Responsive coding loops
Use mini for targeted edits, codebase navigation, front-end generation, test failures, review, and debugging where turnaround matters.
Independent tool overview
GPT-5.4 mini & nano at a glance
GPT-5.4 mini and GPT-5.4 nano are OpenAI's March 2026 small reasoning models for fast, high-volume work. Both offer a 400,000-token context window, 128,000 maximum output, image input, function calling, structured outputs, and reasoning effort from none to xhigh. Mini is the stronger coding, computer-use, and subagent option at $0.75 input and $4.50 output per million tokens; nano targets classification, extraction, ranking, and simple subagents at $0.20 and $1.25. Both remain active, though new systems should also benchmark the newer GPT-5.6 family.
Visit the official GPT-5.4 mini & nano site ↗
- Model IDs
- gpt-5.4-mini and gpt-5.4-nano
- Best use
- Mini for coding and tools; nano for simple high-volume processing
- Context window
- 400,000 tokens for both
- Maximum output
- 128,000 tokens for both
- Knowledge cutoff
- August 31, 2025
- Reasoning effort
- none, low, medium, high, or xhigh
- Mini availability
- API, Codex, and limited/fallback ChatGPT access
- Nano availability
- API only
- Reviewed
- August 31, 2026
Overview
What GPT-5.4 mini & nano is
GPT-5.4 mini and nano compress many GPT-5.4 capabilities into faster, cheaper models. OpenAI positions mini for coding, multimodal understanding, computer use, and subagents, while nano is the smallest GPT-5.4 option for classification, extraction, ranking, and other bounded, high-volume tasks.
The API specifications are unusually generous for small models: both have a 400,000-token context window, 128,000-token maximum output, an August 31, 2025 knowledge cutoff, text and image input, text output, streaming, function calling, structured outputs, and a fixed March 17, 2026 snapshot.
Their tool sets differ. Mini supports web and file search, image generation, code interpreter, hosted shell, apply patch, skills, computer use, MCP, and tool search in the Responses API. Nano supports most of the same tools but does not support computer use or tool search, making it better as a bounded worker than a general interface agent.
Availability also differs. Mini is available in the API, Codex, and ChatGPT; in ChatGPT it is exposed to Free and Go users through Thinking and used as a rate-limit fallback for other plans rather than appearing as an ordinary selectable model. Nano is API-only.
These models are active but no longer the newest cost-optimized OpenAI generation. The current comparison page lists GPT-5.6 Luna at $0.20 input and $1.20 output per million tokens with a 1.05-million-token context and newer cutoff. Developers should benchmark actual task quality, tool reliability, latency, and total cost before defaulting to GPT-5.4 nano or mini.
Use cases
Who GPT-5.4 mini & nano is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Parallel subagents
Use mini or nano for narrow, independently checked supporting tasks while a stronger model handles planning, coordination, and final judgment.
Computer-use assistants
Use mini for screenshot interpretation and bounded interface actions when the workflow has least-privilege access and confirmation gates.
Classification and extraction
Use nano for high-volume tagging, ranking, routing, field extraction, and short transformations with a strict schema and measurable accuracy target.
Document triage
Use either model to sort, summarize, or flag large collections before an expert or stronger model reviews the consequential subset.
Cost-controlled prototypes
Build and evaluate tool workflows at lower token cost before promoting only the difficult cases to a larger model.
Capabilities
Core GPT-5.4 mini & nano features
Five reasoning levels
Both models support none, low, medium, high, and xhigh reasoning effort; none is the default.
Large context
A 400K-token context can hold substantial documents, repositories, or agent state, though long-context accuracy still degrades on difficult retrieval.
Large output ceiling
Up to 128K output tokens supports substantial code, structured data, and document generation, subject to cost and reliability controls.
Vision input
Both accept images alongside text for screenshots, charts, diagrams, document pages, and visual classification.
Function calling
Both can choose developer-defined tools and populate arguments in multi-step workflows.
