OpenAI's GPT-5.4 mini and nano launch - with near flagship performance at much lower cost

Elyse Betters Picaro / ZDNET

Follow ZDNET: Add america arsenic a preferred source on Google.

ZDNET's cardinal takeaways

• GPT-5.4 mini runs much than doubly arsenic accelerated arsenic GPT-5 mini.
• New models purpose astatine agents, coding, and multi-modal workflows.
• Developers tin premix ample readying models with cheaper subagents.

Over the past fewer weeks, we person seen the procreation of OpenAI's flagship ample connection models iterate from GPT-5.3 to GPT-5.4. Think of the exemplary arsenic the motor that powers AI computation. Each generational leap usually results successful accrued show and accuracy.

Also: OpenAI's caller GPT-5.4 clobbers humans connected pro-level enactment successful tests - by 83%

The existent releases tin beryllium a spot hard to way without a scorecard. On March 5, OpenAI released GPT-5.4 Thinking, a high-performance, in-depth reasoning model. Two days earlier, it released GPT-5.3 (not 5.4) Instant, a exemplary that "makes mundane conversations much consistently adjuvant and fluid," but not needfully much accurate.

This week, OpenAI is releasing the GPT-5.4 mini and GPT-5.4 nano models. These models are designed for fast, efficient, high-volume AI workloads. These are fundamentally the fund connection exemplary offerings.

Smaller models for AI workflows

For galore AI workflows, the astir effectual exemplary is 1 that balances beardown show with accelerated responses and reliable instrumentality use.

According to OpenAI, "These models are built for the kinds of workloads wherever latency straight shapes the merchandise experience: coding assistants that request to consciousness responsive, subagents that rapidly implicit supporting tasks, computer-using systems that seizure and construe screenshots, and multimodal applications that tin crushed implicit images successful real-time."

Also: Nvidia's 'ChatGPT moment' for self-driving cars, and different cardinal AI announcements astatine GTC 2026

The institution said, "In these settings, the champion exemplary is often not the largest 1 -- it's the 1 that tin respond quickly, usage tools reliably, and inactive execute good connected analyzable nonrecreational tasks."

Compared to GPT-5 mini, GPT-5.4 mini improves crossed coding, reasoning, multimodal understanding, and instrumentality use. The exemplary runs much than doubly arsenic accelerated arsenic GPT-5 mini.

GPT-5.4 nano is the smallest and fastest model, aimed astatine classification, extraction, ranking, and simpler coding-support tasks.

Performance improvements

When looking astatine the smaller, little costly models, show is the distinguishing factor. Buyers privation to cognize conscionable however overmuch bang for the subordinate they're getting. To exemplify this performance, OpenAI is showing important benefits implicit models released conscionable months earlier:

• GPT-5.4 mini scores 54.38% connected SWE-bench Pro compared with 45.69% for GPT-5 mini.
• On Terminal-Bench 2.0, GPT-5.4 mini reaches 60.00%, versus 38.20% for GPT-5 mini.
• On GPQA Diamond, GPT-5.4 mini scores 88.01%, approaching GPT-5.4's 93.00%.
• OSWorld-Verified results amusement GPT-5.4 mini astatine 72.13%, importantly higher than GPT-5 mini's 42%.

GPT-5.4 mini approaches GPT-5.4-level walk rates portion delivering faster execution. In different words, the smaller, lighter GPT-5.4 mini exemplary performs astir arsenic good arsenic the afloat GPT-5.4 exemplary connected benchmark tests (the "pass rates") that measurement if the exemplary solves problems correctly.

Also: Why encrypted backups whitethorn neglect successful an AI-driven ransomware era

GPT-5.4 nano splits the difference. For example, it scores 52.39% connected SWE-bench Pro and 46.30% connected Terminal Bench 2.0, not arsenic precocious arsenic GPT-5.4 mini but inactive considerably amended than GPT-5 mini.

Customer investigating highlights benefits

Technology specialist Hebbia builds tools that assistance professionals excavation done tremendous collections of documents utilizing earthy language. Their offerings entreaty to users successful sectors specified arsenic finance, law, and research, wherever the quality to analyse and deduce insights from galore documents astatine erstwhile is peculiarly helpful.

According to Aabhas Sharma, CTO astatine Hebbia: "GPT-5.4 mini delivers beardown end-to-end show for a exemplary successful this class. In our evaluations, it matched oregon exceeded competitory models connected respective output tasks and citation callback astatine a overmuch little cost. It besides achieved higher end-to-end walk rates and stronger root attribution than the larger GPT-5.4 model."

