OpenAI nano model
GPT-5.4 Nano
A low-cost GPT-5.4-class model for simple, high-volume tasks such as classification, data extraction, ranking, and tightly scoped subagent work.
- Context window
- 400K
- tokens
- Max output
- 128K
- tokens
- Input
- $0.20
- per 1M tokens
- Cached input
- $0.02
- per 1M tokens
- Output
- $1.25
- per 1M tokens
01 / Overview
What GPT-5.4 Nano Is
GPT-5.4 Nano is OpenAI's lowest-cost GPT-5.4-class model, designed for narrow, well-bounded, high-volume tasks where speed and cost matter more than broad reasoning or complex orchestration.
Built for Repetition, Not Ambiguity
OpenAI positions GPT-5.4 Nano around workloads such as classification, data extraction, ranking, and subagent tasks. These workloads share an important characteristic: the objective can usually be described precisely and the output can be validated against a small set of expected shapes.
That makes Nano different from a general-purpose assistant. It is most useful when the application already knows what the model should do and needs to execute that operation repeatedly at low cost.
OpenAI's small-model guidance reinforces this distinction. GPT-5.4 Nano is intended for narrow, well-bounded tasks, preferably with closed outputs such as labels, enums, short JSON objects, or fixed templates. Ambiguous planning-heavy work should be routed to a stronger model rather than compensated for with increasingly complex prompts.
The model is now deprecated. OpenAI announced the deprecation on October 1, 2026, with API shutdown scheduled for April 1, 2027. GPT-6 Luna is the recommended replacement for new and migrated workloads.
- Lowest-cost GPT-5.4-class model.
- Optimized for simple, high-volume operations.
- Strong fit for classification, extraction, and ranking.
- Supports reasoning from
nonethroughxhigh. - Deprecated with GPT-6 Luna as the recommended replacement.
- Provider
- OpenAI
- Family
- GPT-5.4
- Tier
- Nano
- Status
- Deprecated
- Knowledge cutoff
- Aug 31, 2025
- Input modalities
- Text, Image
- Default reasoning
- None
02 / Use cases
Where GPT-5.4 Nano Fits Best
GPT-5.4 Nano works best when a task can be reduced to a clear input, a constrained decision, and a predictable output that must be produced many times.
Closed Outputs Create the Best Operating Envelope
Classification is an ideal example. A model receives a message, record, or document fragment and maps it to one of a known set of labels. The output is simple, easy to validate, and inexpensive to retry if necessary.
Extraction follows the same pattern. Nano can identify fields in messages, documents, support tickets, product records, or other structured and semi-structured inputs when the expected schema is explicit.
Ranking is another suitable workload. The model can score or order a bounded list according to predefined criteria without needing to plan an open-ended workflow.
Nano can also be used as a small subagent, but the delegated work should remain constrained. OpenAI specifically advises against using it for ambiguous multi-step orchestration unless the entire flow is extremely well defined.
Because the model is deprecated, these use cases are now most relevant for maintaining existing integrations, benchmarking migration behavior, or establishing a baseline before moving to GPT-6 Luna.
- Classification into fixed labels or categories.
- Structured extraction into compact JSON or fields.
- Ranking and scoring against explicit criteria.
- Narrow subagent tasks with tightly defined boundaries.
- Legacy workload benchmarking before migration to GPT-6 Luna.
- 01
Classification
Map text, records, or messages into a fixed set of labels with predictable output.
- 02
Data extraction
Extract specific fields into short JSON objects or application-defined schemas.
- 03
Ranking
Score or order bounded candidates using explicit criteria and a constrained response format.
- 04
Narrow subagents
Delegate simple supporting work when the task steps and final output are already well specified.
03 / Pricing
GPT-5.4 Nano Pricing
GPT-5.4 Nano is priced at $0.20 per million input tokens, $0.02 per million cached input tokens, and $1.25 per million output tokens.
Low Unit Cost Was the Core Design Tradeoff
Nano's pricing made it attractive for workloads where one operation might be cheap but the application needs to execute that operation millions of times.
At $0.20 per million standard input tokens, large batches of small classification or extraction requests can remain inexpensive. Cached input is priced at one tenth of the standard input rate, which can further reduce repeated context cost when requests share reusable prefixes.
Output remains more expensive than input, so closed outputs are economically useful as well as easier to validate. A short enum, compact JSON object, or ranking list typically costs less than a long explanatory answer.
Tool calls can introduce separate fees, and regional processing endpoints carry a 10% uplift. Since the model is deprecated, pricing should now be evaluated primarily in the context of maintaining existing traffic or comparing migration economics with GPT-6 Luna.
