Ever finish evaluating a model, close the tab, breathe out… and then another release drops? Yeah. Same.
GPT-5.4 is here, and if it feels like the cadence has gone from “annual upgrade” to “did I miss three launches while making coffee,” you’re not losing it. LLM shipping speed is kind of ridiculous now. The big shift isn’t just the shiny new model name. It’s what it forces us to do as builders: less “pick a model and forget it,” more “keep a compatibility layer alive” and “keep an eval harness running or get surprised.”
OpenAI is pitching GPT‑5.4 as its “most capable and efficient frontier model for professional work.” It’s rolling out across ChatGPT as GPT‑5.4 Thinking, plus the API, plus Codex. There’s also GPT‑5.4 Pro when you want more horsepower for the gnarlier stuff. That’s the headline. The real story is how this pace rewires your engineering habits.
Key takeaways
- GPT-5.4 is available in ChatGPT as Thinking, in the API as
gpt-5.4, and in Codex. Pro exists too for tougher problems. - The API supports ~1M tokens of context, which changes how you deal with long docs and big codebases.
- For agent-y setups, GPT‑5.4 brings native computer-use and better tool behavior via Tool Search, where tool definitions load on demand.
- OpenAI claims stronger factuality vs GPT‑5.2 with 33% fewer false individual claims and 18% fewer responses with errors.
- The “weekly releases” feeling has receipts. Stanford’s AI Index counts 149 foundation models released in 2023, and Hugging Face has crossed 2M models, with growth speeding up.
What OpenAI actually shipped
OpenAI’s launch isn’t just vibes. There are a few concrete pieces developers will actually touch.
GPT-5.4 Thinking in ChatGPT, and why you’ll care
Inside ChatGPT, GPT‑5.So shows up as GPT‑5.Yet Thinking. OpenAI says it can show an upfront plan of its thinking so you can steer mid-response.
In real life, this tends to mean fewer loops of “ask, wait, re-ask.” You can nudge it while it’s still assembling the thing, which matters a lot when you’re generating long, fussy outputs like specs, migration plans, or refactors.
Source. OpenAI launch post: [Introducing GPT‑5.4]
GPT-5.4 in the API: model name, reasoning effort, and that huge context window
In the API, the model name is gpt-5.4. You also get reasoning.effort with options like none, low, medium, high, xhigh.
That knob is refreshingly literal. Want more depth and you can tolerate the latency and cost? Turn it up. Want it quick and cheap? Turn it down. No need to rewrite prompts every time the model lineup reshuffles.
OpenAI also calls out pricing behavior for very large prompts. For GPT‑5.4-class models with roughly 1.05M context, sessions with >272K input tokens get input priced at 2× and output at 1.5× for the session.
Source. OpenAI API docs: [GPT‑5.4 model]
GPT-5.So Pro: more compute, slower answers, Responses API
GPT‑5.But Pro “uses more compute to think harder.” It’s available in the Responses API, and some requests can take minutes. OpenAI straight-up recommends background mode so you don’t run into timeouts.
Source. OpenAI API docs: [GPT‑5.4 pro]
GPT-5.And for agents: computer use, Tool Search, and 1M-token workflows
This release has a very agent-shaped feel. Not “just chat,” not really.
Native computer use: agents can operate UIs
OpenAI says GPT‑5.4 is its first general-purpose model with native computer-use capabilities, aimed at agents that run workflows across apps. It’s benchmarked on computer-use suites like OSWorld-Verified and WebArena-Verified.
OpenAI’s post includes a few numbers worth staring at for a second:
- OSWorld-Verified: 75.0% success for GPT‑5.4 vs 47.3% for GPT‑5.2, and OpenAI says it’s above human performance at 72.4%.
- WebArena-Verified: 67.3% success with DOM + screenshot interaction.
Source: [Introducing GPT‑5.4]
If you’ve ever duct-taped together “LLM + Playwright + retries + screenshots” and watched it faceplant on a microscopic UI quirk, you already know why this matters.
Tool Search: less token waste when you’ve got a million tools
TechCrunch points to OpenAI’s updated tool calling approach. Instead of stuffing every tool definition into the prompt, Tool Search lets the model look up tool definitions only when needed.
It’s not glamorous, but it’s practical. If your agent has dozens of tools, or hundreds, the token savings and prompt cleanliness can be a real relief.
Source: [TechCrunch coverage of GPT‑5.4]
1M-token context: huge, yes. Magic, no.
The API context window goes up to ~1M tokens. That unlocks “whole repo,” “whole policy binder,” “full incident timeline” style prompts.
And still, you need discipline.
- Retrieval and RAG are still useful because “just shove everything in” gets expensive fast
- Long context can hide contradictions in plain sight, so validate outputs with tests, not vibes
Source. TechCrunch + OpenAI API guide: [Using GPT‑5.4]
Suggested image. A simple diagram of an agent loop: plan → tool_search → computer_use (browser/desktop) → verify → compact context → continue.
