AI updates are not delayed. Literally two days after OpenAI announced a new underlying AI model for ChatGPT called GPT-5.3 Instant, the company announced another even bigger upgrade: GPT-5.4.
There are actually two varieties of GPT-5.4. GPT-5.4 Thinking and GPT-5.4 ProThe latter is designed for the most complex tasks.
Both will be available in OpenAI’s paid application programming interfaces (APIs) and Codex software development applications, with GPT-5.4 Thinking available to all paid subscribers of ChatGPT (Plus, plans priced at $20 per month or more), and Pro reserved for ChatGPT Pro ($200 per month) and Enterprise plan users.
An OpenAI spokesperson said ChatGPT Free users can also experience GPT-5.4, but only if queries are automatically routed to the model.
The big headline of this release is efficiency, with OpenAI reporting that GPT-5.4 uses far fewer tokens (47% less for some tasks) than previous versions, and perhaps even more impressively, New “native” computer usage modes available through APIs and their codexes This allows GPT-5.4 to interact with a user’s computer like a human and work across multiple applications.
The company also New suite of ChatGPT integrations allows you to connect GPT-5.4 directly to your Microsoft Excel and Google Sheets The use of spreadsheets and cells should speed up work across companies by enabling granular analysis and automated task completion, but the impact of similar services from Anthropic’s Claude and its new Cowork application could further raise concerns about white-collar layoffs.
OpenAI says GPT-5.4 will support up to 1 million context tokens in the API and Codex, allowing agents to plan, execute, and validate tasks over time, but inputs exceeding 272,000 tokens will be charged twice the cost per million tokens.
Using native computers: A step towards autonomous workflows
The most important feature highlighted by OpenAI is that GPT-5.4 is the first general-purpose model released with native, cutting-edge computer usage capabilities in Codex and APIs, allowing agents to interact with computers and execute multi-step workflows across applications.
OpenAI says the model can also write code to interact with the computer through libraries like Playwright, or issue mouse and keyboard commands in response to screenshots. OpenAI also claims breakthroughs in web browsing with agents.
Benchmark results show that this is more than just a UI wrapper.
In BrowseComp, which measures whether an AI agent can persistently browse the web and find hard-to-find information, OpenAI reports that GPT-5.4 has an absolute 17% improvement over GPT-5.2, and GPT-5.4 Pro reaches 89.3%, making it a new state-of-the-art technology.
At OSWorld-Verified, which uses screenshots and keyboard and mouse interactions to measure desktop navigation, OpenAI reports a success rate of 75.0% for GPT-5.4 compared to 47.3% for GPT-5.2, and the note reports human performance at 72.4%.
On WebArena-Verified, GPT-5.4 reaches a success rate of 67.3% using both DOM and screenshot-driven interactions, compared to 65.4% for GPT-5.2. On Online-Mind2Web, OpenAI reports 92.8% success using only screenshot-based observations.
OpenAI also links the use of computers to improved vision and word processing. MMMU-Pro says GPT-5.4 reaches an 81.2% success rate without tools, compared to 79.5% for GPT-5.2, and OpenAI uses some of its “thought tokens” to achieve that result.
OmniDocBench reports an average error of 0.109 for GPT-5.4, an improvement from 0.140 for GPT-5.2. This post also describes expanded support for high-fidelity image inputs, including an “original” level of detail up to 10.24 million pixels.
OpenAI positions GPT-5.4 as built for longer, multi-step workflows. This workflow starts to look more like an agent that maintains state across many actions, rather than a chatbot responding once.
Tool search and improved tool orchestration
As the tool ecosystem grows, OpenAI argues that the simple approach of dumping all tool definitions into a prompt incurs taxes on every request, including cost, latency, and context pollution.
GPT-5.4 introduces tool search to the API as a structural fix. Instead of receiving all tool definitions upfront, the model receives a lightweight list of tools and a search function, retrieving the full tool definition only when it is actually needed.
OpenAI uses specific comparisons to explain efficiency gains. When running 250 tasks from Scale’s MCP Atlas benchmark with 36 MCP servers enabled, the tool search configuration reduced total token usage by 47% and achieved the same accuracy as a configuration that exposed all MCP functionality directly in context.