- $0.20 per 1M input tokens.
- $0.02 per 1M cached input tokens.
- $1.25 per 1M output tokens.
- Tool-specific operations may add separate charges.
- Regional processing endpoints carry a 10% uplift.
1M tokens · USD
- Input
- $0.20
- Cached input
- $0.02
- Output
- $1.25
Example: 10K input + 2K output
- Input cost
- $0.0020
- Output cost
- $0.0025
- Estimated total
- $0.0045
04 / Context
A 400K Context Window for Focused Processing
GPT-5.4 Nano provides a 400,000-token context window and supports up to 128,000 output tokens, giving even a nano-tier model room for substantial inputs when the actual operation remains narrowly defined.
Context Capacity Does Not Change the Model's Intended Role
A 400K context window is large enough to process long documents, substantial record collections, or multiple source items in a single request.
That does not mean Nano should be used as a general long-context reasoning model. The strongest fit is still a bounded operation over that context: extract specific entities from a long document, classify sections, identify relevant records, or rank a known set of candidates.
For subagents, the same principle applies. Give Nano the context necessary for one delegated operation rather than asking it to infer a broader plan from an entire project history.
OpenAI's March 2026 release notes also list compaction support for GPT-5.4 Nano. Compaction can help longer workflows preserve useful state while reducing accumulated context, although Nano remains a poor fit for unconstrained orchestration.
- 400,000-token context window.
- Maximum output of 128,000 tokens.
- Suitable for large inputs with narrow processing goals.
- Supports compaction for longer workflows.
- Context should be scoped to one well-defined operation whenever possible.
Context window
400,000
Max output
128,000
A large input can still be a Nano workload when the requested operation is narrow—for example extracting a fixed schema, classifying sections, or ranking a bounded candidate set.
05 / Reasoning
Reasoning from None to XHigh
GPT-5.4 Nano supports none, low, medium, high, and xhigh reasoning effort, with none as the default.
More Reasoning Does Not Turn Nano into a Planning Model
The full reasoning range gives developers room to increase inference effort on a difficult classification, extraction, or scoring problem without immediately moving to another model.
But model selection still matters. OpenAI's guidance says Nano should be used only for narrow, well-bounded tasks and recommends routing ambiguous or planning-heavy work to a stronger model instead of over-prompting Nano.
That distinction is important in evaluation. If a task requires progressively longer instructions, many special cases, several decision branches, and high reasoning just to remain reliable, the workload may no longer fit Nano's intended operating envelope.
For good Nano prompts, OpenAI recommends critical rules first, exact step order, explicit edge-case behavior, a closed output format, and a correct example.
Simple closed task
Start at none for straightforward classification, extraction, ranking, and fixed-template transformations.
Harder bounded task
Increase reasoning only when the operation remains narrowly defined and evaluation shows a measurable quality gain.
06 / Capabilities
Broad Tools with Important Nano-Tier Limits
GPT-5.4 Nano supports streaming, function calling, structured outputs, search, code execution, shell, patching, Skills, MCP, and image input, but it does not support Computer Use or Tool Search.
More Capable Than a Simple Classification Endpoint
Despite its narrow-task positioning, Nano has a broad integration surface.
Web Search and File Search can ground a request in external information. Code Interpreter and Hosted Shell can execute bounded technical operations. Apply Patch can make structured code modifications. Skills and MCP extend the model into reusable capabilities and external systems.
Function calling and structured outputs are particularly relevant to Nano because they reinforce its best use cases: closed, application-controlled workflows with predictable outputs.
Two notable GPT-5.4 capabilities are absent. Computer Use is not supported, and Tool Search is not supported. Those limitations make GPT-5.4 Mini a better fit when the workload depends on interface interaction or dynamic discovery across a large tool catalog.
- Streaming responses.
- Function calling and structured outputs.
- Web Search and File Search.
- Image Generation and Code Interpreter.
- Hosted Shell and Apply Patch.
- Skills and MCP.
- Computer Use is not supported.
- Tool Search is not supported.
- Fine-tuning is not supported.
- Supported
Structured outputs
Return labels, enums, JSON objects, and other application-controlled response shapes.
- Supported
Function calling
Connect narrow model decisions to application-defined actions.
- Supported
Search and files
Use Web Search and File Search when a bounded task requires external information.
- Supported
Code and shell tools
Use Code Interpreter, Hosted Shell, and Apply Patch for constrained technical operations.
