HomeAI models
OpenAI: GPT-5.4
OpenAI
•
GPT-5.4 (Xhigh)
•
Released 
March 2026

OpenAI
GPT-5.4 (Xhigh)

OpenAI's unified frontier model with native computer use, 1M-token context, and state-of-the-art desktop-agent performance.

Modality:
Text
Image
PDF
model ID
openai/gpt-5.4
Start for free
arrow black icon
View API Docs
arrow white icon

Output Speed *

0.00
tok/s

Intelligence Index *

39
/ 100

Context Window *

1050000
tokens

Input price

15
Anytoken

Output price

90
Anytoken
GPT-5.4: The First General-Purpose Model Built to Operate Computers GPT-5.4 is OpenAI's frontier model that folds the GPT-5.3-Codex coding line back into the mainline, unifying reasoning, coding, and agentic workflows in one system. It is positioned as OpenAI's most capable model for professional, document-heavy work. Its defining feature is native computer use: GPT-5.4 operates desktop applications from screenshots and keyboard/mouse actions, reaching a state-of-the-art 75.0% on OSWorld-Verified. Combined with a 1M-token context window and tool search, it targets long-horizon agents that plan, execute, and verify tasks across software, spreadsheets, and documents with less manual intervention. Access GPT-5.4 via the AnyAPI.ai API and start building computer-use agents.

Performance

Why GPT-5.4 Leads on Desktop-Agent Tasks

GPT-5.4's strongest capability is agentic computer use. On OSWorld-Verified, which measures navigating a desktop from screenshots and keyboard/mouse actions, it reaches 75.0%, up from GPT-5.2's 47.3% and above the 72.4% human baseline. Because it combines multiple low-level operations into single control steps and emits shorter action traces, it completes tasks in fewer interaction rounds. In production this means agents that operate real software, fill forms, and drive browsers with less supervision and lower token overhead. It also matches specialist coding models on SWE-Bench Pro while remaining a general-purpose default.

Benchmarks

GPT-5.4 Benchmarks: Computer Use, Coding, and Agentic Tool Use

GPT-5.4 posts strong independent and official results across agentic tasks. It reaches a state-of-the-art 75.0% on OSWorld-Verified, surpassing the reported 72.4% human baseline for desktop navigation. On coding, it scores 57.7% on SWE-Bench Pro (Public), narrowly ahead of the coding-specialist GPT-5.3-Codex at 56.8%. For browser and tool use, it hits 82.7% on BrowseComp and 54.6% on Toolathlon for real-world multi-step tool calling. Artificial Analysis measures roughly 117 tokens/sec output on OpenAI's endpoint at xhigh reasoning, though time to first token is high given the reasoning workload.

Output Speed

*
0.00
tok/s

Intelligence Index

*
39
/ 100

MMLU *

Broad world knowledge and problem-solving
0
%

GPQA *

PhD-level scientific reasoning across physics, biology, chemistry.
92
%

HLE *

Adherence to multi-step structured instructions.
44
%

LiveCodeBench *

Tool-calling reliability in long agentic loops.
0
%

Technical Specifications

What the model supports

GPT-5.4 accepts text and image input and returns text, with a 1.05M-token context window and up to 128K output tokens. The most production-relevant detail is context pricing: prompts above 272K input tokens are charged at 2x input and 1.5x output for the entire session, so large-context sessions carry materially higher per-session cost. It exposes reasoning effort control (medium, high, xhigh), tool calling, structured outputs via JSON schema, prompt caching, and native computer use through the Responses API.
Verified Specifications — 
GPT-5.4 (Xhigh)
*
Input modalities
Text
Image
PDF
output modalities
Text
Context window
1050000
 tokens
Maximum output tokens
128000
Reasoning
Yes
Knowledge cutoff
March 2026
Pricing (standard)
15
 AnyTokens in
 / 
90
 AnyTokens out

Limitations & Trade-offs

Where GPT-5.4 (Xhigh) falls short

1
High latency on reasoning workloads. Independent measurement from Artificial Analysis shows time to first token above 140 seconds on OpenAI's endpoint at xhigh reasoning effort. This is inherent to a reasoning model doing long-horizon planning, but it makes GPT-5.4 poorly suited to latency-sensitive, interactive use like live chat or autocomplete. For those cases, lower-effort settings, GPT-5.4 mini, or a non-reasoning model are better fits.
2
Expensive large-context sessions. GPT-5.4 supports a 1.05M-token context, but prompts exceeding 272K input tokens are billed at 2x input and 1.5x output for the entire session. Teams doing very large codebase or document-collection analysis should expect a step change in cost above that threshold, and should chunk or retrieve selectively rather than defaulting to full-context prompts when it is not necessary.
3
Premium pricing tier. GPT-5.4 is priced higher per token than its predecessor GPT-5.2 to reflect improved capability, and GPT-5.4 Pro sits far above it. Although greater token efficiency reduces total tokens on many tasks, high-volume generation workloads that do not need frontier reasoning or computer use will be more economical on GPT-5.4 mini, nano, or a competing mid-tier model.
4
Instruction-following inconsistency. Independent reviewers praised GPT-5.4 Thinking for a low tendency to hallucinate but noted it sometimes does not follow user prompts accurately. For workflows that depend on strict adherence to detailed instructions or formatting, validate outputs and use structured outputs and confirmation policies rather than assuming compliance.

