GPT-5.4, GPT-5.4 Mini & Nano: The Complete GuideOpenAI's Most Powerful Model Family — March 2026

GPT-5.4 is the latest generation of OpenAI's flagship generative AI models, succeeding GPT-5.2 (released December 2025) and incorporating the agentic coding capabilities first introduced in GPT-5.3-Codex (February 2026). It represents a convergence of OpenAI's previously separate model lines — the reasoning-focused o-series and the general-purpose GPT series — into a unified architecture that scales from lightweight classification to frontier-grade professional work.

The model family comprises four distinct variants, each targeting a different segment of the cost-performance spectrum. GPT-5.4 Standard and GPT-5.4 Pro handle the most demanding tasks with a 1 million token context window. GPT-5.4 Mini and Nano serve high-volume, latency-sensitive applications at substantially lower cost. Together, they give developers and enterprises the flexibility to route different tasks to the most appropriate model — a pattern that has become standard practice in production AI systems.

What sets GPT-5.4 apart from its predecessors is not merely incremental improvement. Three genuinely new capabilities — native computer use, tool search, and the Thinking variant's visible reasoning — represent architectural shifts in how large language models interact with the world. These are not fine-tuned additions bolted onto an existing model; they are capabilities baked into the model's training from the ground up.

GPT-5.4 Standard is the general-purpose workhorse — the model most developers and ChatGPT users will interact with daily. GPT-5.4 Pro is the maximum-compute variant reserved for the hardest professional tasks, available to Pro and Enterprise plan subscribers in ChatGPT and via the API.

Both share the same architecture and core capabilities, but Pro allocates significantly more compute per request. OpenAI's own internal investment banking modelling benchmark illustrates the gap: GPT-5.4 Pro achieves 87.3% accuracy compared to GPT-5.2's 68.4% — a substantial improvement on tasks requiring sustained multi-step quantitative reasoning.

GPT-5.4 Thinking is a reasoning variant that makes the model's thought process visible to users. In ChatGPT, it provides an upfront plan of its thinking, allowing users to see and adjust the model's approach mid-response. This is the evolution of OpenAI's earlier o-series reasoning models (o1, o3, o3-pro, and o4-mini), which were retired from ChatGPT on 13 February 2026.

Rather than maintaining a separate reasoning model line, OpenAI has folded reasoning capabilities directly into the GPT-5.x architecture as a “Thinking” tier. This simplifies the model selection process for developers — instead of choosing between a chat model and a reasoning model, you now choose a thinking depth within a single model family.

For enterprise applications, visible reasoning is particularly valuable in compliance-sensitive domains. When an AI system makes a recommendation — whether it's shortlisting a candidate, scoring a resume, or flagging a risk — auditors and compliance teams can inspect the model's reasoning chain to verify that decisions are based on appropriate criteria and are not influenced by protected characteristics.

Released on 17 March 2026, GPT-5.4 Mini is designed for applications where speed and cost matter as much as intelligence. It runs more than 2x faster than GPT-5 Mini (the previous small model) while delivering significantly better performance across coding, reasoning, multimodal understanding, and tool use.

At $0.75 per million input tokens and $4.50 per million output tokens, Mini occupies a sweet spot between Nano's bare-minimum pricing and Standard's frontier capability. Its 400,000 token context window is generous for a small model, and it has native computer use capabilities — making it viable for agent subprocesses that need to interact with web interfaces or desktop applications.

GPT-5.4 Nano is the smallest and most cost-effective model in the GPT-5.4 family. At $0.20 per million input tokens and $1.25 per million output tokens, it is designed for workloads where speed and cost are the primary considerations — classification, data extraction, ranking, routing, and simple coding tasks.

Unlike Mini, Nano is an API-only model. It is not available through the ChatGPT consumer application. This positions it squarely as a developer tool — the model you reach for when building production pipelines that process millions of items and where every fraction of a cent per request matters.

