Here’s what’s real, what’s marketing, and whether the upgrade actually makes sense for you.
What GPT-5.5 Actually Is
Previous GPT-5.x releases (5.1 through 5.4) were iterative refinements on the GPT-5 base. Post-training tweaks, RLHF rounds, distillation tricks. GPT-5.5 breaks that pattern — it’s a ground-up retrain with a new data mix and training methodology that OpenAI says was specifically optimized for multi-step agentic execution.
The practical translation: GPT-5.5 doesn’t just answer questions better. It’s built to plan tasks, use tools, navigate software interfaces, and chain actions together across long horizons without losing the thread. OpenAI’s internal pitch calls it “a new class of intelligence for real work,” which is exactly the kind of marketing copy you’d expect, but the benchmarks suggest they might not be entirely wrong this time.
It’s available now to Plus ($20/mo), Pro ($100/mo and $200/mo tiers), Business, and Enterprise ChatGPT subscribers. API access is live at $5 per million input tokens and $30 per million output tokens. There’s also GPT-5.5 Pro — the heavier, slower variant for harder problems — reserved for Pro, Business, and Enterprise tiers.
The Benchmarks That Matter
OpenAI touted state-of-the-art results across 14 benchmarks. Most of them, frankly, aren’t worth your time analyzing — incremental gains on saturated evals don’t tell you much. But three benchmarks stand out because they measure things you’d actually care about.
Terminal-Bench 2.0: 82.7%
This is the headline number, and it’s genuinely impressive. Terminal-Bench 2.0 tests complex command-line workflows — the kind where you need to plan a multi-step approach, iterate when things go wrong, coordinate across multiple tools, and maintain context over dozens of actions. GPT-5.5 scored 82.7%, blowing past Opus 4.7 at 69.4% and Gemini 3.1 Pro at 68.5%.
That’s not a marginal lead. A 13-point gap over Claude on an agentic coding benchmark is massive, especially when Opus 4.7 was the reigning champion just last week.
GDPval: 84.9%
GDPval tests whether models can do real knowledge work — not toy benchmarks, but tasks pulled from 44 actual occupations. Writing reports, analyzing spreadsheets, researching and synthesizing information, drafting emails. GPT-5.5’s 84.9% here reinforces the “agent for real work” positioning. This is the benchmark most relevant to anyone using AI as a daily productivity tool.
OSWorld-Verified: 78.7%
This one measures autonomous computer use — can the model actually operate software on a real desktop, clicking buttons, filling forms, navigating between apps? GPT-5.5 hits 78.7%, up from GPT-5.4’s 75%. For context, human performance on this benchmark hovers around 72%. So yes, GPT-5.5 now outperforms the average human at navigating a computer interface. Make of that what you will.
Where the Benchmarks Don’t Tell the Full Story
Here’s what OpenAI’s announcement was quieter about: SWE-bench Pro, the benchmark that tests real-world GitHub issue resolution end-to-end. GPT-5.5 scores 58.6%, which is solid — but Claude Opus 4.7 posts 64.3%. That’s a meaningful gap in the opposite direction. If your primary use case is large-scale codebase refactoring and complex multi-file changes, Claude still has the edge on the hardest software engineering tasks.
Also absent from the spotlight: GPQA Diamond, the graduate-level science reasoning benchmark. OpenAI didn’t highlight a score, which usually means the number isn’t a clear win. For reference, Gemini 3.1 Pro leads this benchmark at 94.3%.
The takeaway? GPT-5.5 dominates agentic execution and computer use. It’s weaker on the hardest pure reasoning and precision coding tasks. That’s a meaningful distinction depending on your workflow.
GPT-5.5 vs GPT-5.4: What Actually Changed
If you’re already on GPT-5.4, here’s the honest diff:
Agentic task completion is substantially better. Multi-step workflows that GPT-5.4 would fumble on step 6 or 7 now complete reliably. The model maintains context and adjusts plans mid-execution in ways that feel qualitatively different, not just incrementally better.
