OpenAI ships GPT-5.4 mini and nano, faster and more capable but up to 4x pricier
Key Points
- OpenAI has launched two new compact models, GPT-5.4 mini and nano, specifically optimized for coding, subagents, and computer control.
- In its Codex platform, OpenAI showcases a subagent architecture where a larger model like GPT-5.4 handles planning and coordination while delegating simpler subtasks to GPT-5.4 mini.
- The new models come at a significantly higher price point than their predecessors: GPT-5.4 mini costs three times more per million input tokens and over twice as much for output compared to GPT-5 mini, while GPT-5.4 nano is up to four times pricier than its predecessor on input.
OpenAI has released two new compact models—GPT-5.4 mini and nano—built for coding assistants, subagents, and computer control. GPT-5.4 mini nearly matches the full model's performance, but both new models come with a steep price hike over their predecessors.
OpenAI says GPT-5.4 mini delivers major improvements over GPT-5 mini in coding, reasoning, multimodal understanding, and tool usage, while running more than twice as fast. The model gets close to the much larger GPT-5.4 across several benchmarks, hitting 54.4 percent vs. 57.7 percent on the coding benchmark SWE-Bench Pro and 72.1 percent vs. 75.0 percent on OSWorld-Verified, which measures computer usage capabilities.
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| SWE-Bench Pro | 57.7% | 54.4% | 52.4% | 45.7% |
| Terminal-Bench 2.0 | 75.1% | 60.0% | 46.3% | 38.2% |
| Toolathlon | 54.6% | 42.9% | 35.5% | 26.9% |
| GPQA Diamond | 93.0% | 88.0% | 82.8% | 81.6% |
| OSWorld-Verified | 75.0% | 72.1% | 39.0% | 42.0% |
GPT-5.4 nano is the smallest and cheapest option. OpenAI recommends it for classification, data extraction, ranking, and coding subagents that handle simpler supporting tasks. This model is also a big step up from GPT-5 nano.
Big brain plans, small brain grinds
The subagent setup OpenAI shows off in Codex is worth a closer look: a larger model like GPT-5.4 handles planning, coordination, and final evaluation, while farming out parallel subtasks to GPT-5.4 mini or nano subagents.
Those subtasks include things like searching a codebase, scanning large files, or processing supporting docs. OpenAI says GPT-5.4 mini burns only 30 percent of the GPT-5.4 quota in Codex, which should cut costs for simpler tasks to roughly a third.
The mini model also shows a huge jump in computer control. GPT-5.4 mini scored 72.1 percent on the OSWorld Verified benchmark, just a hair behind the full GPT-5.4 at 75.0 percent. GPT-5 mini only managed 42.0 percent.
Coding
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| SWE-bench Pro (Public) | 57.7% | 54.4% | 52.4% | 45.7% |
| Terminal-Bench 2.0 | 75.1% | 60.0% | 46.3% | 38.2% |
Tool-calling
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| MCP Atlas | 67.2% | 57.7% | 56.1% | 47.6% |
| Toolathlon | 54.6% | 42.9% | 35.5% | 26.9% |
| τ2-bench (telecom) | 98.9% | 93.4% | 92.5% | 74.1% |
Intelligence
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| GPQA Diamond | 93.0% | 88.0% | 82.8% | 81.6% |
| HLE w/ tool | 52.1% | 41.5% | 37.7% | 31.6% |
| HLE w/o tools | 39.8% | 28.2% | 24.3% | 18.3% |
MM / Vision / CUA
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| OSWorld-Verified | 75.0% | 72.1% | 39.0% | 42.0% |
| MMMUPro w/ Python | 81.5% | 78.0% | 69.5% | 74.1% |
| MMMUPro | 81.2% | 76.6% | 66.1% | 67.5% |
| OmniDocBench 1.5 (no tools, lower is better) | 0.109 | 0.1263 | 0.2419 | 0.1791 |
Long context
| Benchmark | GPT-5.4 | GPT-5.4 mini | GPT-5.4 nano | GPT-5 mini |
|---|---|---|---|---|
| OpenAI MRCR v2 8-needle 64K-128K | 86.0% | 47.7% | 44.2% | 35.1% |
| OpenAI MRCR v2 8-needle 128K-256K | 79.3% | 33.6% | 33.1% | 19.4% |
| Graphwalks BFS 0K-128K | 93.1% | 76.3% | 73.4% | 73.4% |
| Graphwalks parents 0-128K (accuracy) | 89.8% | 71.5% | 50.8% | 64.3% |
Better performance comes at up to 4x the price
GPT-5.4 mini is now available through the API, Codex, and ChatGPT at $0.75 per million input tokens and $4.50 per million output tokens. Nano is API-only at $0.20 per million input tokens and $1.25 per million output tokens. Both models support a 400,000-token context window.
Compared to the previous mini and nano models in the GPT-5 lineup, that's a serious price bump. According to OpenAI's pricing page, GPT-5 mini ran $0.25 per million input tokens and $2.00 per million output tokens. GPT-5 nano was $0.05 input and $0.40 output per million tokens.
| Model | Input (per 1M tokens) | Output (per 1M tokens) | Input markup | Output markup |
|---|---|---|---|---|
| GPT-5.4 mini | $0.75 | $4.50 | 3.0x | 2.25x |
| GPT-5.4 nano | $0.20 | $1.25 | 4.0x | 3.125x |
| GPT-5 mini | $0.25 | $2.00 | - | - |
| GPT-5 nano | $0.05 | $0.40 | - | - |
OpenAI likely justifies the higher prices by pointing to the performance gains, which bring these compact models much closer to the full-size versions that cost significantly more to run.
AI News Without the Hype – Curated by Humans
As a THE DECODER subscriber, you get ad-free reading, our weekly AI newsletter, the exclusive "AI Radar" Frontier Report 6× per year, access to comments, and our complete archive.
Subscribe now