I Spent 12 Hours With GPT-5.4. Here’s What Nobody Is Talking About.

GPT-5.4 shipped on March 5th into the middle of a boycott. Most people were arguing about the Pentagon contract. Meanwhile, a few things quietly changed about how you should be working.

OpenAI launched GPT-5.4 while 2.5 million users were mid-cancellation over a DoD contract. The discourse was all ethics, all politics, all noise.

Which means most people haven’t actually looked at what changed.

Some things genuinely changed. Not hype. Not incremental. Things that shift what’s buildable right now.

This isn’t a hype piece. There are real caveats and I’ll name them. But if you’re building with AI, shipping with AI, or doing knowledge work at volume, here is the honest breakdown before your competitors figure it out.

What Is GPT-5.4 and Which Version Are You Actually Using?

GPT-5.4 shipped as three distinct variants, and the naming is already causing confusion.

GPT-5.4 Thinking goes to Plus, Team, and Pro users inside ChatGPT. It’s the reasoning-focused version. It shows you an upfront plan before executing, which you can interrupt and redirect mid-response. This is what most people interact with daily.

GPT-5.4 Pro is for Pro and Enterprise plans through ChatGPT and the API. Maximum compute. Built for complex, multi-step tasks where shortcuts are not acceptable.

GPT-5.4 via API targets developers. Context window up to 1 million tokens. Priced at $2.50 per million input tokens and $20 per million output. Input costs run 43% higher than GPT-5.2, but more on why that number is misleading in a minute.

Free plan users have no access yet. Plus users who don’t see it: it’s still rolling out as of March 6th.

Did GPT-5.4 Cross a Meaningful Threshold?

Yes. One benchmark in particular is worth understanding: OSWorld-Verified. It measures how well a model can operate a real desktop computer through screenshots, mouse commands, and keyboard inputs.

GPT-5.4 scored 75.0%. The human baseline on the same test is 72.4%. GPT-5.2 scored 47.3%.

That’s not incremental. From 47.3% to 75.0% is a 58% improvement. More importantly, GPT-5.4 is now operating a real computer better than a typical person does.

What that means practically: agentic workflows that involve actual software, not just API calls but real desktop environments, are now viable in production in a way they were not six months ago.

Computer-use capability is native to GPT-5.4 Pro and available through the API. If you’re building agents, this is the variant to test first.

What Is Tool Search and Why Does It Matter for Multi-Tool Agents?

Tool Search is OpenAI’s redesign of how the API handles tool calling. Previously, every request loaded definitions for all tools upfront regardless of which ones the model needed. With Tool Search, tool definitions load dynamically. The model looks up what it needs, when it needs it.

If your agent had access to 40 or more tools, you were paying to load all 40 definitions on every single call. At volume, that’s a token and latency problem with no clean fix.

Mainstay, a property tax workflow company, reported their agents running approximately 3x faster with around 70% fewer tokens compared to GPT-5.2 after switching to Tool Search. For businesses paying per API call at volume, that compression more than offsets the 43% input price increase.

If you’re building multi-tool agents and haven’t hit the definition-bloat problem yet, you will. Build with Tool Search in mind now.

What Can You Actually Do With a 1 Million Token Context Window?

A 1 million token context window means you can pass a full large codebase, an entire contract set, or a long conversation history into a single API call. No chunking. No retrieval gymnastics. No file summarization workarounds.

Three practical unlocks worth knowing:

Codebase analysis. Feed the entire codebase, not just the relevant files. Ask GPT-5.4 to map dependencies, spot inconsistencies, or explain architecture decisions across the whole thing simultaneously.

Document-heavy professional work. Contract review, financial modeling, research compilation. Send the full document set and ask questions across all of it at once. GPT-5.4 scored 91% on BigLaw Bench, which measures accuracy across complex multi-document legal tasks.

Long session continuity. For products where users expect context across long sessions, you can pass significantly more conversation history per call without hitting limits. Not a replacement for a real memory architecture, but useful for prototypes and simpler products.

Honest caveat: a bigger context window amplifies what’s already working. It does not fix bad retrieval design or weak evaluation coverage. Get your fundamentals right first.

Which Professional Workflows Actually Changed This Week?

Beyond the model itself, OpenAI launched ChatGPT for Excel and Google Sheets in beta: an embedded version of ChatGPT living directly inside spreadsheets for building, analyzing, and updating financial models.