Computer use is smoother. The jump from 75% to 78.7% on OSWorld doesn’t sound dramatic, but in practice the error recovery is noticeably better. GPT-5.4 would sometimes get stuck in loops when a UI element didn’t appear where expected. GPT-5.5 handles those edge cases more gracefully.
Token efficiency improved. OpenAI claims GPT-5.5 uses significantly fewer tokens to complete the same Codex tasks as GPT-5.4, while matching per-token latency. If true (and my limited testing suggests it is), you’re getting more done per dollar even at the higher API price.
The API price doubled. Input tokens went from $2.50 to $5 per million. Output tokens went from $15 to $30 per million. GPT-5.5 Pro is $30/$180. That’s a meaningful cost increase, especially at scale. Whether the capability gains justify 2x pricing depends entirely on whether your use case hits GPT-5.5’s sweet spots.
Chat quality is… roughly the same. For standard Q&A, writing assistance, and single-turn tasks, I haven’t noticed a dramatic improvement over GPT-5.4. The gains are concentrated in multi-step, tool-using, agentic scenarios. If you mostly use ChatGPT to draft emails and answer questions, GPT-5.5 won’t feel like a revelation.
How It Stacks Up Against Claude and Gemini
The three-way comparison just got more interesting. Here’s where things stand as of this week:
| Benchmark | GPT-5.5 | Claude Opus 4.7 | Gemini 3.1 Pro |
|---|---|---|---|
| Terminal-Bench 2.0 | 82.7% | 69.4% | 68.5% |
| SWE-bench Pro | 58.6% | 64.3% | — |
| SWE-bench Verified | ~78% | 80.8%* | 78.8% |
| GDPval | 84.9% | — | — |
| OSWorld-Verified | 78.7% | 72.7% | — |
| GPQA Diamond | — | ~93% | 94.3% |
*Opus 4.6 score; Opus 4.7 likely higher but independent verification pending.
GPT-5.5 wins at: Agentic task execution, computer use automation, long-horizon multi-tool workflows, and general-purpose knowledge work. If you need an AI that can operate software autonomously and chain complex tasks together, OpenAI has the strongest offering right now.
Claude Opus 4.7 wins at: Precision software engineering — especially complex multi-file refactoring, large codebase analysis, and the hardest GitHub issue resolution. It also maintains an edge in extended thinking and careful analytical work. And at $5/$25 per million tokens for API, it’s cheaper than GPT-5.5’s $5/$30.
Gemini 3.1 Pro wins at: Graduate-level scientific reasoning, massive context windows for processing enormous documents, and — crucially — price. At $1.25/$5 per million tokens, Gemini costs roughly 6x less than GPT-5.5 on output. For workloads where Gemini’s quality is sufficient, that price gap is hard to ignore.
The pattern from my earlier Opus 4.7 vs GPT-5.4 vs Gemini 3.1 Pro comparison still holds, just with updated numbers: there’s no single best model. The right choice depends on what you’re building.
Pricing and Subscription Tiers: The Confusing Landscape
OpenAI’s pricing structure has gotten… complicated. Here’s the current state:
ChatGPT Free: GPT-5.3 with a 10-message cap every 5 hours. Fine for occasional use, painful for anything regular.
ChatGPT Go ($8/mo): Includes ads, limited features. Honestly, I’m not sure who this tier is for.
ChatGPT Plus ($20/mo): Gets GPT-5.5 Thinking access. This is probably the sweet spot for most individual users. You get the new model without breaking the bank.
ChatGPT Pro ($100/mo): Launched April 9 as a middle tier. Gives 5x Plus usage caps, GPT-5.4 Pro model access, and higher Codex quotas. GPT-5.5 Pro may land here soon.
ChatGPT Pro ($200/mo): The original Pro tier. Full GPT-5.5 Pro access, maximum usage caps, priority during peak times.
Business ($25/user/mo) and Enterprise (custom): Team-oriented with admin controls, compliance features, and GPT-5.5 access.
