The Deployment Economics Revolution
They told you to pick the biggest model for the best results. That advice just became expensive and wrong.
GPT-5.4 Mini and Nano, released March 17, 2026, rewrite the economics of AI deployment. Mini scores 54.4% on SWE-Bench Pro — within 3.3 points of the $2.50/M-token flagship — at $0.75 per million input tokens. That’s 70% cheaper for 94% of the performance.
Nano goes further: $0.20 per million input tokens. Simon Willison described 76,000 photos for $52. For classification, extraction, and routing tasks, Nano is functionally free at scale.
The Subagent Architecture
The real innovation isn’t the models — it’s how they compose. OpenAI designed Mini and Nano as subagents: a flagship GPT-5.4 handles planning, coordination, and judgment while delegating narrower subtasks to Mini or Nano running in parallel.
In Codex, Mini uses only 30% of the GPT-5.4 quota for simpler coding tasks. The math: flagship handles the hard decisions, Mini handles the volume, Nano handles the classification. Three tiers, each priced for its role, costs drop 90% without meaningful quality loss.
This is the pattern that every AI team will adopt by Q3 2026. If you’re routing every request to a frontier model, you’re overpaying by an order of magnitude.
The Benchmarks
| Model | SWE-Bench Pro | OSWorld | Cost (Input) | Speed | |-------|--------------|---------|--------------|-------| | GPT-5.4 | 57.7% | 78.3% | $2.50/M | Baseline | | GPT-5.4 Mini | 54.4% | 72.1% | $0.75/M | 2x faster | | GPT-5.4 Nano | 52.4% | 39.0% | $0.20/M | 3x faster |
Mini is the workhorse. Nano is the classifier. The flagship is the thinker. Together they’re cheaper than any single model doing all the work.
What’s Missing
Nano scores 39% on OSWorld — useless for visual interface tasks. It’s API-only with no ChatGPT access. And the ‘Thinking’ mode that makes Mini shine on reasoning tasks isn’t available on Free or Go tiers.
The 400K context window on Mini (vs 1M on the flagship) also matters for codebase-scale tasks.
The Verdict
GPT-5.4 Mini is the best value in AI right now. Not because it’s the smartest, but because it’s smart enough at a price that changes what you can afford to build. The subagent architecture — flagship for thinking, Mini for doing, Nano for sorting — is the deployment pattern of 2026. Teams that redesign their process around model tiers will outperform teams that prompt-engineer a single expensive model.