benchrPricing breakdown
| Tier | Rate / 1M tokens |
|---|---|
| Standard input | $0.25 |
| Standard output | $2.00 |
| Cached input | $0.025 |
| Batch input | $0.125 |
| Batch output | $1.00 |
| Context window | 128,000 tokens |
The 48% SWE-bench score — what it means in practice
SWE-bench Verified measures model performance on real GitHub issues — the kind that require reading existing code, understanding the test suite, and producing a patch that passes CI. At 48%, GPT-5 Mini resolves fewer than half of these tasks. Compare to GPT-5 at roughly 75% or Claude Haiku 4.5 at 73.3% — the gap is 25+ points.
This matters if you're using GPT-5 Mini for: code generation beyond simple snippets, bug fixing, refactoring, or any task requiring understanding of existing code structure. It doesn't matter for: classifying customer intent, extracting structured data from text, answering factual questions, routing traffic between system components, or generating templated content. Know which category your workload falls into before choosing this model.
Where GPT-5 Mini excels: volume classification
At $0.25/1M input and 160 tok/s, GPT-5 Mini is purpose-built for pipelines where you're making millions of small, well-specified calls. A sentiment classification step at 50M calls/month: $12.50. A routing layer that categorizes incoming customer queries: $8.50/month at 34M queries (assuming 300 tokens average input). These economics make it viable to add LLM intelligence to steps that would be cost-prohibitive with frontier models.
Batch API: the $0.125 route
GPT-5 Mini via the OpenAI Batch API: $0.125/1M input, $1/1M output — the cheapest async LLM pricing from a major provider. For offline document processing, data labeling, eval generation, or any pipeline tolerating 24-hour turnaround, batch Mini is a strong economic choice. At $0.125 per million, a 1B token monthly batch job costs $125.
Cost scenarios
High-volume production at 100M input + 20M output per month: $25 + $40 = $65/month. GPT-5 at the same volume: $125 + $200 = $325/month — GPT-5 is 5× more expensive. Claude Haiku 4.5 at the same volume: $100 + $100 = $200/month — Haiku is 3× more expensive. For classification-only workloads, GPT-5 Mini's cost advantage is the largest of any major provider's offering.
Use-case fit
Best for: Intent classification; sentiment analysis; structured extraction; customer service routing; content moderation; batch document labeling; any high-volume pipeline where LLM calls are made millions of times and the 48% coding floor doesn't apply to your task type.
Skip if: Your tasks involve code generation, bug fixing, or reasoning-intensive analysis — GPT-5 Mini's 48% SWE-bench is a hard ceiling. Also skip for context-heavy tasks exceeding 128K tokens.
Decision checklist
Before choosing GPT-5 Mini: confirm your task type doesn't involve coding or complex reasoning. Run a 100-sample quality eval on your actual task distribution. If accuracy on your task type is acceptable, the cost savings are compelling. If not, Claude Haiku 4.5 at $1/1M or GPT-5 at $1.25/1M are the next steps up.
For batch workloads: calculate monthly token volume and compare $0.125 batch vs $0.25 standard. If turnaround time allows async execution, the batch route cuts costs in half.