TheRouter.aiTheRouter.ai
Back to Models

GPT-5.4 Mini

openaiopenai/gpt-5.4-mini

Strongest mini model for coding, computer use, and subagents. Fast and cost-efficient with reasoning token support, 400K context window and 128K max output.

OpenAI announcement β†—OpenAI model docs β†—OpenAI pricing page β†—

GPT-5.4 Mini is OpenAI's strongest mini model yet, released March 17, 2026 alongside GPT-5.4 Nano. It brings the core capabilities of GPT-5.4 β€” coding, reasoning, multimodal understanding, and tool use β€” into a faster, more cost-efficient package designed for high-volume agentic workloads. OpenAI positions it as the subagent workhorse: fast enough for responsive coding assistants, smart enough for complex tool-calling pipelines, and cheap enough to parallelise across many tasks.

On benchmarks, GPT-5.4 Mini scores 54.4% on SWE-bench Pro and 88.0% on GPQA Diamond at xhigh reasoning effort β€” approaching GPT-5.4's 57.7% and 93.0% respectively while running more than 2x faster. It also delivers an impressive 72.1% on OSWorld-Verified, nearly matching the full GPT-5.4, making it one of the best cost-efficient choices for computer-use AI workflows. In Codex, GPT-5.4 Mini consumes only 30% of the GPT-5.4 quota, allowing developers to delegate simpler coding subagent tasks for roughly one-third the cost.

Best for
  • β€’ Coding subagents in multi-model architectures β€” delegate codebase search, targeted edits, and file reviews to GPT-5.4 Mini while a larger model handles planning and orchestration
  • β€’ Computer-use AI workflows (UI automation, screenshot interpretation) β€” 72.1% on OSWorld-Verified approaches full GPT-5.4 at much lower cost
  • β€’ High-throughput tool-calling pipelines β€” 93.4% on Ο„2-bench (telecom) and strong MCP Atlas performance make it reliable for complex API orchestration
  • β€’ Multimodal apps needing real-time image reasoning β€” handles text + image inputs with low latency
Reach for something else if
  • β€’ Zero-shot complex research or scientific reasoning tasks β€” GPT-5.4 or GPT-5.5 will deliver substantially better results at the cost of more latency and higher price
  • β€’ Long-context retrieval-heavy tasks beyond ~64K tokens β€” GPT-5.4 Mini's MRCR v2 performance drops sharply (47.7% at 64K–128K, 33.6% at 128K–256K); use GPT-5.4 Pro or GPT-5.5 for deeper recall requirements
  • β€’ Workloads where GPT-5.4 Nano suffices β€” for simple classification, data extraction, and ranking, Nano costs $0.20/$1.25 vs Mini's $0.75/$4.50 per 1M tokens
Context Length
400K
Max Output
128K
Input Priceper 1M tokens
$1.62/ 1M tokens
Output Priceper 1M tokens
$4.86/ 1M tokens

Modalities

textimage→text

Capabilities

Vision

Pricing Breakdown

TypeRate
Input$1.62 / 1M tokens
Output$4.86 / 1M tokens

Supported Parameters

temperaturemax_tokenstop_ptoolstool_choiceresponse_formatreasoningstop

Specifications

Release date2026-03-17openai.com β†—verified
OpenAI model idgpt-5.4-minidevelopers.openai.com β†—verified
Training cutoffNot publicly disclosedunknown
Reasoning effort levelsnone (default), low, medium, high, xhighopenai.com β†—verified
Cached input pricing$0.075 / 1M tokens (10x discount vs standard input)developers.openai.com β†—verified
AvailabilityOpenAI API, Codex, ChatGPT (Free/Go/Plus/Team/Pro)openai.com β†—verified
LicenseOpenAI Usage Policy (proprietary, API-only)verified

Benchmarks

BenchmarkDistributionScoreSource
SWE-bench Pro
xhigh reasoning effort; vs GPT-5.4 at 57.7% and GPT-5 mini (high) at 45.7%
54.4%%openai.com β†—
Terminal-Bench 2.0
xhigh reasoning effort; GPT-5.4 at 75.1%, GPT-5 mini at 38.2%
60.0%%openai.com β†—
GPQA Diamond
xhigh reasoning effort; GPT-5.4 at 93.0%, GPT-5 mini at 81.6%
88.0%%openai.com β†—
OSWorld-Verified
xhigh reasoning effort; GPT-5.4 at 75.0%, GPT-5 mini at 42.0%
72.1%%openai.com β†—
Toolathlon
xhigh reasoning effort; GPT-5.4 at 54.6%, GPT-5 mini at 26.9%
42.9%%openai.com β†—
MCP Atlas
xhigh reasoning effort; GPT-5.4 at 67.2%, GPT-5 mini at 47.6%
57.7%%openai.com β†—
Ο„2-bench (telecom)
xhigh reasoning effort; GPT-5.4 at 98.9%, GPT-5 mini at 74.1%
93.4%%openai.com β†—
HLE w/ tool
xhigh reasoning effort; GPT-5.4 at 52.1%, GPT-5 mini at 31.6%
41.5%%openai.com β†—
MMMUPro w/ Python
xhigh reasoning effort; GPT-5.4 at 81.5%, GPT-5 mini at 74.1%
78.0%%openai.com β†—
Graphwalks BFS 0K–128K
Long-context recall; GPT-5.4 at 93.1%, GPT-5 mini at 73.4%
76.3%%openai.com β†—

API Usage Examples

Use the global api.therouter.ai endpoint shown below for new integrations; the legacy China accelerated endpoint is retired.

