GPT-5.4 Just Hit 83% on Pro-Level Work. In One Generation, OpenAI Added Native Computer Control.

The Jump Nobody Predicted: 83% on Real Professional Work

When OpenAI shipped GPT-5.2 in late 2025, the benchmark that got buried in the announcement was GDPval — a 44-occupation professional test where the model's output is compared directly against actual professionals doing their jobs. GPT-5.2 scored 70.9%. Solid, but not alarming.

GPT-5.4, released March 5, 2026, scored 83.0%.

That's not a marginal improvement. That's 12 percentage points in a single model generation — roughly one year — moving from 'useful assistant' territory to 'outperforms professionals in 4 out of 5 tasks.' The benchmark covers 44 occupations: legal, medical, financial, technical, creative. The model now wins the majority of those head-to-head comparisons.

But the benchmark is almost the secondary story. The bigger structural change in GPT-5.4 is what it can do now — not just what it can say.

Native Computer-Use: The Architectural Shift

Every major AI lab has been working on computer-use capabilities — the ability for models to operate real software interfaces, click buttons, fill forms, and execute multi-step workflows autonomously. Until now, those capabilities existed as separate plugins, separate models, or separate product tiers.

GPT-5.4 ships with computer-use built in at the model level, available across both variants and in the standard API.

This isn't a minor feature flag. It means the same model you use to write a document can, in the same session, open a browser, navigate to a SaaS dashboard, extract data, update a spreadsheet, and email a summary — without switching products, chaining separate tools, or writing orchestration code.

OpenAI has already enabled this natively in Codex, and the computer-use API is live for developers. OSWorld-Verified and WebArena-Verified — the two most demanding real-world computer-use benchmarks — both show record scores.

1M Token Context: Long Workflows Without the Workarounds

GPT-5.4's 1 million token context window isn't just a spec-sheet number. It's the threshold at which certain categories of agentic tasks stop requiring chunking, summarization, or external memory workarounds.

At 1M tokens, you can feed an entire large codebase, a full quarter of customer support transcripts, or a company's complete documentation library — and ask questions or execute workflows across all of it in one pass.

For developers building agents, this changes the architecture. Less time engineering around context limits means more time on actual task logic.

33% Fewer Factual Errors

Factual accuracy has been a persistent frustration with large language models. GPT-5.4 reduces factual errors by 33% versus GPT-5.2, verified across internal benchmarks and third-party evaluations. OpenAI attributes this to improved grounding during training and better use of the extended context window to cross-check claims against source material.

In practice, this matters most for high-stakes use cases: legal research, medical information, financial analysis. A model that's wrong 33% less often in those contexts isn't just more useful — it's meaningfully safer to deploy.

Two Variants, One Strategy

GPT-5.4 ships in two configurations:

GPT-5.4 Thinking is the reasoning-focused variant. Extended computation time, more deliberate multi-step problem solving, higher accuracy on complex analytical tasks. This is the variant you'd reach for on hard coding problems, legal analysis, or scientific reasoning.

GPT-5.4 Pro prioritizes throughput — high performance at speed, optimized for production workloads where latency matters. Customer-facing products, real-time agents, high-volume API use.

ChatGPT now defaults to GPT-5.3 Instant for quick queries, GPT-5.4 Thinking for complex tasks, and GPT-5.4 Pro for power users. GPT-5.1 was deprecated the same week GPT-5.4 launched — a signal that OpenAI is compressing its support window as the release cadence accelerates.

What This Means for the Competitive Landscape

Three months ago, the standard benchmark comparisons had Claude 3.7 Sonnet and Gemini 2.0 Pro in competitive parity with GPT-5.2. GPT-5.4's 12-point professional benchmark jump resets that picture.

Anthropic and Google are presumably not standing still. Anthropic has been testing extended computer-use capabilities in Claude for months. Google has Gemini 2.5 in internal evaluation. But GPT-5.4 marks the first time since 2024 that one model has pulled meaningfully ahead on the most practically meaningful benchmark — real professional work output.

The other data point worth watching: OpenAI deprecated GPT-5.1 the same week it shipped GPT-5.4. That's roughly a 6-month support window per model generation. For enterprises and developers building on the API, the pace is accelerating. Planning for migration is no longer optional.

The Developer Implications

For developers building with the API, GPT-5.4 changes the calculus on several architectural decisions:

• External memory systems become optional for most use cases. At 1M tokens, context engineering is about quality, not survival. You don't need RAG to avoid hitting a wall — you need it to retrieve the right 50K tokens from a 10M token corpus.
• Agent tool-calling becomes simpler. Native computer-use reduces the surface area of orchestration code for workflows that involve navigating real software interfaces.
• Factual grounding for high-stakes applications is more viable. 33% fewer errors with the same verification workflow means lower human review burden at scale.

The computer-use API is available now. The 1M context window is live across both variants. GDPval 83% is the number to keep in your head when evaluating whether a workflow is ready to delegate to a model.

The Bottom Line

GPT-5.4 isn't a quiet incremental release. The combination of native computer-use, a 12-point professional benchmark leap, 1M context, and 33% better factual accuracy in a single generation is the kind of capability jump that changes what's buildable today versus six months ago.

The question worth asking isn't whether AI can now handle professional-level work — at 83%, the data says it can, most of the time. The question is what the next model generation looks like when that number starts approaching 90%.