This 47% number specifically refers to the tool search settings in that evaluation, and is not a blanket claim that GPT-5.4 uses 47% fewer tokens for all kinds of tasks.
Improved developer and coding workflows
OpenAI’s coding pitch is that GPT-5.4 combines the coding strength of GPT-5.3-Codex with more powerful tools and computational capabilities that are important when tasks are more than one-off.
GPT-5.4 matches or exceeds GPT-5.3-Codex on SWE-Bench Pro, resulting in lower latency for the entire inference task.
Codex also adds workflow-level knobs. OpenAI says /fast mode delivers up to 1.5x faster performance across supported models, including GPT-5.4, with the same models and intelligence “just faster.”
We also discuss the release of the experimental Codex skill Playwright (Interactive). It aims to demonstrate how coding and computer usage work together, allowing you to visually debug web and Electron apps and test the apps you’re building.
OpenAI for Microsoft Excel and Google Sheets
In parallel with GPT-5.4, OpenAI is announcing ChatGPT’s suite of secure AI products built for enterprises and financial institutions, leveraging GPT-5.4 to enable advanced financial reasoning and Excel-based modeling.
The highlight is ChatGPT for Excel and Google Sheets (beta), OpenAI describes it as ChatGPT embedded directly into spreadsheets to build, analyze, and update complex financial models using the formulas and structures your team already uses.
The suite also includes a new ChatGPT app integration aimed at integrating market, corporate, and internal data into a single workflow. FactSet, MSCI, Third Bridge, Moody’s.
It also introduces reusable “skills” for routine financial tasks such as earnings previews, comparative analysis, DCF analysis, and investment memo creation.
OpenAI strengthens its financial drive with internal benchmark claims. The model performance improved from 43.7% on GPT-5 to 88.0% on GPT-5.4. Consider OpenAI’s internal investment banking benchmark.
Measuring AI performance for professional work
OpenAI is based on benchmarks that aim to resemble real office artifacts, not just puzzle solving. For GDPval, an assessment across “well-defined knowledge work” across 44 occupations, OpenAI reports that GPT-5.4 matches or beats industry experts in 83.0% of comparisons, compared to 71.0% for GPT-5.2.
The company is also focusing on specific improvements in the types of artifacts that tend to expose model weaknesses, such as structured tables, formulas, narrative consistency, and design quality.
In an internal benchmark of spreadsheet modeling tasks that junior investment banking analysts might perform, the average score for GPT-5.4 reached 87.5%, while the average score for GPT-5.2 was 68.4%.
OpenAI also states that for a set of presentation evaluation prompts, human raters preferred GPT-5.4 presentations over GPT-5.2 68.0% of the time, citing stronger aesthetics, visual variety, and more effective use of image generation.
Improved reliability and reduced hallucinations
OpenAI describes GPT-5.4 as the most fact-based model to date and connects that claim to an actionable dataset: anonymized prompts where users have previously flagged factual errors. In that set, OpenAI reports GPT-5.4 personal claim 33% less likely to be false; complete answer 18% less likely to contain errors compared to GPT-5.2.
In a statement provided to VentureBeat by OpenAI and early GPT-5.4 testers, Walleye Capital’s Daniel Swiecki said GPT-5.4 has improved accuracy by 30 percent in internal finance and Excel evaluations, leading to greater automation of model updates and scenario analysis.
Mercor CEO Brendan Foody said GPT-5.4 was the best model the company had tried, and that GPT-5.4 now tops Mercor’s APEX-Agents benchmark in professional services work, highlighting long-term deliverables such as slide decks, financial models, and legal analysis.
