The AI updates aren't slowing down. Actually two days after OpenAI launched a brand new underlying AI mannequin for ChatGPT referred to as GPT-5.3 Prompt, the corporate has unveiled one other, much more large improve: GPT-5.4.
Truly, GPT-5.4 is available in two varieties: GPT-5.4 Considering and GPT-5.4 Professional, the latter designed for probably the most complicated duties.
Each will likely be out there in OpenAI's paid software programming interface (API) and Codex software program growth software, whereas GPT-5.4 Considering will likely be out there to all paid subscribers of ChatGPT (Plus, the $20-per-month plan, and up) and Professional will likely be reserved for ChatGPT Professional ($200 month-to-month) and Enterprise plan customers.
ChatGPT Free customers may also get a style of GPT-5.4, however solely when their queries are auto-routed to the mannequin, based on an OpenAI spokesperson.
The massive headlines on this launch are effectivity, with OpenAI reporting that GPT-5.4 makes use of far fewer tokens (47% fewer on some duties) than its predecessors, and, arguably much more impressively, a brand new "native" Pc Use mode out there via the API and its Codex that lets GPT-5.4 navigate a customers' pc like a human and work throughout functions.
The corporate can be releasing a new suite of ChatGPT integrations permitting GPT-5.4 to be plugged straight into customers' Microsoft Excel and Google Sheets spreadsheets and cells, enabling granular evaluation and automatic process completion that ought to pace up work throughout the enterprise, however might make fears of white collar layoffs much more pronounced on the heels of related choices from Anthropic's Claude and its new Cowork software.
OpenAI says GPT-5.4 helps as much as 1 million tokens of context within the API and Codex, enabling brokers to plan, execute, and confirm duties throughout lengthy horizons— nevertheless, it fees double the associated fee per 1 million tokens as soon as the enter exceeds 272,000 tokens.
Native pc use: a step towards autonomous workflows
Probably the most consequential functionality OpenAI highlights is that GPT-5.4 is its first general-purpose mannequin launched with native, state-of-the-art computer-use capabilities in Codex and the API, enabling brokers to function computer systems and perform multi-step workflows throughout functions.
OpenAI says the mannequin can each write code to function computer systems by way of libraries like Playwright and challenge mouse and keyboard instructions in response to screenshots. OpenAI additionally claims a leap in agentic internet looking.
Benchmark outcomes are offered as proof that this isn’t merely a UI wrapper.
On BrowseComp, which measures how properly AI brokers can persistently browse the online to seek out hard-to-locate info, OpenAI experiences GPT-5.4 bettering by 17% absolute over GPT-5.2, and GPT-5.4 Professional reaching 89.3%, described as a brand new state-of-the-art.
On OSWorld-Verified, which measures desktop navigation utilizing screenshots plus keyboard and mouse actions, OpenAI experiences GPT-5.4 at 75.0% success, in comparison with 47.3% for GPT-5.2, and notes reported human efficiency at 72.4%.
On WebArena-Verified, GPT-5.4 reaches 67.3% success utilizing each DOM- and screenshot-driven interplay, in comparison with 65.4% for GPT-5.2. On On-line-Mind2Web, OpenAI experiences 92.8% success utilizing screenshot-based observations alone.
OpenAI additionally hyperlinks pc use to enhancements in imaginative and prescient and doc dealing with. On MMMU-Professional, GPT-5.4 reaches 81.2% success with out instrument use, in contrast with 79.5% for GPT-5.2, and OpenAI says it achieves that consequence utilizing a fraction of the “considering tokens.”
On OmniDocBench, GPT-5.4’s common error is reported at 0.109, improved from 0.140 for GPT-5.2. The put up additionally describes expanded help for high-fidelity picture inputs, together with an “authentic” element stage as much as 10.24M pixels.
OpenAI positions GPT-5.4 as constructed for longer, multi-step workflows—work that more and more appears to be like like an agent maintaining state throughout many actions relatively than a chatbot responding as soon as.
Software search and improved instrument orchestration
As instrument ecosystems get bigger, OpenAI argues that the naive method—dumping each instrument definition into the immediate—creates a tax paid on each request: price, latency, and context air pollution.
