What OpenAI Released
On March 17, 2026, OpenAI released GPT-5.4 mini and GPT-5.4 nano — two smaller, faster variants of the GPT-5.4 flagship that launched on March 5. Both are available in the API immediately. Mini is also available in ChatGPT; nano is API-only.
What Free ChatGPT Users Actually Get
The rollout announcement is worth reading carefully:
"GPT-5.4 mini is available to Free and Go users via the 'Thinking' feature in the + menu. For all other users, GPT-5.4 mini is available as a rate limit fallback for GPT-5.4 Thinking." — OpenAI ChatGPT Release Notes, March 17, 2026
In plain terms:
- Free and Go users: access GPT-5.4 mini through the Thinking toggle in the
+menu. It will not appear in the standard model picker. - Plus, Pro, Enterprise users: GPT-5.4 mini is a fallback — it kicks in automatically when you hit your GPT-5.4 Thinking rate limit, not as a primary choice.
- API developers: full access with no special toggles.
Note: GPT-5.4 nano is not available in ChatGPT at all. It is API-only.
Benchmark Numbers
| Benchmark | GPT-5.4 mini | GPT-5 mini | GPT-5.4 (full) |
|---|---|---|---|
| SWE-Bench Pro | 54.4% | 45.7% | 57.7% |
| OSWorld-Verified (computer use) | 72.1% | 42.0% | 75.0% |
| tau2-bench (tool use) | 93.4% | 74.1% | — |
| MCP Atlas | 57.7% | 47.6% | — |
| Terminal-Bench 2.0 | 60.0% | — | — |
| Context window | 400K tokens | 128K tokens | 1M tokens |
The OSWorld jump — from 42.0% to 72.1% — is the headline result. OSWorld measures whether a model can navigate operating systems and complete real desktop tasks from screenshots. For reference, humans score 72.4% on the same benchmark. GPT-5.4 mini is just below human baseline; GPT-5 mini was not close.
SWE-Bench Pro measures real GitHub software engineering tasks. Mini's 54.4% closes most of the gap with the full model's 57.7%. For comparison, Claude Opus 4.6 scores ~80% on SWE-Bench Verified (a related but slightly different benchmark).
Source: OpenAI official release and Artificial Analysis, March 17, 2026.
API Pricing
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Batch discount |
|---|---|---|---|
| GPT-5.4 mini | $0.75 | $4.50 | 50% off |
| GPT-5.4 nano | $0.20 | $1.25 | 50% off |
| GPT-5.4 (full) | $2.50 | $20.00 | 50% off |
| GPT-5 mini (previous) | $0.30 | $1.20 | 50% off |
Two things worth noting:
Mini costs 2.5x more than GPT-5 mini ($0.75 vs $0.30 input). You get meaningfully more capability, but this is not a straight price-for-price upgrade at scale.
Codex pricing: inside Codex, GPT-5.4 mini consumes 30% of a GPT-5.4 quota. If you are building multi-agent Codex pipelines where GPT-5.4 handles planning and mini handles subtasks, mini's effective cost drops significantly.
GPT-5.4 Nano: What It Is For
Nano is the cheapest model in the GPT-5.4 family at $0.20/$1.25 per million tokens. It scores 39.0% on OSWorld-Verified — lower than GPT-5 mini's 42.0% — and 52.4% on SWE-Bench Pro. On OSWorld, it is not an improvement over its predecessor.
Where it is designed to win: classification, data extraction, ranking, and background subagents handling simple tasks. If you are running a pipeline where an agent needs to categorize 10,000 support tickets overnight, nano's batch pricing ($0.10/$0.625) makes it the cheapest option in the GPT family for that workload.
The Subagent Pattern OpenAI Is Pushing
OpenAI explicitly designed both models for orchestration hierarchies:
- GPT-5.4 (full) handles planning, coordination, and final judgment
- GPT-5.4 mini handles parallel subtasks: codebase search, file review, document processing
- GPT-5.4 nano handles the simplest background tasks: classification, extraction
GitHub Copilot rolled GPT-5.4 mini into general availability on the same day it launched — March 17, 2026. That signal reflects confidence in the model for interactive coding workflows where response latency matters.
What Mini Does Not Do Well
The long-context benchmark tells a clear story: on OpenAI MRCR v2 with 8 needles at 64K–128K context, GPT-5.4 mini scores 47.7% versus the full model's 86.0%. If your task requires tracking many details across a long document, the full GPT-5.4 still has a large advantage.
Frontend code generation also remains a weak point — the benchmark improvements in other areas do not carry over to that specific task.
FAQ
Is GPT-5.4 mini better than GPT-5 mini for coding?
Yes, clearly. SWE-Bench Pro goes from 45.7% to 54.4%, and tool-calling reliability improves significantly (tau2-bench: 74.1% → 93.4%). For agentic coding workflows, the upgrade is worth it if your costs allow for the 2.5x price increase.
Can I select GPT-5.4 mini directly in ChatGPT?
No — it will not appear in the model picker for any tier. Free/Go users access it via the Thinking toggle. Paid users get it as an automatic fallback when rate limits hit. The only way to call it directly is through the API using the model ID gpt-5.4-mini.
When is GPT-5 Thinking mini being retired?
OpenAI stated it will be retired as a selectable option within 30 days of the March 17, 2026 announcement — so approximately April 17, 2026.
How does nano compare to Gemini Flash-Lite on price?
Gemini 3.1 Flash-Lite is positioned at the ultra-cheap inference tier. Nano's $0.20 input pricing is competitive with Google's cheapest models. For classification and extraction tasks, benchmark quality is more relevant than price — run both on your specific data before committing.
Sources
Next step: If you are currently using GPT-5 mini in a production pipeline, pull the OpenAI API benchmark comparison and run your own eval on 50–100 samples from your actual workload. The OSWorld and tool-calling improvements are real, but the SWE-Bench long-context gap means the upgrade is not automatic for every use case.