OpenAI GPT-5.4 release: unified reasoning, coding, and computer-use model with 1M-token context

TL;DR

OpenAI released GPT‑5.4, its most capable and efficient frontier model for professional work, now available in ChatGPT (as Thinking), the API, and Codex, with a higher‑performance Pro variant. The model adds native computer‑use, up to 1 M‑token context, better tool search, and stronger reasoning, coding, and knowledge‑work abilities, delivering higher accuracy, lower token usage, and faster speeds across a wide range of benchmarks.

Overview of the Release

OpenAI announced the rollout of GPT‑5.4 on March 5 2026. The model is deployed in three product lines:

  • ChatGPT – as GPT‑5.4 Thinking (standard) and GPT‑5.4 Pro (maximum performance).
  • API – accessible via the gpt-5.4 and gpt-5.4-pro endpoints.
  • Codex – integrated with the new coding‑focused capabilities.

GPT‑5.4 unifies recent advances in reasoning, coding, and agentic workflows, inheriting the coding strengths of GPT‑5.3‑Codex while extending performance on spreadsheets, presentations, documents, and tool‑rich environments.

Key Technical Improvements

Unified Reasoning and Planning

  • Up‑front planning – In ChatGPT, the model now outlines its thinking before generating a response, allowing users to adjust the plan mid‑response.
  • Longer context – Supports up to 1 M tokens, enabling agents to maintain coherence over very long tasks.
  • Reduced token consumption – Uses significantly fewer tokens than GPT‑5.2 for comparable reasoning, lowering cost and latency.

Native Computer‑Use

  • First general‑purpose model with built‑in computer‑use capabilities, allowing agents to control applications via libraries such as Playwright and to issue mouse/keyboard actions from screenshots.
  • Tool search – Instead of loading all tool definitions upfront, the model can retrieve a tool’s definition on demand, cutting token usage by ≈47 % in large tool‑heavy workflows.
  • Tool calling accuracy – Improves multi‑step tool usage on benchmarks like Toolathlon and τ2‑bench, achieving higher accuracy with fewer interaction turns.

Vision and Perception

  • Introduces an original image‑detail level (up to 10.24 M pixels) and expands the high level to 2.56 M pixels, improving localization, click accuracy, and dense‑image understanding.
  • Achieves 81.2 % on MMMU‑Pro (no tools) and 82.1 % with tools, surpassing GPT‑5.2.

Coding Enhancements

  • Merges GPT‑5.3‑Codex coding strengths with the new reasoning and tool capabilities.
  • /fast mode in Codex delivers up to 1.5× faster token velocity without sacrificing intelligence.
  • Introduced the experimental Playwright (Interactive) skill for visual debugging and automated testing of web/Electron apps.

Benchmark Performance

Benchmark GPT‑5.4 GPT‑5.3‑Codex GPT‑5.2
GDPval (knowledge work) 83.0 % – 70.9 %
SWE‑Bench Pro (coding) 57.7 % – 55.6 %
OSWorld‑Verified (desktop computer use) 75.0 % 74.0 % 47.3 %
WebArena‑Verified (browser use) 67.3 % – 65.4 %
BrowseComp (web search) 82.7 % (standard) / 89.3 % (Pro) 77.3 % 65.8 %
MMMU‑Pro (vision) 81.2 % (no tools) / 82.1 % (with tools) – 79.5 %
Toolathlon (tool use) 54.6 % – 45.7 %
Investment‑banking spreadsheet task (internal) 87.3 % – 68.4 %
Presentation quality (human raters) 68.0 % preferred over GPT‑5.2

All evaluations were run with reasoning effort set to xhigh unless noted otherwise.

Real‑World Capabilities

Professional Knowledge Work

  • Matches or exceeds industry professionals on 83 % of GDPval comparisons across 44 occupations.
  • Reduces factual errors: individual claims are 33 % less likely to be false, and full responses 18 % less likely to contain any error versus GPT‑5.2.
  • Generates higher‑quality spreadsheets, presentations, and documents, with human raters preferring GPT‑5.4 outputs for aesthetics and functional variety.

Computer‑Use Agents

  • Achieves state‑of‑the‑art success rates on desktop navigation (75 % vs. 47 % for GPT‑5.2) and on web interaction benchmarks (up to 92.8 % on Online‑Mind2Web).
  • Supports coordinated multi‑tool workflows with reduced latency thanks to tool‑search and parallel tool yields.
  • Demonstrated 95 % first‑attempt success on property‑tax portal tasks, completing sessions ~3× faster while using ~70 % fewer tokens.

Coding & Development

  • Outperforms GPT‑5.3‑Codex on SWE‑Bench Pro while offering lower latency.
  • The Playwright (Interactive) skill enables visual debugging and automated testing during code generation, illustrated by a full theme‑park simulation built from a single prompt.

Safety and Governance

  • Classified as High cyber capability under OpenAI’s Preparedness Framework, with expanded monitoring, trusted access controls, and asynchronous blocking for high‑risk requests on Zero Data Retention (ZDR) surfaces.
  • Introduced the open‑source CoT controllability evaluation, showing GPT‑5.4’s low ability to hide chain‑of‑thought reasoning, reinforcing the effectiveness of CoT monitoring.
  • Ongoing refinement may cause occasional false‑positive refusals for ZDR customers as classifiers improve.

Availability and Pricing

  • ChatGPT – GPT‑5.4 Thinking is live for Plus, Team, and Pro users; GPT‑5.4 Pro is available for Pro and Enterprise plans. Legacy GPT‑5.2 Thinking remains until June 5 2026.
  • API – gpt-5.4 (standard) and gpt-5.4-pro (maximum performance) are live. Pricing per M tokens: $2.50 input, $0.25 cached input, $15 output (standard); $30 input, $180 output for the Pro variant.
  • Context window – Default 272 K tokens; experimental 1 M‑token window can be enabled via model_context_window and model_auto_compact_token_limit, billed at 2× the normal rate for usage beyond 272 K.

Implications for Developers and Enterprises

  • Higher productivity – The model’s planning, longer context, and efficient tool use reduce the number of interaction turns needed for complex tasks.
  • Cost efficiency – Despite higher per‑token rates, overall token consumption drops because GPT‑5.4 solves problems with fewer tokens.
  • Broader tool ecosystems – Tool search enables agents to operate with thousands of possible tools without incurring prohibitive token overhead.
  • Enhanced reliability – Lower hallucination rates and stronger safety controls make GPT‑5.4 suitable for high‑stakes professional domains such as legal analysis, financial modeling, and medical scheduling.

Conclusion

OpenAI’s GPT‑5.4 represents a significant step forward in frontier AI, combining state‑of‑the‑art reasoning, coding, vision, and native computer‑use into a single model. Its performance gains across professional, coding, and computer‑use benchmarks, together with improved token efficiency and safety features, make it a compelling choice for enterprises and developers seeking high‑quality, cost‑effective AI assistance for complex, real‑world workflows.

Sources

Sources