OpenAI ships GPT-5.5: what actually changes for builders

OpenAI shipped GPT-5.5 on April 24. The launch post leads with benchmark numbers, but the practical changes for anyone running production AI live in three places: latency, pricing, and the agent loop.

What actually changed

Three things matter, in order.

Latency dropped meaningfully. First-token latency on the API is roughly 35% lower than GPT-5, and total time-to-completion on long outputs is down about 20%. For agent products that chain 8–12 tool calls per task, this is the change that takes them from "feels slow" to "feels fluid." If you've been running parallel calls to mask latency, you can probably pull that complexity back out.

Pricing came down. Input tokens are now 30% cheaper, output tokens 20% cheaper. For agent-style workloads where context windows balloon across iterations, this is bigger than it sounds. A workflow that costs $0.40 per task this morning costs around $0.28 tomorrow. Roll that across millions of agent tasks per month and the math is real.

Agent-loop reliability is up. OpenAI's own evals show a 14-point jump on tool-use tasks that involve more than five sequential calls. We've been testing it against our own internal agent harness for two days, and the qualitative read is that it's noticeably less likely to lose its place in long workflows or hallucinate tool arguments. The model card doesn't break this out cleanly, but it's there.

Where it still loses

GPT-5.5 does not move the needle on coding. Our internal SWE-style eval shows it roughly tied with GPT-5 on real-world bug fixes, and noticeably behind Claude on multi-file refactors. If you're running a coding agent today, this release is not a reason to switch back.

It also still has the same long-context degradation curve OpenAI has had since the GPT-5 launch — recall above 128k tokens drops faster than the marketing suggests. If your workload depends on stable retrieval across 200k+ contexts, do your own evals before migrating.

Why this matters

The bigger story isn't the release itself — it's the price-performance shift on agent-style workloads. The economics of "chain a frontier model 10 times and use it like a function" just got significantly better. That's the workload OpenAI most wants to win, and they're pricing it that way.

For anyone running an agent product on the OpenAI API, the upgrade is essentially free — same context window, same tool calling interface, lower price, faster output. The only reason not to switch is if you've tuned prompts heavily against GPT-5's specific failure modes.

For anyone running on Claude or open-weights, this doesn't change the calculus much yet. The bigger story for self-hosters is the Llama 4 release on the same day — covered in the open-weights story.

What we'd do

If you're already on the OpenAI API: switch the default model to GPT-5.5 today. Re-run your evals before the end of the week. Watch your agent task cost in production — it should drop.

If you're on Claude for coding: stay there. There's nothing in this release that changes the coding picture.

If you're self-hosting: the more interesting decision this week is whether Llama 4 405B is good enough to displace your hosted spend. We dug into that in a separate story.