Best-Fit Workloads

Where this model earns its place

01

Desktop and browser automation agents

‍
GPT-5.4's native computer use and 75.0% OSWorld-Verified score make it the strongest option for agents that operate real software from screenshots and keyboard/mouse actions. It also leads browser benchmarks (82.7% BrowseComp, 92.8% Online-Mind2Web). Behavior is steerable via developer messages and configurable confirmation policies, so agents can be tuned for specific risk tolerances. Best for automating multi-application workflows; expect high per-task latency.

02
Long-horizon coding agents

‍
With 57.7% on SWE-Bench Pro (Public), GPT-5.4 matches the specialist GPT-5.3-Codex while adding computer use and a 1M-token context that fits large codebases in a single request. Tool search lets it navigate large tool ecosystems efficiently. It suits agents that plan, edit, run, and verify across a repository. For pure terminal-agent tasks, GPT-5.3-Codex still holds a small edge on Terminal-Bench 2.0.

03
Document-heavy knowledge work

‍
GPT-5.4 is tuned for professional deliverables such as spreadsheets, presentations, and legal or financial analysis, where spreadsheet-modeling scores rose sharply over GPT-5.2 and it topped an independent professional-services agent benchmark. The 1M-token context supports synthesizing across long document sets. It fits report generation, financial modeling, and legal review, provided large-context sessions are managed against the higher above-272K pricing tier.

04
Multi-tool orchestration

‍
Tool search lets GPT-5.4 receive a lightweight tool list and fetch full definitions on demand, cutting token usage while operating across large connector ecosystems. It scores 54.6% on Toolathlon for real-world multi-step tool calling. This suits agents integrating many external APIs and services where prompt bloat from full tool schemas would otherwise degrade performance and inflate cost.

Pricing in anytokens via AnyAPI
Input
15
₳
Output
90
₳
Cache write
—
₳
Cache read
1.5
₳
Get API access
arrow black icon
Read Docs

on this page

PerformanceBenchmarksTechnical SpecificationsLimitations & Trade-offsBest-Fit WorkloadsIntegrationFAQ

Integration

Access GPT-5.4 (Xhigh) via AnyAPI.ai

Access GPT-5.4 (Xhigh) through AnyAPI.ai using a unified API built for multi-model AI applications. Integrate GPT-5.4 (Xhigh) without maintaining a separate provider-specific connection, and keep the flexibility to test, switch, or combine models as your application requirements evolve.

01

One API integration

Access GPT-5.4 (Xhigh) and other AI models through the same API workflow instead of maintaining separate integrations for every provider.

02

Easy model switching

Test GPT-5.4 (Xhigh) against alternative models or switch models as your performance, capability, or cost requirements change without rebuilding your application around another provider API.

03

Flexible for production

Use GPT-5.4 (Xhigh) from experimentation through production while keeping your AI stack flexible as workloads, traffic, and model requirements evolve.

04

Multi-model applications

Use GPT-5.4 (Xhigh) for the workloads where it performs best and combine it with other models for tasks that require different capabilities, performance, or efficiency.

Start free
arrow black icon
Read Docs

Frequently Asked Questions

Answers to common questions about integrating and using this AI model via AnyAPI.ai

GPT-5.4 has a 1,050,000-token context window with a maximum output of 128,000 tokens. Prompts exceeding 272K input tokens are priced at 2x input and 1.5x output for the full session, so very large-context requests carry meaningfully higher cost.

GPT-5.4 is OpenAI's first general-purpose model with native, state-of-the-art computer use, scoring 75.0% on OSWorld-Verified and surpassing the reported 72.4% human baseline. It unifies coding, reasoning, and agentic workflows in one model, folding the GPT-5.3-Codex coding line back into the mainline.

Yes. GPT-5.4 scores 57.7% on SWE-Bench Pro (Public), slightly ahead of the coding specialist GPT-5.3-Codex at 56.8%, while adding computer use and a 1M-token context. For narrowly terminal-agent coding, GPT-5.3-Codex still holds a small edge on Terminal-Bench 2.0.

Yes. GPT-5.4 exposes a reasoning-effort control with medium (default), high, and xhigh settings. It supports tool calling with tool search, structured outputs via JSON schema, prompt caching, and native computer use through the OpenAI Responses API.

GPT-5.4 accepts text and image input and produces text output. Its knowledge cutoff is August 31, 2025. Unlike some sibling or consumer models, it is a text-output model, so it does not natively generate images or audio through this API model.

* Benchmark data source: Artificial Analysis artificialanalysis.ai