OpenAI GPT-5 mini
by OpenAI
OpenAI's cost-efficient GPT-5 variant: a 400K context window at a tenth of frontier input pricing
GPT-5 mini is OpenAI's faster, cheaper GPT-5 variant, built for well-defined tasks and precise prompts where latency and volume matter more than frontier reasoning. It pairs a 400,000-token context window with $0.25 per million input tokens, making it the default workhorse for classification, extraction and routing at scale.
GPT-5 mini is OpenAI's cost-efficient variant of GPT-5, documented as a faster, more economical version aimed at well-defined tasks and precise prompts, particularly workloads that need low latency at high volume. It was released on 7 August 2025 and takes text and image input, returning text. Its context window is 400,000 tokens, of which up to 272,000 may be input and up to 128,000 output, and its knowledge cutoff is 31 May 2024 — by some distance the oldest of OpenAI's current mainline models and now roughly two years stale, which matters for any task touching recent events, current pricing or newly released software. Supported API features cover streaming, function calling, file search, file uploads, web search, structured outputs and image input, so it drops into most existing GPT-5 integrations without code changes. List pricing is $0.25 per million input tokens, $0.025 per million cached input tokens and $2.00 per million output tokens, putting input at one tenth of GPT-5.4's rate and cached input at one tenth of that again. Independent measurement by Artificial Analysis places it at an intelligence index of 25, ranking 51st of 162 models and described as above average among comparable models, with output at 90.9 tokens per second against a 107.5 median for peer reasoning models. The same source flags two caveats buyers should price in: time-to-first-token measured at 94.84 seconds at high reasoning effort against a 1.84-second median, and output pricing of $2.00 per million that sits well above the $0.88 median for models in its intelligence band, so the model is cheap on input and comparatively expensive on output. It does not appear on OpenAI's published deprecation schedule.
Platform teams running millions of well-specified LLM calls a day — classification, extraction, routing, summarisation — where a frontier model's per-token cost is the dominant line item.
Roughly a tenth of frontier input pricing with a 400,000-token window, so high-volume pipelines stay affordable without dropping to a materially weaker model.
At a Glance
- Category
- AI Models & APIs
- Pricing
- Usage-based
- Target Market
- CTOs, Enterprise Developers, Data Scientists, Platform Engineers
- Deployment
- API-based, Cloud-only
- Founded
- 2015
- Headquarters
- San Francisco, United States
- Team Size
- 500+
Key Features
- ✓400,000-token context window
Accepts up to 272,000 input and 128,000 output tokens, enough for long documents without chunking, at budget pricing.
- ✓$0.25 per million input tokens
One tenth of GPT-5.4's input rate, which is what makes million-call-per-day pipelines financially viable at all.
- ✓Prompt caching at $0.025 per million
A 90 percent discount on repeated context, so long shared system prompts cost almost nothing across high-volume traffic.
- ✓Structured outputs and function calling
Returns schema-conformant JSON and calls tools directly, removing brittle parsing from extraction and routing pipelines.
- ✓Multimodal text and image input
Processes screenshots, scans and diagrams alongside text, so document pipelines do not need a separate vision model.
- ✓Hosted web search and file search
Server-side retrieval tools partly offset the May 2024 knowledge cutoff without building separate search infrastructure.
Capabilities
Use Cases
- •Support-ticket classification at volume
Route millions of inbound tickets to the right queue where per-token cost, not reasoning depth, is the binding constraint.
- •Document-to-JSON extraction
Pull structured fields from invoices, contracts or forms using structured outputs, with the schema enforced by the API.
- •Tiered inference front end
Handle the routine majority of requests cheaply and escalate only low-confidence cases to a frontier model.
- •Long-document summarisation
Condense reports or transcripts inside a 400,000-token window without splitting the source into overlapping chunks.
- •Content moderation and policy screening
Screen large volumes of user-generated text against fixed policy rules where the task is narrow and well specified.
Ideal For
Best For
- ✓High-volume classification, tagging and routing where prompts are precise and the task is well defined
- ✓Structured extraction from documents into JSON via structured outputs, at input pricing a tenth of GPT-5.4's
- ✓First-pass triage in a tiered inference architecture, escalating only hard cases to a frontier model
- ✓Long-document summarisation that needs a 400,000-token window but not frontier-level reasoning
- ✓Cost-constrained internal tooling and prototypes where per-token spend decides whether a feature ships
Not Ideal For
- ✗Anything depending on knowledge after May 2024 — the cutoff is roughly two years stale, so recent events, current pricing and new library APIs are unreliable without retrieval
- ✗Hard multi-step reasoning: Artificial Analysis scores it at intelligence index 25, and work that fails here gets re-run on a frontier model, erasing the saving
- ✗Interactive, latency-sensitive UX at high reasoning effort, where Artificial Analysis measured 94.84 seconds to first token against a 1.84-second median
- ✗Output-heavy generation workloads — at $2.00 per million output tokens it is well above the $0.88 median for its intelligence tier, so long generations lose the cost advantage
- ✗Agentic desktop or computer-use tasks, which are handled by the GPT-5.4 line rather than the mini variants
Integrations
Deployment
Market Analysis
Pros
- ✓Input at $0.25 per million is a tenth of GPT-5.4's rate, which is what makes high-volume production pipelines economic
- ✓400,000-token context is unusually large for a budget-tier model and removes most chunking logic
- ✓Artificial Analysis rates it above average on intelligence for comparable models, at index 25 and 51st of 162
- ✓Full GPT-5 feature parity — structured outputs, function calling, web search, file search, image input — so workloads move down-tier without a rewrite
- ✓Not on OpenAI's deprecation schedule as of August 2026, unlike the o3 family
Cons
- ✗Knowledge cutoff of 31 May 2024 is roughly two years stale, the oldest of OpenAI's mainline models, so anything time-sensitive needs retrieval bolted on
- ✗Artificial Analysis measured 94.84 seconds to first token at high reasoning effort against a 1.84-second median — at odds with the low-latency positioning
- ✗Output at $2.00 per million is more than double the $0.88 median for its intelligence tier, so generation-heavy work loses the cost advantage
- ✗Output speed of 90.9 tokens per second is below the 107.5 median for comparable reasoning models
- ✗Intelligence index of 25 means hard reasoning fails and gets retried on a frontier model, which can erase the intended saving
- ✗No computer use or agentic desktop capability — those stay on the GPT-5.4 line
Pricing
Standard API
From $0.25 per 1M input tokens
- ✓$0.25 per 1M input tokens
- ✓$0.025 per 1M cached input tokens
- ✓$2.00 per 1M output tokens
- ✓400,000-token context window
- ✓Structured outputs, function calling, web search and file search
Pure per-token metering with no seats or minimums: $0.25 per million input, $0.025 cached and $2.00 output. Input is a tenth of GPT-5.4's rate, which is the whole reason to choose it, but output is not proportionally discounted — Artificial Analysis puts $2.00 per million well above the $0.88 median for models at this intelligence level, so the economics favour large-input, small-output workloads such as classification and extraction rather than long-form generation.
Security & Compliance
Connect
Sources
This page was written from 6 sources, 4 on domains other than developers.openai.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
TwelveLabs
Video intelligence API that makes every hour of enterprise footage searchable, analyzable and agent-ready.
Engram
A learned memory layer that makes AI actually know your organization — at up to 100x fewer tokens.
Mistral OCR 4
Structure-aware document AI that returns bounding boxes, typed blocks, and per-word confidence scores.
Hugging Face
The AI community building the future.