cURL
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-mini",
    "messages": [
      {"role": "user", "content": "Summarize the key points from this input."}
    ]
  }'

API guide

Full API reference β†’

Chat completion

GPT-5.4 Mini supports the full chat completions API including tool calls, structured output, streaming, and reasoning effort control.

cURL
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-mini",
    "messages": [
      {"role": "system", "content": "You are a helpful coding assistant."},
      {"role": "user", "content": "Write a Python function to merge two sorted lists."}
    ],
    "reasoning_effort": "medium"
  }'

More from openai

openai/gpt-5.4Full-sized flagship β€” stronger across the board, but slower and more expensiveopenai/gpt-5.4-nanoSmallest GPT-5.4 variant β€” cheapest option for classification, extraction, and simple subagentsopenai/gpt-5.5Latest GPT flagship β€” same per-token latency as GPT-5.4 with higher intelligence and ~40% fewer output tokensopenai/gpt-5.4-pro1.05M context window β€” enterprise-grade recall for long-document and codebase tasks

Similar models

Cross-provider sibling models
anthropic/claude-sonnet-4.6
Anthropic's mid-range reasoning model β€” strong competitor for coding and agent workflows
google/gemini-2.5-pro
Google's flagship reasoning model β€” 1M+ context and strong multimodal
deepseek/deepseek-v3.2
Cost-effective reasoning alternative β€” competitive on coding benchmarks
moonshot/kimi-k2
Long-context reasoning model from Moonshot β€” strong at agentic coding tasks

News & changes

2026-03-17

OpenAI launches GPT-5.4 Mini and Nano

OpenAI released GPT-5.4 Mini and Nano, positioning them as fast, affordable models purpose-built for subagent workloads. GPT-5.4 Mini scores 54.4% on SWE-bench Pro and 88.0% on GPQA Diamond at xhigh reasoning effort, approaching the full GPT-5.4 while being more than 2x faster. The models are available via the API, Codex, and ChatGPT (Mini is available to Free/Go users via the Thinking feature).

re-authored by TheRouteropenai.com β†—

Frequently asked

How does GPT-5.4 Mini compare to GPT-5.4?

GPT-5.4 Mini delivers approximately 90-95% of GPT-5.4's benchmark performance on coding (SWE-bench Pro: 54.4% vs 57.7%), reasoning (GPQA: 88.0% vs 93.0%), and computer-use tasks (OSWorld: 72.1% vs 75.0%) while being more than 2x faster and costing significantly less ($0.75/$4.50 per 1M tokens vs $2.50/$10.00). The Mini's main weakness is long-context retrieval β€” MRCR v2 scores drop sharply past 64K tokens, whereas GPT-5.4 maintains strong recall at 1M tokens.

openai.com β†—
What reasoning effort levels does GPT-5.4 Mini support?

GPT-5.4 Mini supports five reasoning effort levels: none (default), low, medium, high, and xhigh. The default is 'none', which returns the fastest responses. For complex coding or reasoning tasks, use 'xhigh' to match the benchmarks published by OpenAI. Unlike GPT-5.4, there is no separate 'xhigh' mode description in the API β€” the string 'xhigh' is passed directly via the reasoning_effort parameter.

openai.com β†—
Is GPT-5.4 Mini available on ChatGPT Free?

Yes. GPT-5.4 Mini is available to Free and Go users via the 'Thinking' feature in the + menu of ChatGPT. For Plus, Team, and Pro subscribers, GPT-5.4 Mini serves as a rate-limit fallback when GPT-5.4 Thinking is at capacity.

help.openai.com β†—
Can I use GPT-5.4 Mini in Codex?

Yes. GPT-5.4 Mini is available across the Codex app, CLI, IDE extension, and web. It consumes only 30% of the GPT-5.4 quota, meaning developers can complete simpler coding tasks in Codex for roughly one-third the cost. Codex can also delegate to GPT-5.4 Mini subagents so that less reasoning-intensive work runs on the cheaper model.

openai.com β†—
Fact ledger β€” every claim on this page traces here
sourceURLretrieved
Release dateopenai.com β†—2026-05-29verified
OpenAI model iddevelopers.openai.com β†—2026-05-29verified
Training cutoffβ€”β€”unknown
Reasoning effort levelsopenai.com β†—2026-05-29verified
Cached input pricingdevelopers.openai.com β†—2026-05-29verified
Availabilityopenai.com β†—2026-05-29verified
Licenseβ€”β€”verified
SWE-bench Proopenai.com β†—2026-05-29verified
Terminal-Bench 2.0openai.com β†—2026-05-29verified
GPQA Diamondopenai.com β†—2026-05-29verified
OSWorld-Verifiedopenai.com β†—2026-05-29verified
Toolathlonopenai.com β†—2026-05-29verified
MCP Atlasopenai.com β†—2026-05-29verified
Ο„2-bench (telecom)openai.com β†—2026-05-29verified
HLE w/ toolopenai.com β†—2026-05-29verified
MMMUPro w/ Pythonopenai.com β†—2026-05-29verified
Graphwalks BFS 0K–128Kopenai.com β†—2026-05-29verified
OpenAI launches GPT-5.4 Mini and Nanoopenai.com β†—2026-05-29verified
How does GPT-5.4 Mini compare to GPT-5.4?openai.com β†—2026-05-29to verify
What reasoning effort levels does GPT-5.4 Mini support?openai.com β†—2026-05-29to verify
Is GPT-5.4 Mini available on ChatGPT Free?help.openai.com β†—2026-05-29to verify
Can I use GPT-5.4 Mini in Codex?openai.com β†—2026-05-29to verify
Help & contact