Digital workspace Notion is the darling of internet-based productivity wonks. I'm penning this nonfiction successful my Notion workspace. The exertion provides a location for some structured and unstructured data. You tin besides usage Notion to physique no-code mini applications for accusation management. I usage Notion to way my nonfiction production, interior projects, video plans, improvement projects, and more.

Also: As AI agents spread, 1Password's caller instrumentality tackles a rising information threat

Abhisek Modi, AI engineering pb astatine Notion, said: "GPT-5.4 mini handles focused, well-defined tasks with awesome precision. For editing pages specifically, it matched and often exceeded GPT-5.2 connected handling analyzable formatting astatine a fraction of the compute."

Modi continued: "Until recently, lone the astir costly models could reliably navigate agentic instrumentality calling. Today, smaller models similar GPT-5.4 mini and nano tin easy grip it, which volition fto our users physique Custom Agents connected Notion prime precisely the magnitude of quality they need."

I haven't been super-impressed by Notion's AI. Hopefully, by incorporating these caller models, Notion AI's show volition amended considerably.

Subagents and multimodal tasks

When you commencement to look astatine however agents acceptable into the wide ecosystem, it becomes evident that AI tin beryllium structured to reflector real-world quality operations. For example, you tin harvester a much almighty AI exemplary (like GPT-5.4 Thinking) with faster, cheaper models similar GPT-5.4 mini successful the aforesaid mode you mightiness person a elder technologist managing a squad of inferior engineers.

Also: Nvidia wants to ain your AI information halfway from extremity to end

Agentic systems tin harvester models of antithetic sizes, with larger models readying tasks and smaller models executing subtasks. In this context, GPT-5.4 mini tin grip subagent work, specified arsenic searching codebases, reviewing files, and processing documents.

OpenAI said: "GPT-5.4 mini is besides beardown connected multimodal tasks, peculiarly those related to machine use. The exemplary tin rapidly construe screenshots of dense idiosyncratic interfaces to implicit machine usage tasks with speed."

Availability and pricing

GPT-5.4 mini is disposable successful API, Codex, and ChatGPT versions. For Free and Go tier users, GPT-5.4 mini is accessible via the "Thinking" enactment successful the positive menu. OpenAI said: "For each different users, GPT-5.4 mini is disposable arsenic a complaint bounds fallback for GPT-5.4 Thinking."

Also: I utilized GPT-5.2-Codex to find a enigma bug and hosting nightmare - it was beyond fast

The institution said that for programmers, GPT-5.4 mini is disposable crossed the Codex app, CLI, IDE extension, and web. OpenAI said that the mini exemplary "Uses lone 30% of the GPT-5.4 quota, letting developers rapidly grip simpler coding tasks successful Codex for astir one-third the cost." Additionally, Codex tin besides delegate to GPT-5.4 mini subagents truthful that little reasoning-intensive enactment runs connected the little costly model.

You tin spot however costs comparison erstwhile you look astatine them broadside by side:

• GPT-5.4 mini pricing is $0.75 per cardinal input tokens and $4.50 per cardinal output tokens with a 400k discourse window.
• GPT-5.4 nano is API-only and costs $0.20 per cardinal input tokens and $1.25 per cardinal output tokens.

By comparison, GPT-5.4 is priced astatine $2.50 per cardinal input tokens and $15.00 per cardinal output tokens. That's a lot much expensive. It makes consciousness that if you're trying to support costs down and don't request the other processing power, it's amended to usage the mini and nano models.

What astir you?

Have you experimented with smaller AI models, similar GPT-5.4 mini oregon nano, successful your ain workflows? Do you similar utilizing the largest models available, oregon bash you find faster, cheaper models are often "good enough" for real-time tasks similar coding, papers analysis, oregon cause workflows?

If you physique AI-powered tools, however bash you determine erstwhile to usage a afloat reasoning exemplary versus a lightweight subagent model? Let america cognize what you're seeing successful signifier and remark below.

You tin travel my day-to-day task updates connected societal media. Be definite to subscribe to my play update newsletter, and travel maine connected Twitter/X astatine @DavidGewirtz, connected Facebook astatine Facebook.com/DavidGewirtz, connected Instagram astatine Instagram.com/DavidGewirtz, connected Bluesky astatine @DavidGewirtz.com, and connected YouTube astatine YouTube.com/DavidGewirtzTV.