They also added integrations with FactSet, MSCI, Third Bridge, and Moody’s, letting teams pull market data and company data into a single workflow inside ChatGPT.

These are not consumer features. They are a direct push into enterprise knowledge work and a direct shot at Anthropic’s Claude for Financial Services launch last year.

If you’re building B2B AI products in finance, legal, or professional services: the competitive landscape shifted this week. Your potential clients now have a natively integrated option from OpenAI. Factor that into your positioning before your next sales conversation.

What Is Playwright Interactive and Why Is Nobody Talking About It?

Playwright Interactive is an experimental Codex skill that allows GPT-5.4 to visually debug web and Electron apps. More importantly, it lets the model test an app it is building as it builds it.

OpenAI’s demo shows a fully playable theme park simulation game built from a single lightly specified prompt. The model used Playwright Interactive to browse-test its own output mid-generation and correct it.

That is a model completing a feedback loop that previously required human eyes: build, test visually, identify failure, fix. All in one generation cycle.

It is experimental. It will not work reliably on complex projects yet. But if you’re building frontend-heavy apps with Codex, play with it now while most people haven’t noticed it exists.

Should You Migrate From GPT-5.2 or GPT-5.3-Codex Right Now?

Depends on your primary use case.

Coding performance moved from 56.8% to 57.7% on SWE-Bench Pro. If you’re on GPT-5.3-Codex doing primarily coding work, this is not a “drop everything and migrate” moment. The gains on that specific dimension are marginal.

What you cannot ignore: GPT-5.2 Thinking enters legacy status on June 5th, 2026. If you have production workflows pinned to it, start migration planning now, not in May.

The staggered rollout is also real. As of March 6th, not all users see GPT-5.4 yet. Don’t change production systems based on access you don’t have.

Who Should Actually Care About GPT-5.4 Right Now?

Multi-tool agentic workflow builders. Tool Search is worth testing immediately. The token efficiency gains are real and documented at scale.

Knowledge workers doing heavyweight document tasks. Contract review, financial modeling, research compilation. The 1 million token context window and 91% BigLaw Bench score make this genuinely useful for production work.

Frontend developers on Codex. Playwright Interactive is experimental and worth exploring now before it becomes standard workflow everyone knows about.

B2B product teams in finance or legal. The Excel integration and enterprise data partnerships changed your competitive landscape this week. Understand what your clients now have natively before your next sales conversation.

Developers on 5.3-Codex doing primarily coding work. You can probably wait. The benchmark gains are not meaningful enough to justify migration disruption right now.

Frequently Asked Questions About GPT-5.4

What is the difference between GPT-5.4 Thinking, GPT-5.4 Pro, and GPT-5.4 API?

GPT-5.4 Thinking is the reasoning-focused consumer version for Plus, Team, and Pro ChatGPT users. GPT-5.4 Pro is maximum-compute for Pro and Enterprise plans. GPT-5.4 via API targets developers with a 1 million token context window and native tool-use capabilities.

How much does GPT-5.4 cost compared to GPT-5.2?

API input tokens are priced 43% higher than GPT-5.2 at $2.50 per million input tokens and $20 per million output tokens. Token efficiency gains from Tool Search can offset this at scale for multi-tool agent workflows.

Is GPT-5.4 better than Claude Sonnet 4.6 for coding?

On SWE-Bench Pro, GPT-5.4 scored 57.7%, up from 56.8% on GPT-5.3-Codex. Head-to-head benchmarks against Claude Sonnet 4.6 are still emerging. For pure coding at this tier, workflow fit matters more than benchmark points.

When is GPT-5.2 Thinking being deprecated?

GPT-5.2 Thinking enters legacy status on June 5th, 2026. Production workflows pinned to it need migration planning before that date.

What is OSWorld-Verified and why does the GPT-5.4 score matter?

OSWorld-Verified measures how well a model operates a real desktop computer through screenshots and input commands. GPT-5.4 scored 75.0% against a human baseline of 72.4% and GPT-5.2’s 47.3%. It is the first model to exceed human-level performance on this benchmark, which makes real desktop-environment agentic workflows viable for production use.

The discourse around GPT-5.4 this week has been almost entirely about OpenAI’s ethics, the QuitGPT boycott, and the Pentagon deal.

Those conversations matter.

But the model shipped. It is in production. Native computer use crossing human-level performance, Tool Search eliminating definition bloat at scale, a 1 million token window for document-heavy work: these are real changes to what’s buildable right now.

Understand the tool. Decide what you do with the company separately.