GPT-5.4 introduces instrument search within the API as a structural repair. As a substitute of receiving all instrument definitions upfront, the mannequin receives a light-weight record of instruments plus a search functionality, and it retrieves full instrument definitions solely once they’re truly wanted.
OpenAI describes the effectivity win with a concrete comparability: on 250 duties from Scale’s MCP Atlas benchmark, working with 36 MCP servers enabled, the tool-search configuration lowered complete token utilization by 47% whereas reaching the identical accuracy as a configuration that uncovered all MCP features straight in context.
That 47% determine is particularly concerning the tool-search setup in that analysis—not a blanket declare that GPT-5.4 makes use of 47% fewer tokens for each type of process.
Enhancements for builders and coding workflows
OpenAI’s coding pitch is that GPT-5.4 combines the coding strengths of GPT-5.3-Codex with stronger instrument and computer-use capabilities that matter when duties aren’t single-shot.
GPT-5.4 matches or outperforms GPT-5.3-Codex on SWE-Bench Professional whereas being decrease latency throughout reasoning efforts.
Codex additionally will get workflow-level knobs. OpenAI says /quick mode delivers as much as 1.5× sooner efficiency throughout supported fashions, together with GPT-5.4, describing it as the identical mannequin and intelligence “simply sooner.”
And it describes releasing an experimental Codex ability, “Playwright (Interactive)”, meant to display how coding and pc use can work in tandem—visually debugging internet and Electron apps and testing an app because it’s being constructed.
OpenAI for Microsoft Excel and Google Sheets
Alongside GPT-5.4, OpenAI is saying a collection of safe AI merchandise in ChatGPT constructed for enterprises and monetary establishments, powered by GPT-5.4 for superior monetary reasoning and Excel-based modeling.
The centerpiece is ChatGPT for Excel and Google Sheets (beta), which OpenAI describes as ChatGPT embedded straight in spreadsheets to construct, analyze, and replace complicated monetary fashions utilizing the formulation and constructions groups already depend on.
The suite additionally consists of new ChatGPT app integrations supposed to unify market, firm, and inside knowledge right into a single workflow, naming FactSet, MSCI, Third Bridge, and Moody’s.
And it introduces reusable “Expertise” for recurring finance work similar to earnings previews, comparables evaluation, DCF evaluation, and funding memo drafting.
OpenAI anchors the finance push with an inside benchmark declare: mannequin efficiency elevated from 43.7% with GPT-5 to 88.0% with GPT-5.4 Considering on an OpenAI inside funding banking benchmark.
Measuring AI efficiency towards skilled work
OpenAI leans on benchmarks supposed to resemble actual workplace deliverables, not simply puzzle-solving. On GDPval, an analysis spanning “well-specified information work” throughout 44 occupations, OpenAI experiences that GPT-5.4 matches or exceeds business professionals in 83.0% of comparisons, in comparison with 71.0% for GPT-5.2.
The corporate additionally highlights particular enhancements within the sorts of artifacts that have a tendency to show mannequin weaknesses: structured tables, formulation, narrative coherence, and design high quality.
In an inside benchmark of spreadsheet modeling duties modeled after what a junior funding banking analyst may do, GPT-5.4 reaches a imply rating of 87.5%, in comparison with 68.4% for GPT-5.2.
And on a set of presentation analysis prompts, OpenAI says human raters most popular GPT-5.4’s displays 68.0% of the time over GPT-5.2’s, citing stronger aesthetics, higher visible selection, and simpler use of picture era.
Bettering reliability and lowering hallucinations
OpenAI describes GPT-5.4 as its most factual mannequin but and connects that declare to a sensible dataset: de-identified prompts the place customers beforehand flagged factual errors. On that set, OpenAI experiences GPT-5.4’s particular person claims are 33% much less more likely to be false and its full responses are 18% much less more likely to include any errors in comparison with GPT-5.2.
In statements offered to VentureBeat from OpenAI and attributed early GPT-5.4 testers, Daniel Swiecki of Walleye Capital says that on inside finance and Excel evaluations, GPT-5.4 improved accuracy by 30 proportion factors, which he hyperlinks to expanded automation for mannequin updates and state of affairs evaluation.
Brendan Foody, CEO of Mercor, calls GPT-5.4 the perfect mannequin the corporate has tried and says it’s now high of Mercor’s APEX-Brokers benchmark for skilled providers work, emphasizing long-horizon deliverables like slide decks, monetary fashions, and authorized evaluation.
