OpenAI just shipped a model that can use your computer better than most humans can.

That’s not hyperbole. GPT-5.4 hit a 75% success rate on OSWorld-Verified, a benchmark that measures an AI’s ability to navigate a real desktop using screenshots, mouse clicks, and keyboard actions. The human benchmark on the same test? 72.4%. This is the first time an OpenAI model has crossed that threshold.

GPT-5.4 launched March 5, 2026, rolling out across ChatGPT (as GPT-5.4 Thinking), the API, Codex, and GitHub Copilot. It’s available in three flavors: standard, Thinking, and Pro. I’ve been running it since launch day. Here’s what’s actually different, what the benchmarks mean in practice, and whether you should care.

Quick Verdict: GPT-5.4

Aspect	Rating
Overall Score	★★★★½ (4.5/5)
Best For	Agentic workflows, long-document analysis, professional coding
Pricing (API)	$2.50/1M input, $15/1M output (standard context)
Context Window	1M tokens (API), largest OpenAI has offered
Computer Use	75% OSWorld-Verified (vs 72.4% human baseline)
Hallucination Improvement	33% fewer false claims vs GPT-5.2
Variants	GPT-5.4, GPT-5.4 Thinking, GPT-5.4 Pro

Bottom line: The computer-use capability is genuinely new territory for OpenAI. The 1M context window closes the gap with competitors. Pricing is competitive. For agentic and long-horizon work, this is OpenAI’s strongest play yet. For everyday ChatGPT use, the improvements over GPT-5.3 Instant are less dramatic.

Try GPT-5.4 in ChatGPT

What Makes GPT-5.4 Different

Three things set this release apart from everything OpenAI has shipped before.

Native computer use. Not a plugin, not a wrapper, not an API bolt-on. GPT-5.4 can look at a screenshot of your desktop, understand what’s on screen, move the mouse, click buttons, type text, and navigate multi-step workflows across applications. This capability is baked into the model at the architecture level, available through Codex and the API.

1 million token context window. OpenAI’s previous ceiling was 400K tokens with GPT-5.3 Instant. GPT-5.4 jumps to 1M in the API, matching what other frontier models offer and getting closer to Gemini’s 2M. For long-horizon agent tasks that need to maintain context across extended workflows, this matters.

Unified reasoning + coding + agentic model. Previous OpenAI releases separated fast models (Instant series) from smart models (Pro/Codex). GPT-5.4 combines reasoning, coding, and computer-use into a single system. You don’t have to choose which capability you need before picking your model.

Computer Use: The Headline Feature

This is the capability that justifies GPT-5.4’s existence as a new release rather than an incremental update.

What it actually does: GPT-5.4 can interpret screenshots of a desktop environment, identify UI elements (buttons, text fields, menus, icons), and issue mouse/keyboard commands to accomplish tasks. Think of it as an AI that can sit at your computer and operate it the way you would.

The numbers:

Benchmark	GPT-5.2	GPT-5.4	Human Baseline
OSWorld-Verified	47.3%	75.0%	72.4%
WebArena Verified	Not reported	Record score	N/A

That jump from 47.3% to 75.0% on OSWorld is enormous. It’s a 58% relative improvement, and it crosses the human performance threshold. This isn’t marginal progress.

What this means in practice: An AI agent running GPT-5.4 can navigate web applications, fill out forms, move data between tools, operate desktop software, and execute multi-step workflows that previously required a human clicking through them. For anyone building automation that involves interacting with software that doesn’t have a clean API, this is a shift.

The caveats: 75% success rate means 1 in 4 attempts still fails. That’s not reliable enough for unsupervised production use on critical workflows. You need human oversight, error handling, and retry logic. But for semi-automated workflows where a human reviews the output? Highly practical.

For context on how this compares to what Anthropic and Google are doing with agentic AI, see our best AI agents guide.

1M Token Context: Closing the Gap

OpenAI has been behind on context windows for a while. GPT-5’s original 128K felt cramped. The 400K bump with GPT-5.3 Instant helped but still trailed competitors. GPT-5.4’s 1M tokens in the API is OpenAI finally playing catch-up.

What 1M tokens buys you:

• A full enterprise codebase (~400K-600K tokens)
• Hundreds of pages of legal documents
• An entire book series for analysis
• Extended agent sessions that maintain context for hours

The pricing structure reflects the context jump. Standard rates apply up to 272K tokens ($2.50/1M input). Past that threshold, input pricing doubles to $5.00/1M. That tiered approach is new for OpenAI and worth factoring into cost projections for high-context workloads.

For coding tasks specifically, the 1M window means GPT-5.4 can hold an entire project in context while making changes, something that matters when you compare it against Claude Opus 4.6 for agentic coding work.

Professional Performance: The GDPval Numbers

OpenAI claims GPT-5.4 matches or exceeds industry professionals in 83% of comparisons across 44 occupations on GDPval, their professional competency benchmark.