If you’re on Plus and debating the jump to Pro $100, my advice: wait a couple of weeks. GPT-5.5 is available on Plus, and the main Pro $100 advantage (GPT-5.4 Pro access) is now somewhat obsolete. OpenAI will likely update the Pro tiers to include GPT-5.5 Pro, and that’s when the upgrade conversation gets interesting.
API users should do the math carefully. At $5/$30 per million tokens, GPT-5.5 costs twice as much as GPT-5.4. If your workloads are cost-sensitive, consider whether GPT-5.4 at half the price still meets your needs — for many use cases, it will. Or test whether the improved token efficiency actually offsets the per-token price increase for your specific tasks.
The “Super App” Angle: What OpenAI Is Really Doing
The more interesting story isn’t the model itself — it’s the strategy. OpenAI explicitly positioned GPT-5.5 as an agent runtime rather than a chat model. They’re merging ChatGPT, Codex, and their AI browser into what Sam Altman calls a “super app.”
This matters because it signals where OpenAI is headed: they don’t want to sell you a chatbot. They want to sell you an AI employee that operates your computer, writes your code, manages your email, and handles your research — all within one platform. GPT-5.5’s benchmark profile makes sense through this lens. They optimized for the skills an autonomous digital worker would need, not for academic reasoning puzzles.
Whether that vision excites or terrifies you probably depends on your job description. But from a product strategy perspective, it’s the clearest statement yet about what “AGI” means to OpenAI: not some abstract superintelligence, but a reliable digital worker you can delegate real tasks to.
The Safety Question Nobody’s Talking About
OpenAI confirmed GPT-5.5 meets their “High” internal risk classification, one notch below “Critical.” They didn’t elaborate much on what specific capabilities push it into that category, but a model scoring 78.7% on autonomous computer use — beating average human performance — introduces real surface area for misuse.
When your AI can navigate websites, fill in forms, and chain actions together autonomously, the potential for things to go wrong scales up significantly. OpenAI says they’ve implemented “robust guardrails,” which is the corporate equivalent of “trust us.” I’d like to see more transparency about what those guardrails actually look like in practice, especially for the computer-use capabilities.
Should You Upgrade? A Practical Framework
Upgrade to GPT-5.5 if you:
- Rely heavily on agentic workflows — task automation, multi-step tool use, Codex
- Use computer-use features and need reliable autonomous operation
- Build products on the OpenAI API and need the best agentic performance
- Already hit limitations with GPT-5.4 on complex, multi-step tasks
Stay on GPT-5.4 if you:
- Primarily use ChatGPT for chat, writing, and single-turn tasks (you won’t notice much difference)
- Are cost-sensitive on API usage (GPT-5.4 at $2.50/$15 is half the price)
- Don’t use agentic or computer-use features regularly
Consider Claude Opus 4.7 instead if you:
- Your primary workflow is software engineering — especially large codebase refactoring
- You value precision and careful reasoning over speed and automation breadth
- You want cheaper API pricing ($5/$25 vs $5/$30) with arguably better coding performance
Consider Gemini 3.1 Pro instead if you:
- Cost is your primary constraint ($1.25/$5 per million tokens — dramatically cheaper)
- You need massive context windows for long documents
- Your use case is in the Google ecosystem
What Comes Next
GPT-5.5 launched less than 24 hours ago, so independent benchmarking is still in early stages. The numbers I’ve cited come from OpenAI’s own evaluation plus early third-party reports from MarkTechPost, VentureBeat, and Artificial Analysis. Expect more nuanced results over the coming weeks as developers stress-test the model on real-world workloads.
Also worth watching: Anthropic has Claude Mythos in preview, which reportedly matches or beats GPT-5.5 on Terminal-Bench 2.0. If that ships to general availability, the leaderboard reshuffles again. Google has been quiet lately, which in this market usually means they’re about to drop something.
The pace is relentless. Six weeks between GPT-5.4 and 5.5. A week between Opus 4.7 and GPT-5.5. If you’re exhausted by the upgrade treadmill, that’s a rational response. Pick the model that fits your actual workflow, ignore the FOMO, and revisit in a month when the dust settles and real-world benchmarks paint a clearer picture.