GPT-5.4: Computer Use, Tool Search, Benchmarks, Pricing

What Is GPT-5.4

GPT-5.4 is OpenAI's new flagship reasoning model, designed for professional work that requires sustained tool use, multi-step execution, and high-quality outputs across documents, code, and software environments. It replaces GPT-5.2 Thinking in ChatGPT and unifies the coding strengths of GPT-5.3-Codex with broader knowledge work and computer-use capabilities.

The naming jump from GPT-5.3-Codex to GPT-5.4 reflects the scope of improvement. OpenAI explicitly states this is their first mainline reasoning model that incorporates frontier coding capabilities while being deployed across all surfaces simultaneously. In the API, it is available as gpt-5.4 with a Pro variant at gpt-5.4-pro.

Knowledge Work Performance

GPT-5.4 sets a new state of the art on GDPval, OpenAI's benchmark that tests agents' abilities to produce well-specified knowledge work across 44 occupations from the top 9 industries contributing to U.S. GDP. Tasks include real work products such as sales presentations, accounting spreadsheets, urgent care schedules, manufacturing diagrams, and short videos.

OpenAI put particular focus on improving spreadsheets, presentations, and documents. On an internal benchmark of spreadsheet modeling tasks a junior investment banking analyst might handle, GPT-5.4 achieved 87.3% compared to 68.4% for GPT-5.2. Human raters preferred GPT-5.4 presentations 68% of the time due to stronger aesthetics, greater visual variety, and more effective use of image generation.

GPT-5.4 is also the most factual model OpenAI has released. On a set of de-identified prompts where users flagged factual errors, individual claims are 33% less likely to be false and full responses are 18% less likely to contain any errors compared to GPT-5.2. For businesses that rely on accurate AI-generated content, this is a meaningful improvement in production reliability.

Computer Use and Vision

GPT-5.4 is the first general-purpose model OpenAI has released with native computer-use capabilities. It can write code to operate computers via libraries like Playwright and issue mouse and keyboard commands in response to screenshots. This dual approach makes it the strongest model currently available for developers building agents that complete real tasks across websites and software systems.

The OSWorld result is particularly notable: GPT-5.4 surpasses human performance at 72.4% on desktop navigation tasks involving screenshots and keyboard/mouse actions. The jump from GPT-5.2's 47.3% to 75.0% represents a generational improvement in agentic desktop capabilities.

Developers can access computer-use capabilities through the updated computer tool in the API. Behavior is steerable via developer messages, meaning agents can be tuned for specific use cases. Developers can also configure custom confirmation policies that specify different levels of risk tolerance for automated actions.

Coding Capabilities

GPT-5.4 combines the coding strengths of GPT-5.3-Codex with its broader knowledge work and computer-use capabilities. It matches or outperforms GPT-5.3-Codex on SWE-Bench Pro while offering lower latency across reasoning effort levels.

In Codex, the /fast mode delivers up to 1.5x faster token velocity with GPT-5.4. It uses the same model and intelligence, just with faster output. Developers can access the same speeds via the API using priority processing.

OpenAI demonstrated this with a theme park simulation game built from a single prompt, using Playwright Interactive for browser playtesting and image generation for isometric assets. The simulation includes tile-based path placement, ride construction, guest pathfinding, and live park metrics. For teams building web applications, this combination of visual debugging and code generation points toward a future where AI agents can iterate on frontend work with minimal human intervention.

Tool Search and Agentic Capabilities

GPT-5.4 introduces tool search in the API, a structural change to how models work with external tools. Previously, all tool definitions were included in the prompt upfront. For systems with many tools, this could add thousands or tens of thousands of tokens to every request, increasing cost, slowing responses, and crowding the context with information the model might never use.

On 250 tasks from Scale's MCP Atlas benchmark with all 36 MCP servers enabled, tool search reduced total token usage by 47% while achieving the same accuracy. For MCP servers that may contain tens of thousands of tokens of tool definitions, the efficiency gains are substantial. This is particularly relevant for enterprise deployments where agents need access to large connector ecosystems.

Agentic Tool Calling

GPT-5.4 achieves higher accuracy in fewer turns on Toolathlon, a benchmark testing how well AI agents use real-world tools and APIs to complete multi-step tasks like reading emails, grading assignments, and recording results in spreadsheets. The BrowseComp improvement is especially striking: a 17-point jump over GPT-5.2, with GPT-5.4 Pro setting a new state of the art at 89.3%.

Steerability and Mid-Response Guidance

In ChatGPT, GPT-5.4 Thinking introduces a preamble feature that outlines the model's planned approach before executing on complex queries. Users can add instructions or adjust direction mid-response, steering the model toward the exact outcome they want without starting over or requiring multiple additional turns.

This feature is available now on chatgpt.com and the Android app, with iOS support coming soon. For professionals using ChatGPT for complex document creation, research synthesis, or multi-step analysis, the ability to steer mid-response means fewer wasted iterations and faster convergence on the desired output.

Pricing and Availability

GPT-5.4 is priced higher per token than GPT-5.2 to reflect its improved capabilities, while its greater token efficiency helps reduce the total number of tokens required for many tasks. Batch and Flex pricing are available at half the standard API rate, while Priority processing is available at twice the standard rate.

The per-token input price increase from $1.75 to $2.50 (a 43% bump) is partially offset by GPT-5.4's improved token efficiency. OpenAI states it uses significantly fewer reasoning tokens than GPT-5.2 to solve equivalent problems. For workloads that were previously expensive due to long reasoning chains, the total cost may be comparable or lower. Developers should benchmark their specific use cases to determine net cost impact.

Full Benchmark Comparison

The table below consolidates GPT-5.4's performance across academic, reasoning, and long-context benchmarks. All evaluations were run with reasoning effort set to xhigh unless noted otherwise.

The abstract reasoning improvements are substantial. ARC-AGI-2 jumps from 52.9% to 73.3% (GPT-5.4 Pro reaches 83.3%), demonstrating that GPT-5.4 represents a genuine reasoning advance, not just a tool-use wrapper around GPT-5.3-Codex. FrontierMath Tier 4, the hardest tier of mathematical reasoning, improves from 18.8% to 27.1%, with Pro reaching 38.0%.

Long Context Performance

GPT-5.4 is the first OpenAI model to support context lengths beyond 256K tokens. On Graphwalks BFS at the 0-128K range, it achieves 93.0% accuracy, comparable to GPT-5.2's 94.0%. At the new 256K-1M range that only GPT-5.4 supports, it achieves 21.4% on Graphwalks BFS and 32.4% on Graphwalks parents, reflecting the challenges of retrieval at extreme context lengths. On the MRCR needle-retrieval benchmark, performance remains strong through 128K (86.0%) and declines at the 512K-1M range (36.6%), which is typical of current-generation models operating at these scales.

Frequently Asked Questions