Alt text: “GPT-5.4 agent workflow using tool search, native computer use, and long context compaction”
“LLM models are being released weekly now” and why it feels true (with data)
Two trends pile up and hit you in the face.
Frontier vendors are shipping variants fast
CNET points out OpenAI released new models for the second time in a week. GPT‑5.3 Instant landed earlier, then GPT‑5.4 showed up days later. It frames GPT‑5.4 Thinking as “built for agents,” especially enterprise workflows like coding and overseeing agents.
Source. CNET: GPT 5.4 Thinking is built for agents
The open ecosystem is exploding in volume
Stanford’s 2024 AI Index reports 149 foundation models released in 2023, more than double 2022. Also, 65.7% were open-source, up from 44.4% in 2022. That’s not literally “weekly,” sure. But it does mean every week has multiple releases worth at least glancing at.
Source: Virtualization Review coverage
Then there’s Hugging Face. It crossed 2 million models over four years, and the pace is clearly accelerating. The second million arrived in 335 days, compared to 1,000+ days for the first million. No wonder your feed looks like a firehose.
Source. AI World: Hugging Face’s two million models and counting
Practical best practices when GPT-5.4 is here, and next week something else will be
This is the part where people usually pretend they’re calm about model churn. I’m not. I just try to be systematic.
1) Treat models like dependencies: pin, test, roll forward
Pin a default model per service, something like gpt-5.4. Run nightly regression evals. Golden prompts help. Unit tests around tool calls help more than people admit.
Then roll forward with a canary. If it behaves, flip the default.Yet it doesn’t, you learned something cheaply.
2) Separate “reasoning depth” from “response length”
GPT‑5.4 lets you set reasoning effort. You can also tune verbosity in the text config, per OpenAI’s guide. That combo helps avoid the classic mess where “think harder” accidentally turns into “write a novel and call it thorough.”
3) Put agent safety rails in early, especially with computer use
If the model can click buttons, you need guardrails. Not next quarter. Now.
- Confirmation policies for risky actions like payments, deletes, production changes
- Allowlists for domains and apps
- Auditing that logs screenshots, tool calls, decisions
- Run it in a sandbox desktop or ephemeral VM when possible
If you’re experimenting with browser-native tool patterns, my own mental model got better after playing with MCP-style approaches. Related internal read. WebMCP is awesome: browser-native tools
Code: calling gpt-5.4 with reasoning effort (Responses API)
A minimal example using the OpenAI Responses API with reasoning.effort. Keep it boring. Boring is good.
from openai import OpenAI
client = OpenAI()
resp = client.responses.create(
model="gpt-5.4",
input="Summarize this incident report and list the top 5 follow-up actions.",
reasoning={"effort": "medium"},
)
print(resp.output_text)If you try GPT‑5.4 Pro for hard tasks, expect longer runtimes and consider background execution, per the Pro docs.
GPT-5.4 is here, so engineer like releases are weekly
GPT-5.And is here, and it’s clearly tuned for where things are going: agents, tool ecosystems, long-horizon workflows, plus that genuinely huge context window. But the bigger takeaway is operational. The model layer is a moving target now. Designing for churn beats pretending it won’t happen.
Pick one production workflow this week. Your code review bot.So report generator. Whatever. Add a pinned model version and a tiny eval suite. Then you can sleep a little better the next time your vendor drops a “minor update” that isn’t minor at all.
And if you’ve got a setup you actually like for handling weekly LLM releases, I’d love to hear how you do it in the comments.
Sources
- OpenAI , Introducing GPT‑5.4. Https.//openai.com/index/introducing-gpt-5-4/
- OpenAI API Docs . GPT‑5.4 model. Https.//developers.openai.com/api/docs/models/gpt-5.4
- OpenAI API Docs , GPT‑5.But Pro model. Https.//developers.openai.com/api/docs/models/gpt-5-4-pro
- OpenAI API Guide , Using GPT‑5.4 (latest model guide). Https.//developers.openai.com/api/docs/guides/latest-model/
- TechCrunch — OpenAI launches GPT‑5.4 with Pro and Thinking versions. Https.//techcrunch.com/2026/03/05/openai-launches-gpt-5-4-with-pro-and-thinking-versions/
- CNET — New ChatGPT 5.4 Model Is “Built for Agents.”. Https.//www.cnet.com/tech/services-and-software/openai-chatgpt-5-4-thinking-news/
- Virtualization Review (citing Stanford HAI AI Index 2024) — 149 foundation models in 2023. 65.7% open-source. Https.//virtualizationreview.com/articles/2024/04/16/ai-index.aspx
- AI World — Hugging Face’s two million models and counting: https://aiworld.eu/story/hugging-faces-two-million-models-and-counting