Structured outputs
Both support schema-constrained responses for classification, extraction, routing, and machine-consumed results.
Mini computer use
Mini can use the Responses API computer-use tool and interpret dense interface screenshots; nano does not support computer use.
Search and files
Both support web search, file search, image generation, code interpreter, hosted shell, apply patch, skills, and MCP in the Responses API.
Tool search
Mini can dynamically discover relevant tools through tool search, while nano does not support that feature.
Stable snapshots
Pin gpt-5.4-mini-2026-03-17 or gpt-5.4-nano-2026-03-17 for repeatable evaluations and controlled releases.
Codex efficiency
OpenAI says mini uses 30% of GPT-5.4 quota in Codex and can serve as a delegated worker across app, CLI, IDE, and web.
Process
How the GPT-5.4 mini & nano workflow works
Step 1
Split the task by difficulty
Define which steps are bounded enough for nano, which need mini's stronger coding or computer use, and which require a current flagship or human expert.
Step 2
Create a labeled evaluation set
Include representative inputs, edge cases, adversarial instructions, expected schema, source truth, and acceptable error thresholds for every route.
Step 3
Start with nano for narrow work
Try nano on classification, extraction, ranking, and transformation first. Promote failures or low-confidence cases to mini or a larger model.
Step 4
Use mini for interaction and code
Route code edits, multimodal interpretation, tool discovery, and computer-use steps to mini, but keep final approval outside the model.
Step 5
Set reasoning deliberately
Begin with none or low for simple tasks and increase only when evaluation gains justify added latency and generated-token cost.
Step 6
Constrain every machine result
Use structured outputs, enums, bounds, allowlists, and server-side validation. Reject missing, extra, malformed, or impossible values.
Step 7
Treat content as data, not instructions
Defend against prompt injection in documents, websites, repositories, screenshots, and tool responses. Never let retrieved text expand permissions.
Step 8
Sandbox tools
Use read-only access and isolated environments by default. Require confirmation for communication, money movement, deletion, publishing, credentials, identity, or production changes.
Step 9
Verify outputs
Run tests, reconcile extracted fields, sample classifications, inspect citations, compare screenshots, and independently recalculate consequential numbers.
Step 10
Track route-level economics
Measure input, cached input, reasoning, output, tools, retries, failures, reviewer time, and escalation rate—not only the advertised token price.
Step 11
Benchmark newer small models
Compare against GPT-5.6 Luna and other current candidates on the same task set. Migrate with shadow traffic, canaries, pinned snapshots, and rollback.
Cost
GPT-5.4 mini & nano pricing and free plan
API pricing is per million text tokens. GPT-5.4 mini costs $0.75 input, $0.075 cached input, and $4.50 output; GPT-5.4 nano costs $0.20 input, $0.02 cached input, and $1.25 output. Regional-processing endpoints add a 10% uplift. Tool calls can add separate charges. In Codex, mini uses 30% of GPT-5.4 quota. ChatGPT access to mini is plan- and routing-dependent; nano is not available there.
GPT-5.4 mini API
$0.75 input / $4.50 output per 1M tokens
Stronger small model for coding, computer use, multimodal analysis, tools, and subagents.
- $0.075 per 1M cached input tokens
- 400K context and 128K maximum output
- Five reasoning-effort levels
- Computer use and tool search supported
- 10% regional-processing uplift
- Tool-specific fees can apply
GPT-5.4 nano API
$0.20 input / $1.25 output per 1M tokens
Cheapest GPT-5.4-class model for bounded classification, extraction, ranking, and simple subagents.
- $0.02 per 1M cached input tokens
- 400K context and 128K maximum output
- Five reasoning-effort levels
- No computer use or tool search
- 10% regional-processing uplift
- Tool-specific fees can apply
Codex with GPT-5.4 mini
30% of GPT-5.4 quota
Mini is available in Codex as a direct model and delegated subagent option.
