What is GPT 5.4 Mini?
GPT 5.4 Mini is OpenAI's March 17, 2026 small model, built for coding, computer use, tool calling, and subagents. It sits between the full GPT 5.4 and the cheaper GPT 5.4 Nano, and runs on Unifically as openai/gpt-5.4-mini. It accepts text and image input and outputs text, with a 400k token context window and up to 128k output tokens. The pitch is simple: results close to GPT 5.4 on several agent benchmarks, at more than twice the speed of GPT-5 mini and a much lower price than the full model. Use it where you would run many model calls, not one careful one.
Key features of GPT 5.4 Mini
Near GPT 5.4 coding results
It scores 54.4% on SWE-Bench Pro, close to the full GPT 5.4 at 57.7% and far ahead of GPT-5 mini at 45.7%. Cost per solved task is a fraction of the larger model.
Strong computer use and tool calling
It reaches 72.1% on OSWorld-Verified, near GPT 5.4 at 75.0% and well above Claude Haiku 4.5 at 50.7%. On MCP Atlas tool calling it scores 57.7%.
More than 2x faster than GPT-5 mini
Latency drops by more than half while accuracy rises across coding, reasoning, and tool use. That matters most for subagents that run in parallel.
400k context, 128k output
Large codebases, long reports, or a batch of documents fit in one context, with enough output room for full file rewrites and long structured answers.
Adjustable reasoning effort
Reasoning effort runs from none, the default for instant replies, through low, medium, high, and xhigh. One model covers both quick lookups and hard multi-step problems.
Image input for real-time reasoning
It reads screenshots, charts, and scanned documents alongside text, and scores 88.0% on GPQA Diamond for science questions without tools.
Best for
Coding subagents
Test runs, refactors, and code review passes that fan out in parallel under a larger planner model.
Computer-use agents
Form filling, UI navigation, and desktop-style tasks, with clear action limits.
High-volume tool calling
Function calling, web search, and file search across many requests per minute.
Screenshot and chart reading
Image-plus-text requests where speed matters, like support tickets with screenshots.
Long document work
Extraction and summaries across big files held in the 400k context.
Scaled chat assistants
User-facing chat where latency and per-token cost decide the architecture.
Use cases
Build a coding agent that hands the planning to a larger model and sends GPT 5.4 Mini the subtasks: write the test, fix the lint errors, review the diff. Its speed makes fan-out cheap. Use it for computer-use bots that fill forms, click through internal tools, and read the screen as they go. It also fits document pipelines that pull fields out of contracts, invoices, or scanned pages at scale, with image input handling the messy ones. For user-facing products, it is a good default chat model: fast enough for live typing, smart enough to call tools and cite files.
Limitations
The gap to the full GPT 5.4 is widest on long terminal sessions. GPT 5.4 Mini scores 60.0% on Terminal-Bench 2.0 versus 75.1% for GPT 5.4, so hard multi-step shell work still belongs to the larger model.
Input is text and images only. There is no audio or video input, so voice agents need a separate transcription step.
Training data ends August 2025. For anything current, wire up the web search tool rather than trusting the model's memory.
OSWorld is a controlled benchmark. Real desktops and web apps are messier, so give computer-use agents clear action limits and review steps.
GPT 5.4 Mini vs GPT-5 mini
This is a clean generational jump. GPT 5.4 Mini scores 54.4% on SWE-Bench Pro versus 45.7% for GPT-5 mini, 72.1% on OSWorld-Verified versus 42.0%, 60.0% on Terminal-Bench 2.0 versus 38.2%, and 88.0% on GPQA Diamond versus 81.6%. It does all that while running more than twice as fast. The computer-use gain is the headline: GPT-5 mini was not a serious option for UI agents, and GPT 5.4 Mini is.
When to use GPT 5.4 Mini
Use GPT 5.4 Mini when call volume, latency, or parallelism drives the design: subagents, tool-calling loops, chat products, and document pipelines. Move up to the full GPT 5.4 when a single hard task needs the best answer, especially long terminal work. Drop down to GPT 5.4 Nano for classification, extraction, and other simple tasks where cost per call matters most.
API examples
Call GPT 5.4 Mini from any language by POSTing to /v1/chat/completions, the OpenAI-compatible endpoint shared by every language model on the platform. Full parameter docs live at docs.unifically.com/models/llm/openai/gpt-5.4-mini.
curl -X POST https://api.unifically.com/v1/chat/completions \
-H "Content-Type: application/json" \
-H "Authorization: Bearer YOUR_API_KEY" \
-d '{
"model": "openai/gpt-5.4-mini",
"messages": [
{ "role": "user", "content": "Review this diff and list any bugs it introduces." }
]
}'
The response comes back synchronously with the completion. Set "stream": true to receive tokens as they generate.
FAQs
People also ask
openai/gpt-5.4-mini, called through the OpenAI-compatible POST /v1/chat/completions endpoint with your Unifically API key.
Text and image input with text output. It reads screenshots, charts, and documents in the same request as code or prose. Audio and video input are not supported.
400k tokens, with a max output of 128k tokens. That fits large codebases, long reports, or big document batches in one run.
Yes. It scores 54.4% on SWE-Bench Pro, close to the full GPT 5.4 at 57.7%, and 60.0% on Terminal-Bench 2.0. It runs faster and costs less per token than the larger model.
Yes. It scores 72.1% on OSWorld-Verified, near GPT 5.4 at 75.0% and well above Claude Haiku 4.5 at 50.7%. Real desktops and web apps are messier than the benchmark, so set clear action limits.
Yes. Reasoning effort runs from none (the default, for instant replies) through low, medium, high, and xhigh for the hardest problems.
More than 2x faster than GPT-5 mini, while scoring higher on coding, computer use, and science benchmarks.
Long terminal-heavy agent runs are where the gap to the full GPT 5.4 is widest. Training data ends August 2025, so pair it with a search tool for current facts.