Pricing and availability
Within the API, OpenAI says GPT-5.4 Considering is out there as gpt-5.4 and GPT-5.4 Professional as gpt-5.4-pro. Pricing is as follows:
GPT-5.4: $2.50 / 1M enter tokens; $15 / 1M output tokens
GPT-5.4 Professional: $30 / 1M enter tokens; $180 / 1M output tokens
Batch + Flex: half-rate; Precedence processing: 2× charge
This makes GPT-5.4 among the many costlier fashions to run over API in comparison with your complete discipline, as seen within the desk beneath.
Mannequin | Enter | Output | Complete Price | Supply |
Qwen 3 Turbo | $0.05 | $0.20 | $0.25 | |
Qwen3.5-Flash | $0.10 | $0.40 | $0.50 | |
deepseek-chat (V3.2-Exp) | $0.28 | $0.42 | $0.70 | |
deepseek-reasoner (V3.2-Exp) | $0.28 | $0.42 | $0.70 | |
Grok 4.1 Quick (reasoning) | $0.20 | $0.50 | $0.70 | |
Grok 4.1 Quick (non-reasoning) | $0.20 | $0.50 | $0.70 | |
MiniMax M2.5 | $0.15 | $1.20 | $1.35 | |
Gemini 3.1 Flash-Lite | $0.25 | $1.50 | $1.75 | |
MiniMax M2.5-Lightning | $0.30 | $2.40 | $2.70 | |
Gemini 3 Flash Preview | $0.50 | $3.00 | $3.50 | |
Kimi-k2.5 | $0.60 | $3.00 | $3.60 | |
GLM-5 | $1.00 | $3.20 | $4.20 | |
ERNIE 5.0 | $0.85 | $3.40 | $4.25 | |
Claude Haiku 4.5 | $1.00 | $5.00 | $6.00 | |
Qwen3-Max (2026-01-23) | $1.20 | $6.00 | $7.20 | |
Gemini 3 Professional (≤200K) | $2.00 | $12.00 | $14.00 | |
GPT-5.2 | $1.75 | $14.00 | $15.75 | |
Claude Sonnet 4.6 | $3.00 | $15.00 | $18.00 | |
GPT-5.4 | $2.50 | $15.00 | $17.50 | |
Gemini 3 Professional (>200K) | $4.00 | $18.00 | $22.00 | |
Claude Opus 4.6 | $5.00 | $25.00 | $30.00 | |
GPT-5.2 Professional | $21.00 | $168.00 | $189.00 | |
GPT-5.4 Professional | $30.00 | $180.00 | $210.00 |
One other essential observe: with GPT-5.4, requests that exceed 272,000 enter tokens are billed at 2X the conventional charge, reflecting the flexibility to ship prompts bigger than earlier fashions supported.
In Codex, compaction defaults to 272k tokens, and the upper long-context pricing applies solely when the enter exceeds 272k—that means builders can maintain sending prompts at or below that measurement with out triggering the upper charge, however can choose into bigger prompts by elevating the compaction restrict, with solely these bigger requests billed in a different way.
An OpenAI spokesperson stated that within the API the utmost output is 128,000 tokens, the identical as earlier fashions.
Lastly, on why GPT-5.4 is priced larger at baseline, the spokesperson attributed it to a few elements: larger functionality on complicated duties (together with coding, pc use, deep analysis, superior doc era, and gear use), main analysis enhancements from OpenAI’s roadmap, and extra environment friendly reasoning that makes use of fewer reasoning tokens for comparable duties—including that OpenAI believes GPT-5.4 stays beneath comparable frontier fashions on pricing even with the rise.
The broader shift
Throughout the discharge and the follow-up clarifications, GPT-5.4 is positioned as a mannequin meant to maneuver past “reply era” and into sustained skilled workflows—ones that require instrument orchestration, pc interplay, lengthy context, and outputs that appear to be the artifacts folks truly use at work.
OpenAI’s emphasis on token effectivity, instrument search, native pc use, and lowered user-flagged factual errors all level in the identical route: making agentic techniques extra viable in manufacturing by reducing the price of retries—whether or not that retry is a human re-prompting, an agent calling one other instrument, or a workflow re-running as a result of the primary cross didn’t stick.