GPT-5.4 enters a fiercely competitive landscape. Anthropic's Claude Opus 4.6 (released February 2026) leads on software engineering benchmarks. Google DeepMind's Gemini 3.1 Pro (also February 2026) offers the best price-to-performance ratio and full multimodal support. Understanding where each model excels is essential for making informed architecture decisions.

GPT-5.4's combination of massive context, computer use, and improved accuracy makes it particularly well-suited for enterprise recruitment workflows. The ability to process entire candidate portfolios in a single context window eliminates the need for complex chunking strategies, while native computer use enables direct interaction with legacy ATS platforms and job boards.

AI-powered recruitment platforms are already leveraging multi-model architectures to optimise cost and performance. For example, FluxHire.AIuses frontier models for complex candidate analysis and matching, while routing high-volume tasks like resume parsing and initial screening to more cost-effective variants. The GPT-5.4 model family's four-tier pricing structure is tailor-made for this kind of intelligent routing.

Australian enterprises evaluating GPT-5.4 should consider several factors specific to the local regulatory and business environment.

Privacy Act 1988 Compliance:Any AI system processing personal information of Australian individuals must comply with the Australian Privacy Principles (APPs), particularly APP 11 (security of personal information) and APP 8 (cross-border disclosure). GPT-5.4's API includes data processing agreements, but organisations should ensure their use cases comply with data minimisation principles and have appropriate consent mechanisms in place.

Data Residency:OpenAI's regional data residency option (available at a 10% pricing uplift) allows organisations to ensure API requests are processed within specific geographic regions. Australian enterprises handling sensitive data should evaluate whether this meets their data sovereignty requirements.

Fair Work Act Considerations:When using GPT-5.4 for recruitment decisions (screening, shortlisting, scoring), organisations must ensure that AI-assisted decisions do not discriminate on the basis of protected attributes under the Fair Work Act 2009. The Thinking variant's visible reasoning provides an audit trail that can help demonstrate compliance.

Cost Advantage for Australian SMEs:The Nano model at $0.20/MTok makes frontier AI accessible to smaller Australian recruitment firms that previously could not justify the cost of API-based AI. A recruitment agency processing 10,000 candidate profiles per month would spend approximately $10–$50 on Nano, compared to $125–$750 on Standard.

With GPT-5.2 Thinking classified as a Legacy Model and scheduled for permanent retirement on 5 June 2026, organisations currently running GPT-5.2 in production need to plan their migration. Here are the key considerations.

Alongside GPT-5.4, OpenAI launched ChatGPT for Excel and Google Sheets in beta. This embeds ChatGPT directly within spreadsheet applications, allowing users to invoke GPT-5.4 capabilities without leaving their workflow.

For recruitment teams, this has immediate practical applications: analysing candidate data, generating interview scorecards, creating salary benchmarking reports, and automating repetitive spreadsheet tasks. The integration runs on GPT-5.4 Mini by default, keeping costs manageable for frequent use.

This is significant for the broader adoption of AI in enterprises. Many HR professionals and recruiters spend a substantial portion of their day in spreadsheets. Meeting them where they already work — rather than requiring them to switch to a separate AI tool — reduces friction and accelerates adoption.

The pace of OpenAI's releases — five major model versions in seven months — shows no sign of slowing. GPT-5.4 is not an endpoint but a milestone in a rapid evolution. For enterprises, the strategic imperative is clear: build architectures that can adapt to new models quickly, invest in evaluation frameworks that can benchmark new releases against your specific use cases, and maintain model-agnostic abstractions in your codebase.

The AI recruitment space, in particular, is evolving rapidly. Platforms that integrate frontier models for complex decision-making while leveraging cost-effective models for high-volume processing are best positioned to deliver value. The GPT-5.4 family's four-tier structure makes this kind of intelligent routing more accessible than ever. To explore how AI-powered recruitment automation is transforming the industry, see our guide to OpenAI agents in recruitment automation.