- Available in Codex app, CLI, IDE extension, and web
- Designed for faster, less reasoning-intensive tasks
- Quota and plan limits apply
- Not a direct dollar price comparison with API billing
ChatGPT with GPT-5.4 mini
Included with eligible access
Mini is available through Thinking for Free and Go and as a fallback for GPT-5.4 Thinking on other plans.
- Does not appear as a normal selectable model in the picker
- Paid users can receive it after GPT-5.4 Thinking rate limits
- Enterprise can retain an Auto-routing preference
- GPT-5.4 nano is API-only
Pricing checked . Check current pricing at the source ↗
Assessment
GPT-5.4 mini & nano strengths and limitations
Where it stands out
- Both models provide a large 400K context and 128K output ceiling at small-model prices.
- Mini is a strong balance of coding quality, multimodal understanding, tool breadth, speed, and cost.
- Nano makes high-volume extraction, routing, and classification affordable when the task is tightly bounded.
- Five reasoning settings let developers tune latency and quality per route.
- Structured outputs and function calling make both models practical application components.
- Mini supports computer use, tool search, code execution, shell, patching, skills, MCP, web search, and file search.
- Nano keeps most tool support while lowering token cost for simple delegated work.
- Fixed snapshots support repeatable regression tests and controlled deployments.
- Cached-input rates reduce the cost of repeated instructions and shared context.
- Codex quota treatment makes mini useful for parallel supporting tasks.
- API business data is not used for training by default, with approved retention-control options available.
What to consider
- Neither model is the newest cost-optimized OpenAI generation; GPT-5.6 Luna has newer knowledge, a larger context, and a slightly lower output price than GPT-5.4 nano in the current catalog.
- The August 31, 2025 knowledge cutoff is stale for current events, laws, products, prices, security issues, and people unless the workflow supplies live sources.
- Small models can be more brittle than frontier models on ambiguous requirements, deep planning, long chains of inference, subtle exceptions, and adversarial inputs.
- A 400K context limit does not imply reliable use of every token. OpenAI's own long-context benchmarks show substantial accuracy gaps versus full GPT-5.4.
- Nano is substantially weaker than mini on computer-use and complex tool benchmarks and does not support the computer-use or tool-search tools.
- Mini's ChatGPT presence is limited and routing-dependent; it is not an ordinary selectable model for most paid users.
- Nano is API-only and cannot be selected in ChatGPT or Codex as described for mini.
- Both can hallucinate facts, citations, code behavior, extracted fields, and tool results.
- Higher reasoning effort adds latency and token use without guaranteeing correctness.
- Tool calling can select the wrong tool, use incorrect arguments, repeat an action, or follow malicious instructions in retrieved content.
- Computer use can misread a screenshot or interact with the wrong control; every consequential action needs preview, scope limits, and confirmation.
- Structured output guarantees shape, not truth. A valid JSON object can still contain wrong labels, dates, totals, or identifiers.
- Fine-tuning, native audio input/output, and native video input/output are not supported.
- Regional processing costs 10% more, and search, code, computer-use, image, or other tools can add charges beyond tokens.
- Default API abuse-monitoring logs may retain customer content for up to 30 days even though API data is not used for training by default; verify endpoint and retention controls.
- Classification and ranking can encode bias or proxy discrimination in employment, credit, education, housing, health, insurance, safety, or access decisions.
- Neither model should independently approve code deployment, payments, messages, account changes, medical or legal conclusions, hiring, credit, eligibility, or other high-impact outcomes.
Compare
GPT-5.4 mini & nano alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Consumer
GPT 5.4
Choose full GPT-5.4 when difficult planning, long-context reasoning, and computer-use quality matter more than mini-model cost and latency.
Explore GPT 5.4 →Consumer
GPT 5.5
Choose GPT-5.5 for a newer frontier model when professional reasoning and coding quality outweigh higher token cost.
Explore GPT 5.5 →Coding
GPT-5.3-Codex-Spark
Choose GPT-5.3 Codex Spark for a coding-specific low-latency workflow where its specialized interaction model is the better fit.