That’s a bold claim. Here’s how to interpret it.

GDPval measures whether model outputs are preferred over work produced by domain professionals in blind evaluations. 83% across 44 occupations means the model’s output was rated equal or better in roughly 36-37 of those occupation categories.

Where this matters most: Document drafting, data analysis, code review, research synthesis, and structured problem-solving. These are tasks where LLMs have been steadily improving, and GPT-5.4 appears to represent another step forward.

Where skepticism is warranted: The benchmark doesn’t measure reliability under edge cases, creative originality, or the ability to push back on flawed premises. Professional competency and professional judgment aren’t the same thing. An AI that writes better first drafts than 83% of professionals still needs a professional to know when the draft is wrong.

Hallucination and Error Reduction

OpenAI reports 33% fewer false claims and 18% fewer errors per response compared to GPT-5.2. These are significant improvements if they hold up in production use.

For comparison, GPT-5.3 Instant claimed a 26.8% hallucination reduction over GPT-5.2 with web access. GPT-5.4’s 33% figure against the same baseline suggests continued progress on factual reliability.

In my early testing: The model is noticeably more cautious about making claims it can’t support. It hedges more on uncertain topics rather than confabulating. That’s the right tradeoff for professional use. I’d rather see “I’m not confident about this” than a plausible-sounding fabrication.

The error reduction is harder to verify independently this early. I’ll update this section as community benchmarks come in over the next few weeks.

GPT-5.4 Variants: Standard vs Thinking vs Pro

Three versions, three use cases.

Variant	Access	Best For	Key Difference
GPT-5.4	API, Codex	Developers, agentic pipelines	Full 1M context, computer use
GPT-5.4 Thinking	ChatGPT Plus ($20/mo), 80 msgs/3hr	Daily professional use	Extended reasoning chain
GPT-5.4 Pro	ChatGPT Pro ($200/mo), Enterprise	Heavy professional workloads	Dedicated GPU slice, unlimited access

GPT-5.4 Thinking is what most ChatGPT subscribers will interact with. It uses an extended reasoning chain before responding, similar to how the o-series models work but integrated into the main GPT architecture. The 80-message limit every 3 hours is restrictive for power users.

GPT-5.4 Pro removes the message cap and adds a dedicated GPU allocation, meaning faster inference with no queue. At $200/month, it’s positioned squarely at professionals who bill by the hour and need the model available instantly. The API pricing for Pro is $30/1M input tokens, a significant premium over standard.

Standard GPT-5.4 through the API is where most developers will work. The $2.50/1M input pricing is competitive with the market, and the 1M context window is the largest OpenAI has ever offered.

GPT-5.4 vs Claude Opus 4.6: The Honest Comparison

This is the matchup everyone is watching.

Capability	GPT-5.4	Claude Opus 4.6
Computer Use	75% OSWorld	Claude had computer use first, similar capability
Context Window	1M tokens	200K standard (1M beta)
Coding	Strong, improving	Still the benchmark leader
API Pricing (input)	$2.50/1M	$15/1M
Reasoning	Very strong (Thinking variant)	Excellent
Agentic Tasks	Native computer use + long context	Strong agent framework via MCP

Where GPT-5.4 wins: Pricing (6x cheaper on input tokens), context window size in production, native computer-use benchmark scores, and availability across ChatGPT/Copilot/Codex simultaneously.

Where Claude still leads: Coding quality at the frontier (especially multi-file agentic coding), nuanced reasoning on ambiguous problems, and a more mature agent ecosystem through MCP servers.

My honest take: GPT-5.4 closes the gap on Claude for agentic work. The computer-use numbers are genuinely competitive. But for pure coding and deep reasoning, Claude Opus 4.6 still has an edge that matters if those are your primary workloads. The pricing advantage tilts toward OpenAI for high-volume use.

For the full breakdown of how these two compare across all dimensions, see our ChatGPT vs Claude comparison.

Where GPT-5.4 Struggles

No cheerleading. Here’s what doesn’t work.

Computer use reliability. 75% sounds impressive until you’re building a production workflow. One-in-four failures means you need solid error handling, human review loops, and retry logic. This isn’t “set and forget” automation yet.

Context window cost scaling. The doubled pricing past 272K tokens means that 1M context window gets expensive fast for sustained use. A full 1M-token request costs roughly $5 in input tokens alone. For high-frequency, large-context workloads, the math adds up.

ChatGPT message limits. 80 messages every 3 hours on the Thinking variant is tight for serious work. The Pro plan at $200/month removes this cap but that’s a steep price for unlimited access.

