Smallest and fastest GPT-5.4 variant for lightweight agentic tasks. 400K context window and 128K max output with reasoning support.
GPT-5.4 Nano is the smallest, cheapest, and fastest model in OpenAI's GPT-5.4 family, released March 17, 2026 alongside GPT-5.4 Mini. At $0.20 per million input tokens and $1.25 per million output tokens (OpenAI published pricing), it is designed for latency-sensitive, high-throughput workloads where every millisecond and penny matters. Despite its nano-class cost, it inherits the GPT-5.4 architecture stack including reasoning effort levels (none through xhigh), function calling, structured JSON output, and vision input β a first for OpenAI's nano tier.
From a TheRouter operator's view, GPT-5.4 Nano fills a distinct niche below GPT-5.4 Mini. Where Mini ($0.75/$4.50) handles agent pipelines and mid-tier reasoning, Nano ($0.30/$1.75 via TheRouter) is the right model for classification, extraction, ranking, simple sub-agents, and supporting tasks that a larger orchestrator delegates. It is 2-3x faster than GPT-5.4 Standard at matched effort levels and a fraction of the cost. For workloads that do not need the 54.4% SWE-Bench Pro of Mini, Nano's 52.4% is still competitive while being roughly 2.5x cheaper on output. In the nano-tier market it competes directly with GPT-4.1 Nano ($0.10/$0.40) which is cheaper but lacks reasoning effort β for tasks that benefit from even minimal reasoning, Nano's multi-effort support is a meaningful advantage.
| Type | Rate |
|---|---|
| Input | $0.216 / 1M tokens |
| Output | $1.35 / 1M tokens |
| Release date | 2026-03-17openai.com β | verified |
| Training cutoff | 2025-08-31 (OpenAI developer docs)developers.openai.com β | verified |
| OpenAI model id | gpt-5.4-nanodevelopers.openai.com β | verified |
| Reasoning effort | none (default), low, medium, high, xhighopenai.com β | verified |
| Key API features | Structured outputs, function calling, tool use, streaming, vision input, reasoning effort, prompt cachingopenai.com β | verified |
| License | OpenAI Usage Policy (proprietary, API-only) | verified |
| Benchmark | Distribution | Score | Source |
|---|---|---|---|
SWE-bench Pro (Public) Published by OpenAI (xhigh reasoning effort). Compare: GPT-5.4 Mini 54.4%, GPT-5.4 Standard 57.7%, GPT-5 mini (high) 45.7%. | 52.4%% | openai.com β | |
Terminal-Bench 2.0 Published by OpenAI (xhigh reasoning effort). Compare: Mini 60.0%, Standard 75.1%, GPT-5 mini 38.2%. | 46.3%% | openai.com β | |
Toolathlon Published by OpenAI (xhigh reasoning effort). Compare: Mini 42.9%, Standard 54.6%, GPT-5 mini 26.9%. | 35.5%% | openai.com β | |
MCP Atlas Published by OpenAI (xhigh reasoning effort). Compare: Mini 57.7%, Standard 67.2%, GPT-5 mini 47.6%. | 56.1%% | openai.com β | |
Ο2-bench (telecom) Published by OpenAI (xhigh reasoning effort). Compare: Mini 93.4%, Standard 98.9%, GPT-5 mini 74.1%. | 92.5%% | openai.com β | |
GPQA Diamond Published by OpenAI (xhigh reasoning effort). Compare: Mini 88.0%, Standard 93.0%, GPT-5 mini 81.6% (high). | 82.8%% | openai.com β | |
HLE w/ tool Published by OpenAI (xhigh reasoning effort). Compare: Mini 41.5%, Standard 52.1%, GPT-5 mini 31.6%. | 37.7%% | openai.com β | |
MMMUPro Published by OpenAI (xhigh reasoning effort). Compare: Mini 76.6%, Standard 81.2%, GPT-5 mini 67.5%. | 66.1%% | openai.com β | |
MMMUPro w/ Python Published by OpenAI (xhigh reasoning effort). Compare: Mini 78.0%, Standard 81.5%, GPT-5 mini 74.1%. | 69.5%% | openai.com β | |
OSWorld-Verified Published by OpenAI (xhigh reasoning effort). This is Nano's weakest area β far behind Mini (72.1%) and Standard (75.0%). Not recommended for computer-use tasks. | 39.0%% | openai.com β | |
MRCR v2 8-needle 64K-128K Published by OpenAI (xhigh reasoning effort). Long-context retrieval accuracy. Compare: Mini 47.7%, Standard 86.0%, GPT-5 mini 35.1%. | 44.2%% | openai.com β | |
MRCR v2 8-needle 128K-256K Published by OpenAI (xhigh reasoning effort). Compare: Mini 33.6%, Standard 79.3%, GPT-5 mini 19.4%. | 33.1%% | openai.com β | |
Graphwalks BFS 0K-128K Published by OpenAI (xhigh reasoning effort). Graph traversal accuracy. Compare: Mini 76.3%, Standard 93.1%, GPT-5 mini 73.4%. | 73.4%% | openai.com β | |
Artificial Analysis Intelligence Index (xhigh) Artificial Analysis composite benchmark (reasoning, knowledge, math, coding). The median variant (medium effort) scores 38. Both figures were measured by Artificial Analysis, not OpenAI. | 44 | artificialanalysis.ai β |
Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.