- Not listed
Computer use
OpenAI currently lists Computer Use as unsupported for GPT-5.4 Nano.
- Not listed
Tool Search
Dynamic Tool Search is not supported; tool-heavy dynamic orchestration is better suited to a stronger model.
07 / Evaluation
Strengths and Limitations
GPT-5.4 Nano was built around one clear tradeoff: deliver useful GPT-5.4-class behavior at very low cost for simple, high-volume tasks, while leaving ambiguous planning and richer agent execution to larger models.
Strengths
Low token cost
At $0.20 input and $1.25 output per million tokens, Nano was designed for repeated operations where unit economics matter.
Strong fit for closed outputs
Classification, extraction, ranking, enums, short JSON, and fixed templates align closely with OpenAI's recommended operating envelope.
400K context window
The model can process substantial source material even when the requested operation remains narrow.
Broad integration surface
Function calling, structured outputs, search, code tools, shell, patching, Skills, and MCP support constrained production automation.
What to consider
Deprecated lifecycle
GPT-5.4 Nano was deprecated on October 1, 2026 and is scheduled to shut down on April 1, 2027. OpenAI recommends GPT-6 Luna.
Poor fit for ambiguity
OpenAI recommends routing planning-heavy or ambiguous tasks to a stronger model instead of compensating with increasingly complex prompts.
No Computer Use
The model cannot use OpenAI's built-in Computer Use capability.
No Tool Search
GPT-5.4 Nano cannot dynamically discover tools through Tool Search, limiting its role in large, flexible agent tool ecosystems.
Evaluate before migration
Measure GPT-5.4 Nano against its replacement
Use your real classification, extraction, ranking, and narrow automation prompts to capture a baseline, then compare quality, token usage, and cost before moving production traffic to a newer model.
Start FreeGPT-5.4 Nano is deprecated. For new speed- and cost-sensitive workloads, OpenAI recommends migrating to GPT-6 Luna before the April 1, 2027 shutdown.
Common Questions
What is GPT-5.4 Nano?
GPT-5.4 Nano is OpenAI's lowest-cost GPT-5.4-class model for simple, high-volume tasks such as classification, data extraction, ranking, and tightly scoped subagent work.
Is GPT-5.4 Nano deprecated?
Yes. OpenAI deprecated GPT-5.4 Nano on October 1, 2026 and plans to remove it from the API on April 1, 2027.
What should replace GPT-5.4 Nano?
OpenAI lists GPT-6 Luna as the recommended replacement for GPT-5.4 Nano.
How much does GPT-5.4 Nano cost?
Standard pricing is $0.20 per 1M input tokens, $0.02 per 1M cached input tokens, and $1.25 per 1M output tokens. Tool-specific charges may apply separately.
What is the context window of GPT-5.4 Nano?
GPT-5.4 Nano has a 400,000-token context window and supports up to 128,000 output tokens.
What reasoning levels does GPT-5.4 Nano support?
GPT-5.4 Nano supports none, low, medium, high, and xhigh reasoning effort. None is the default.
What is the knowledge cutoff for GPT-5.4 Nano?
OpenAI lists August 31, 2025 as the knowledge cutoff for GPT-5.4 Nano.
What tasks are a good fit for GPT-5.4 Nano?
The best fits are narrow, well-bounded, high-volume tasks such as classification, structured extraction, ranking, short fixed-template transformations, and constrained subagent work.
How should GPT-5.4 Nano prompts be written?
OpenAI recommends critical rules first, exact step order, explicit edge-case behavior, closed outputs such as labels or short JSON, and a correct example. Ambiguous planning-heavy tasks should be routed to a stronger model.
What tools does GPT-5.4 Nano support?
OpenAI lists Web Search, File Search, Image Generation, Code Interpreter, Hosted Shell, Apply Patch, Skills, and MCP as supported Responses API tools.
Does GPT-5.4 Nano support Computer Use?
No. OpenAI currently lists Computer Use as unsupported for GPT-5.4 Nano.
Does GPT-5.4 Nano support Tool Search?
No. Tool Search is not supported for GPT-5.4 Nano.
Can GPT-5.4 Nano be fine-tuned?
No. The current OpenAI model documentation lists fine-tuning as unsupported for GPT-5.4 Nano.
Model information
Last updated
The specifications, pricing, capabilities, and lifecycle information on this page are based on the official OpenAI documentation for GPT-5.4 Nano and OpenAI's deprecation notice. GPT-5.4 Nano was deprecated on October 1, 2026 and is scheduled to shut down on April 1, 2027, with GPT-6 Luna listed as the recommended replacement.