Six weeks to ship a new flagship model
In mid-March 2026, OpenAI released GPT-5.4. Six weeks later, on 23 April, GPT-5.5 is already taking its place. To put things in perspective, that is a release cycle two to three times shorter than what the industry was doing a year ago. And according to Greg Brockman, OpenAI's president and co-founder, that pace is going to continue.
On the menu for this latest wave: a new text model, a completely rethought image generator, and a clear strategy for merging ChatGPT, Codex and the Atlas browser into a single "super-app".
What is GPT-5.5's agent mode?
To understand the difference from previous versions, let's take a concrete example. Before GPT-5.5, if you wanted an AI to book a train ticket, you had to break the task down yourself: "open the SNCF site, search Paris-Lyon for Friday, pick the 2pm train, pay with my saved card." You were steering every step.
With GPT-5.5, you simply tell it: "find me a Paris-Lyon ticket on Friday afternoon that isn't too expensive". The model plans the steps on its own, opens the site, navigates the menus, checks the times, compares the fares, and only asks you to confirm when it needs you. This is what OpenAI calls an autonomous agent: an AI that manages on its own until the task is done.
The technical leap is measurable. On Terminal-Bench 2.0, a test that measures an AI's ability to chain complex computer commands without human help, GPT-5.5 reaches 82.7%. For comparison, Claude Opus 4.7 tops out at 69.4%.
GPT-5.5 is the first model fully retrained since GPT-4.5. All the intermediate versions (5.1, 5.2, 5.3, 5.4) were improvements built on the same base. This time, OpenAI redid everything from scratch: architecture, training data, objectives. That explains why the perceived jump in quality is more pronounced than usual.
How does GPT-5.5 compare with Claude Opus 4.7 and DeepSeek V4?
Three flagship models came out in the same week of April 2026: Claude Opus 4.7 on the 16th, GPT-5.5 and DeepSeek V4-Pro on the 23rd-24th. Here is the verdict on the benchmarks directly comparable across the three:
| Benchmark | GPT-5.5 | Claude Opus 4.7 | DeepSeek V4-Pro |
|---|---|---|---|
| Terminal-Bench 2.0 (command-line tasks) | 82.7% | 69.4% | 67.9% |
| SWE-Bench Pro (fixing real GitHub bugs) | 58.6% | 64.3% | 55.4% |
| BrowseComp (autonomous web browsing) | 84.4% | 79.3% | 83.4% |
| GPQA Diamond (scientific reasoning) | 93.6% | 94.2% | 90.1% |
| HLE without tools (multi-domain expert questions) | 41.4% | 46.9% | 37.7% |
The conclusion is nuanced: GPT-5.5 clearly dominates on automated workflows and web browsing, Claude Opus 4.7 retains the edge on production-quality code and complex reasoning, and DeepSeek V4-Pro remains competitive on web browsing while being 7 times cheaper. The three models are in fact complementary rather than strictly comparable.
Images 2.0: 99% accuracy on generated text
Two days before GPT-5.5, OpenAI launched ChatGPT Images 2.0 (technical name: gpt-image-2). If you have ever tried to generate an image with an AI and found that all the text looked like badly copied Chinese, for example "ENCHITTAS", "BURRTO" or "PIZZARIA", you will immediately understand why this update matters.
Images 2.0 achieves a text rendering accuracy of close to 99%, and works just as well in the Latin alphabet as in Chinese, Japanese, Hindi or Bengali. The model can generate a print-ready restaurant menu, a complete infographic, or a multi-panel comic strip, something that was completely impossible six months ago.
Images 2.0 is OpenAI's first image generator to incorporate reasoning before generation. Note that it is not the first in the world to do so: ByteDance's Seedream 4.0 already offered this principle in September 2025, and Seedream 5.0 took it further. The model can search the internet for references, plan the composition, and correct itself midway. This is what OpenAI calls "thinking" mode, reserved for paying subscribers. The "instant" mode, faster and cheaper, is more than enough for 80% of everyday uses.
Images 2.0 versus Google's Nano Banana Pro
Images 2.0's real direct competitor is not Seedream, it is Nano Banana Pro, Google's image generator released in November 2025 and based on Gemini 3 Pro Image. And here, the verdict from independent tests is unequivocal.
On the public LM Arena leaderboard (where anonymous testers vote for the best image between two proposals without knowing which model generated it), GPT Image 2 finished with an Elo score of 1,512 points barely 12 hours after its launch. Nano Banana Pro tops out at 1,270. A gap of 242 points, the equivalent of the gulf that separated Nano Banana Pro from the old DALL-E 3. An entire generation leap in a single release.
But the battle is not lost for Google. Nano Banana Pro remains better on certain specific points: photorealistic portraits, consistency across multiple faces, and it is already integrated into Adobe Firefly, Photoshop, Figma and Canva with legal protections on copyright. For many companies, switching suppliers would cost more than the gain in quality.
| Criterion | GPT Image 2 | Nano Banana Pro |
|---|---|---|
| LM Arena score (Elo) | 1,512 | 1,270 |
| Accuracy of generated text | 99% | ~85% |
| Photorealistic portraits | Very good | Slightly better |
| Average speed | 3 seconds | 10 to 15 seconds |