curl https://api.therouter.ai/v1/chat/completions -H "Content-Type: application/json" -H "Authorization: Bearer $THE_ROUTER_API_KEY" -d '{
"model": "openai/gpt-5.4-nano",
"messages": [
{"role": "user", "content": "Summarize the key points from this input."}
]
}'Standard OpenAI-compatible chat. Supports reasoning.effort, tools, streaming, vision, and structured JSON output.
curl https://api.therouter.ai/v1/chat/completions \
-H "Authorization: Bearer $THEROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{"model":"openai/gpt-5.4-nano","messages":[{"role":"user","content":"Classify this review: \"Amazing product, fast shipping!\""}]}'OpenAI added AI search to GPT-5.4 models in ChatGPT. While Nano is API-only, the broader investment in the GPT-5.4 architecture pipeline benefits Nano via shared infrastructure and ongoing weight updates.
OpenAI released GPT Image 2 with native reasoning, expanding the GPT-5.4 ecosystem. Nano can be used alongside GPT Image 2 for vision-to-text classification pipelines.
OpenAI launched GPT-5.4 Mini ($0.75/$4.50) and GPT-5.4 Nano ($0.20/$1.25), bringing reasoning effort, vision, tool use, and structured outputs to the small-model tier for the first time. The gpt-5.4-mini and gpt-5.4-nano ids became available in the API immediately.
Yes β GPT-5.4 Nano is the first OpenAI nano-tier model to support reasoning effort levels (none, low, medium, high, xhigh). Previous nano models like GPT-5 Nano and GPT-4.1 Nano did not support this parameter. This makes Nano suitable for tasks that benefit from minimal-to-moderate reasoning.
No β computer_use with native mouse-and-keyboard desktop automation is only available on GPT-5.4 Standard and GPT-5.4 Mini. Nano's 39.0% on OSWorld-Verified confirms it is not designed for this workload.
GPT-5.4 Nano via TheRouter costs $0.30/M input and $1.75/M output. GPT-4.1 Nano costs $0.10/M input and $0.40/M output. GPT-5.4 Nano is ~3x more expensive on input, but adds reasoning effort support, structured outputs, and significantly stronger benchmarks (52.4% vs ~40% SWE-Bench Pro estimated).
Yes β GPT-5.4 Nano accepts image inputs alongside text (modality: text+image β text). It supports reasoning over images for classification, OCR, and UI element analysis, achieving 66.1% on MMMUPro.
No β GPT-5.4 Nano is API-only. GPT-5.4 Mini is available in ChatGPT (Free and Go users via the 'Thinking' feature; as a rate-limit fallback for other tiers). Nano is designed for programmatic usage via the API.
GPT-5.4 Nano has a 400,000 token context window and supports up to 128,000 output tokens β the same as GPT-5.4 Mini. This is sufficient for processing large documents, long conversations, and batch classification workloads.
| source | URL | retrieved | |
|---|---|---|---|
| Release date | openai.com β | 2026-05-29 | verified |
| Training cutoff | developers.openai.com β | 2026-05-29 | verified |
| OpenAI model id | developers.openai.com β | 2026-05-29 | verified |
| Reasoning effort | openai.com β | 2026-05-29 | verified |
| Key API features | openai.com β | 2026-05-29 | verified |
| License | β | β | verified |
| SWE-bench Pro (Public) | openai.com β | 2026-05-29 | verified |
| Terminal-Bench 2.0 | openai.com β | 2026-05-29 | verified |
| Toolathlon | openai.com β | 2026-05-29 | verified |
| MCP Atlas | openai.com β | 2026-05-29 | verified |
| Ο2-bench (telecom) | openai.com β | 2026-05-29 | verified |
| GPQA Diamond | openai.com β | 2026-05-29 | verified |
| HLE w/ tool | openai.com β | 2026-05-29 | verified |
| MMMUPro | openai.com β | 2026-05-29 | verified |
| MMMUPro w/ Python | openai.com β | 2026-05-29 | verified |
| OSWorld-Verified | openai.com β | 2026-05-29 | verified |
| MRCR v2 8-needle 64K-128K | openai.com β | 2026-05-29 | verified |
| MRCR v2 8-needle 128K-256K | openai.com β | 2026-05-29 | verified |
| Graphwalks BFS 0K-128K | openai.com β | 2026-05-29 | verified |
| Artificial Analysis Intelligence Index (xhigh) | artificialanalysis.ai β | 2026-05-29 | to verify |
| GPT-5.4 now AI search on ChatGPT β Nano benefits from ecosystem | openai.com β | 2026-05-29 | verified |
| GPT-5.4 family expands β GPT Image 2 launched | openai.com β | 2026-05-29 | verified |
| GPT-5.4 Mini and Nano released | openai.com β | 2026-05-29 | verified |
| Does GPT-5.4 Nano support reasoning effort levels? | openai.com β | 2026-05-29 | to verify |
| Does GPT-5.4 Nano support computer use? | openai.com β | 2026-05-29 | to verify |
| How does GPT-5.4 Nano compare to GPT-4.1 Nano on pricing? | developers.openai.com β | 2026-05-29 | to verify |
| Does GPT-5.4 Nano support image input? | openai.com β | 2026-05-29 | to verify |
| Is GPT-5.4 Nano available in ChatGPT? | openai.com β | 2026-05-29 | to verify |
| What is the context window and max output for GPT-5.4 Nano? | developers.openai.com β | 2026-05-29 | to verify |