Pricing Breakdown

Access Method	Cost	Notes
ChatGPT Plus	$20/month	GPT-5.4 Thinking, 80 msgs/3hr
ChatGPT Pro	$200/month	Unlimited GPT-5.4 Pro, dedicated GPU
API (standard, <272K)	$2.50 input / $15 output per 1M	Competitive with market
API (standard, >272K)	$5.00 input / $15 output per 1M	Doubled input rate
API (Pro)	$30.00 input per 1M	Premium reasoning tier
GitHub Copilot	Included in plan	Generally available since March 5

For current rates, check OpenAI’s pricing page.

Who Should Use GPT-5.4

Strong fit:

• Developers building AI agents that need to interact with real software through screen-based interfaces
• Teams processing large document sets that benefit from the 1M context window
• GitHub Copilot users who get GPT-5.4 automatically and want stronger reasoning
• API-heavy workloads where the $2.50/1M input pricing offers cost savings over Claude

Wait and evaluate:

• Current GPT-5.3 Instant users who primarily need fast chat and don’t need computer use or 1M context
• Teams with existing Claude workflows for coding, where the switching cost may not be justified yet
• Anyone relying on computer-use for critical automation until reliability improves past the 75% threshold

Look elsewhere:

• Budget-sensitive teams doing high-volume inference (see our AI models comparison for cheaper alternatives)
• Coding-first teams that need the absolute best code generation (Claude Opus 4.6 still leads)
• Teams needing 2M+ context (Gemini 3.1 Pro still has the largest window)

How to Get Started

ChatGPT users:

1. Open ChatGPT
2. Select GPT-5.4 Thinking from the model picker
3. Plus subscribers get 80 messages every 3 hours; Pro subscribers get unlimited access

API users:

1. Update your model parameter to the GPT-5.4 model ID
2. Set context window limits based on your pricing tolerance (watch the 272K threshold)
3. Enable computer-use capabilities through the Codex integration for agentic workflows
4. Test against existing evaluation suites before migrating production traffic

GitHub Copilot users:

1. GPT-5.4 is already generally available as of March 5
2. No action needed; it’s the default model

The Bottom Line

GPT-5.4 is OpenAI’s most significant release in months. The computer-use capability at 75% OSWorld success genuinely changes what’s possible for AI agents interacting with software. The 1M context window eliminates one of OpenAI’s biggest disadvantages. The pricing is competitive.

But the nuance matters. Computer use at 75% reliability isn’t production-ready without human oversight. The context window gets expensive past 272K tokens. And for pure coding, Claude still holds the crown.

My verdict: If you’re building agentic workflows that interact with real desktop and web applications, GPT-5.4 is the model to evaluate first. The computer-use benchmarks represent a step function improvement for OpenAI. For everything else, evaluate based on your specific workload.

---

Frequently Asked Questions

What is GPT-5.4’s computer-use capability?

GPT-5.4 can interpret screenshots of a desktop environment, identify UI elements, and issue mouse and keyboard commands to accomplish tasks. It achieves a 75% success rate on OSWorld-Verified, surpassing the 72.4% human baseline. This is available through Codex and the API, not the standard ChatGPT interface.

How much does GPT-5.4 cost?

ChatGPT Plus ($20/month) includes GPT-5.4 Thinking with 80 messages every 3 hours. ChatGPT Pro ($200/month) offers unlimited GPT-5.4 Pro access. API pricing starts at $2.50 per million input tokens and $15 per million output tokens, with input pricing doubling past 272K tokens.

What’s the difference between GPT-5.4, GPT-5.4 Thinking, and GPT-5.4 Pro?

Standard GPT-5.4 is the API/Codex model with full 1M context and computer use. GPT-5.4 Thinking adds an extended reasoning chain and is available in ChatGPT. GPT-5.4 Pro is the highest-performance variant with a dedicated GPU slice, available on the $200/month Pro plan.

Is GPT-5.4 better than Claude Opus 4.6?

It depends on the workload. GPT-5.4 leads on computer-use benchmarks, has a larger production context window (1M vs 200K), and is significantly cheaper per token. Claude Opus 4.6 still leads on coding benchmarks and nuanced reasoning. For agentic automation, GPT-5.4 has the edge. For code generation, Claude maintains its advantage.

How much better is GPT-5.4 at avoiding hallucinations?

OpenAI claims 33% fewer false claims and 18% fewer errors per response compared to GPT-5.2. These are meaningful improvements if verified by independent testing. For reference, GPT-5.3 Instant showed a 26.8% hallucination reduction over the same baseline.

Can GPT-5.4 really use a computer better than a human?

On the specific OSWorld-Verified benchmark, yes: 75% vs 72.4% for humans. But benchmarks measure structured tasks in controlled environments. Real-world computer use involves ambiguity, unexpected states, and edge cases that benchmarks don’t capture. The 25% failure rate means human oversight is still essential for production workflows.

---

Last updated: March 6, 2026. Benchmarks sourced from OpenAI’s GPT-5.4 announcement and system card. Pricing verified against openai.com/api/pricing. Computer-use success rates from OSWorld-Verified evaluation.