Price and availability
In the API, OpenAI states that GPT-5.4 Thinking is available as follows: gpt-5.4 and as GPT-5.4 Pro gpt-5.4-pro. Prices are as follows:
-
GPT-5.4: $2.50 / 1 million input tokens; $15 / 1 million output tokens
-
GPT-5.4 Pro: $30 / 1 million input tokens; $180 / 1 million output tokens
-
Batch + Flex: half. Priority processing: double rate
This makes GPT-5.4 one of the more expensive models to run on the API compared to the entire field, as shown in the table below.
|
model |
input |
output |
total cost |
sauce |
|
Quen 3 Turbo |
$0.05 |
$0.20 |
$0.25 |
alibaba cloud |
|
Qwen3.5-Flash |
$0.10 |
$0.40 |
$0.50 |
alibaba cloud |
|
Deepseek Chat (V3.2-Exp) |
$0.28 |
$0.42 |
$0.70 |
deep seek |
|
Deep Seek Reasoner (V3.2-Exp) |
$0.28 |
$0.42 |
$0.70 |
deep seek |
|
Grok 4.1 Fast (Inference) |
$0.20 |
$0.50 |
$0.70 |
xAI |
|
Grok 4.1 fast (non-inferential) |
$0.20 |
$0.50 |
$0.70 |
xAI |
|
Minimax M2.5 |
$0.15 |
$1.20 |
$1.35 |
mini max |
|
gemini 3.1 flashlight |
$0.25 |
$1.50 |
$1.75 |
|
|
MiniMax M2.5-Lightning |
$0.30 |
$2.40 |
$2.70 |
mini max |
|
Gemini 3 Flash Preview |
$0.50 |
$3.00 |
$3.50 |
|
|
Kimi-k2.5 |
$0.60 |
$3.00 |
$3.60 |
moonshot |
|
GLM-5 |
$1.00 |
$3.20 |
$4.20 |
Zai |
|
Ernie 5.0 |
$0.85 |
$3.40 |
$4.25 |
Baidu |
|
Claude Haiku 4.5 |
$1.00 |
$5.00 |
$6.00 |
human |
|
Qwen3-Max (2026-01-23) |
$1.20 |
$6.00 |
$7.20 |
alibaba cloud |
|
Gemini 3 Pro (≤200K) |
$2.00 |
$12.00 |
$14.00 |
|
|
GPT-5.2 |
$1.75 |
$14.00 |
$15.75 |
OpenAI |
|
Claude Sonnet 4.6 |
$3.00 |
$15.00 |
$18.00 |
human |
|
GPT-5.4 |
$2.50 |
$15.00 |
$17.50 |
OpenAI |
|
Gemini 3 Pro (>200K) |
$4.00 |
$18.00 |
$22.00 |
|
|
Claude Op. 4.6 |
$5.00 |
$25.00 |
$30.00 |
human |
|
GPT-5.2 Pro |
$21.00 |
$168.00 |
$189.00 |
OpenAI |
|
GPT-5.4 Pro |
$30.00 |
$180.00 |
$210.00 |
OpenAI |
Another important note: In GPT-5.4, requests with more than 272,000 input tokens are charged twice as much as normal, reflecting the ability to send larger prompts than previous supported models.
In Codex, compaction defaults to 272k tokens, and higher long context fees apply only if the input exceeds 272k. This means that developers can continue to send prompts of that size or smaller without triggering a higher rate, but they can choose larger prompts by increasing the compaction limit, and only larger requests will be charged separately.
According to an OpenAI spokesperson, the API’s maximum output is 128,000 tokens, the same as previous models.
Finally, as to why GPT-5.4’s baseline price is so high, the spokesperson points out its increased ability to perform complex tasks (including coding, computer usage, deep research, advanced document generation, and tool usage), and the large amount of research that OpenAI’s roadmap allows for. This is due to three factors: broader improvements, more efficient inference using fewer inference tokens for comparable tasks, and added that OpenAI believes that even if the price of GPT-5.4 increases, it will remain below comparable Frontier models.
broader shift
Through its release and subsequent commentary, GPT-5.4 has been positioned as a model for moving beyond “answer generation” to sustained professional workflows that require orchestration of tools, interaction with computers, extended context, and outputs that resemble artifacts that people actually use in their work.
OpenAI’s emphasis on token efficiency, tool discovery, native computer usage, and reducing user-reported factual errors all point in the same direction. This means increasing the viability of your agent system in production by reducing the cost of retries. Whether that retry is a human re-prompting, an agent invoking another tool, or re-running a workflow because the first pass didn’t stick.
