GPT-5.5 Is Out: What AI Builders Need to Know

OpenAI released GPT-5.5 yesterday. It's available in ChatGPT and Codex for paid tiers only. Free tier doesn't get it. There's no API yet, but they say it's coming soon while they work through safety filters first.

Pricing is roughly 5 $/M for input tokens and 30 $/M for output tokens. That's about 2-3× what GPT-5.4 costs.

What actually changed

GPT-5.5 handles multi-step tasks with less babysitting. You give it a problem and it works through it without stopping every few steps. It also uses fewer tokens to get the same work done, which matters if you're running long sequences and watching the bill.

Where this shows up in practice

Coding agents: give it a GitHub issue, wait 15 minutes, come back to an open PR. The Cursor team noticed it doesn't stop early as much. Every time an agent restarts, you lose context and burn time, so staying on task is actually useful.

Data analysis work: keeping track of folder structure, dataset names, variable references across a 3-step pipeline. GPT-5.4 would lose the thread. This one doesn't. You see it on GeneBench and BixBench scores.

Clicking around actual computers: OSWorld tasks went from "doesn't work reliably" to "actually works." It navigates, finds things, makes changes.

Figuring out what broke: when something fails in a longer sequence, it can usually reason through why instead of just stopping and asking what to do next.

Where it probably doesn't help much

Single-turn questions or answers are faster but not massively cheaper to run. Don't expect your Q&A spend to drop.

Anything in cybersecurity or biomedical research hits friction. OpenAI flagged these as "high-risk," which means they're filtering more aggressively. You'll hit denials on red-team prompts until you apply for a "trusted" badge. If you're doing legitimate security or research work, you can get around this. But it's an extra step.

Pro vs base tier

Pro gets a stronger model (90% on BrowseComp vs 82% on base). If you're in Cursor 8 hours a day, it might be worth it. Otherwise, base is fine.

Should you migrate production now?

Only if you're running long-horizon agent work, it's token-hungry, and you can absorb the cost hit until the API arrives and prices stabilize. Everyone else: set a reminder to check the API announcement. When it drops, try it on whatever you're building and see if it actually works better on your code. Don't assume the benchmarks match your reality.

On the benchmarks

58.6% on SWE-Bench Pro (real GitHub issues), 80.5% on BixBench, 82.7% on Terminal-Bench. These are real numbers, but they're not your code. What matters is whether it stays focused on your system without restarting. Early reports say yes, but you'll know better once you test it.

OpenAI also did a custom experiment where it helped find a new proof in combinatorics. That's interesting as a proof-of-concept, but it's not the standard model. It was tuned specifically for that task.

The standard release is solid. Don't expect magic, but don't ignore it either.

OpenAI's full announcement

View tool: Claude Design

Promoted

Sponsored

Claude Design

Claude Design turns conversation into polished prototypes, slide decks, and one-pagers. Describe what you need, Claude builds a first version, and you refine through inline comments, edits, or sliders — kept on-brand via…

View tool