Two weeks, two heavyweight releases
On April 16, Anthropic shipped Claude Opus 4.7. One week later, on April 23, OpenAI dropped GPT-5.5. Two frontier models, both claiming the crown, both available right now if you pay for them.
So the obvious question showed up in my inbox about forty times since yesterday: which one should I actually use?
I spent almost 10 hours running both side by side on real work. Coding projects, writing tasks, agent runs, the kind of messy stuff you’d do on a normal Friday. No cherry-picked prompts, no benchmark gymnastics, just real output I could actually ship. Here’s what I found.
Claude Opus 4.7 is Anthropic’s strongest publicly available model. It follows instructions more literally than 4.6, reads images at three times higher resolution, and it’s tuned for long autonomous work. If you’ve been using Claude Code or Cursor lately, this is probably already your default.
GPT-5.5 is something different. It’s the first full retraining of OpenAI’s base model since GPT-4.5. Internally they called it Spud, which I love. It’s natively multimodal, meaning text, image, audio, and video all run through the same architecture instead of being stitched together from separate systems. It’s available now inside ChatGPT and Codex for paid users. The API is listed as coming “very soon,” which is OpenAI-speak for whenever the safety team signs off.
Two models, two different philosophies. That’s really the whole story.
This is where things get uncomfortable.
Claude Opus 4.7 costs 5 dollars per million input tokens and 25 dollars per million output tokens.
GPT-5.5 costs 5 dollars per million input tokens and 30 dollars per million output tokens.
So GPT-5.5 is about 17% more expensive on output, which matters a lot on long generations. And here’s the kicker: that’s double what GPT-5.4 used to cost. OpenAI straight up doubled the price.
Their pitch is that GPT-5.5 uses roughly 40% fewer output tokens to finish the same task, so your real bill only climbs by about 20%. That’s their math. Your billing dashboard will tell you if it’s honest.
There’s also GPT-5.5 Pro at 30 dollars per million input and 180 dollars per million output. That one is for hard research work where getting the right answer once is worth more than any amount of compute. Do not use Pro for daily coding. You will feel it in your credit card statement.
One thing worth flagging on the Claude side. Opus 4.7 ships with a new tokenizer that packs slightly more tokens per message than 4.6. Anywhere from 1.0 to 1.35 times depending on your content. The per-token price stayed the same, but the same prompt can cost you a bit more in raw tokens than it used to. Worth measuring before you panic about a sudden bill jump.
The locked part has the real meat: four build tests, runtime gaps, token bills, SWE-Bench vs Terminal-Bench, where Claude still feels smarter, where GPT-5.5 feels faster, the long-context problem, and the practical rule I now use instead of asking which model is best.