Explore GPT-5.3-Codex-Spark →Project Management
Claude
Choose Claude when Anthropic's models perform better on your document, coding, safety, or tool-use evaluations.
Explore Claude →Consumer
GLM-5.1
Choose GLM-5.1 when an open-weight option, self-hosting path, or its current API economics better match the deployment.
Explore GLM-5.1 →Questions
GPT-5.4 mini & nano FAQs
What is the difference between GPT-5.4 mini and nano?
Mini is the stronger model for coding, multimodal work, computer use, and capable subagents. Nano is cheaper and intended for bounded high-volume tasks such as classification, extraction, ranking, and simple supporting agents.
How much does GPT-5.4 mini cost?
The API price is $0.75 per million input tokens, $0.075 per million cached input tokens, and $4.50 per million output tokens. Regional processing adds 10%, and tool fees can apply.
How much does GPT-5.4 nano cost?
The API price is $0.20 per million input tokens, $0.02 per million cached input tokens, and $1.25 per million output tokens. Regional processing and tool charges can add to the total.
What context window do the models have?
Both have a 400,000-token context window and a 128,000-token maximum output. That capacity does not guarantee perfect retrieval or reasoning across the entire context.
Do GPT-5.4 mini and nano support images?
Yes. Both accept text and image input and return text. Native audio and video modalities are not supported.
Which tools do they support?
Both support function calling, structured outputs, web and file search, image generation, code interpreter, hosted shell, apply patch, skills, and MCP in the Responses API. Mini additionally supports computer use and tool search.
Are GPT-5.4 mini and nano available in ChatGPT?
Mini is available to Free and Go users through Thinking and as a GPT-5.4 Thinking fallback for other plans, but it generally is not a selectable picker model. Nano is API-only.
Are they available in Codex?
GPT-5.4 mini is available across Codex surfaces and uses 30% of GPT-5.4 quota. It can also be used for delegated subagent tasks. OpenAI describes nano as API-only.
Which reasoning levels are available?
Both support none, low, medium, high, and xhigh, with none as the default. Evaluate the lowest setting that meets the route's accuracy target.
Should I use nano for every high-volume task?
No. First compare it with GPT-5.6 Luna and other current small models. Route only tasks that meet measured accuracy, latency, cost, and safety thresholds, and escalate ambiguous cases.
Can mini safely operate a computer?
It can use the computer-use tool, but that does not make actions safe. Isolate the environment, minimize permissions, preview steps, defend against prompt injection, and require confirmation for consequential changes.
Does OpenAI train on API data from these models?
OpenAI says API inputs and outputs are not used to train models by default unless the customer opts in. Default abuse-monitoring retention and feature-specific application state still apply, with approved additional controls available.
Bottom line
Our GPT-5.4 mini & nano verdict
GPT-5.4 mini is the more versatile choice: it offers capable coding, broad tools, computer use, and strong subagent economics without paying full-frontier prices. GPT-5.4 nano is useful when the job is narrow, high-volume, schema-constrained, and easy to audit. Neither should be selected from price alone. The newer GPT-5.6 Luna changes the cost comparison, and small-model mistakes can become expensive when multiplied across thousands of tool calls or records. Benchmark on real traffic, route by difficulty, pin snapshots, validate outputs, sandbox tools, and keep humans responsible for consequential results.
Visit GPT-5.4 mini & nano website ↗Similar Consumer Tools
View all Consumer AI tools →
Mistral 4 Small
Mistral 4 Small - Mistral's open-source 119B MoE model combining reasoning, coding, and vision in one package
GLM-5-Turbo
GLM-5-Turbo - Z AI's high-speed agentic model built specifically for OpenClaw
Nemotron 3 Super
Nemotron 3 Super - NVIDIA's open 120B reasoning model with 1M token context for agentic workflows
Minimax M2.7
MiniMax M2.7 - MiniMax's new 'self-evolving' model with strong coding and agentic benchmarks
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.