GPT-5 mini

Overview

GPT-5 mini launched on August 7, 2025, alongside full GPT-5 and a third, smaller GPT-5 nano, giving developers three sizes of the same generation to trade off against cost and latency. It's the middle tier: a reasoning model built on the same training recipe as flagship GPT-5, priced at a fifth of the input cost and a fifth of the output cost, with the 400K-token context window carried over unchanged.

A year on, it's easy to think of GPT-5 mini as old news - OpenAI has since shipped GPT-5.1, GPT-5.2, and a whole line of mini variants through GPT-5.4 mini. But the original mini never went away. It's still live in the API at its original price, which is why a COLM 2026 paper on wasted compute in autoresearch agents picked it as a fixed backbone this month: cheap enough to run hundreds of tree-search branches across nine Kaggle-style competitions without burning through a research budget.

Key Specifications

Benchmark Performance

OpenAI published a full side-by-side table across the GPT-5 family and the o-series at launch, run at high reasoning effort. Mini consistently lands a few points under the flagship and comfortably ahead of nano and its predecessor o4-mini:

The pattern holds across nearly every category in OpenAI's benchmark table: mini trades 3-4 points on math and science reasoning for a large price cut, and the gap widens on harder agentic evals. Tau2-bench telecom is the outlier worth flagging - a 22.6-point drop from flagship GPT-5, the steepest fall in the table, suggesting mini strains on multi-turn tool-calling scenarios that require holding a lot of state. One place mini edges out the flagship: OpenAI's internal hard instruction-following eval, where it scores 65.8% against GPT-5's 64.0%, a reminder that smaller doesn't mean worse on every axis.

For where mini lands against the wider field, see our coding benchmarks leaderboard and reasoning benchmarks leaderboard.

GPT-5 mini ships with the same reasoning_effort and verbosity API parameters as flagship GPT-5, letting developers dial down latency for well-defined coding tasks. Source: unsplash.com

Key Capabilities

Reasoning at a discount. Mini inherits GPT-5's core mechanics: a reasoning_effort parameter (minimal, low, medium, high) and a verbosity parameter for controlling answer length, plus support for custom tools that let the model emit plaintext instead of JSON for tool calls. None of that's scaled back for the smaller model - it's the same knob set, just cheaper to turn.

Coding and agentic tasks. At 71.0% on SWE-bench Verified, mini handles real GitHub issue-to-patch workflows competently, though the 71.6% Aider polyglot score (against GPT-5's 88.0%) signals it struggles more on multi-language diff editing than on single-file bug fixes. It's a reasonable fit for AI coding assistants that need to process a high volume of small, well-scoped requests rather than one very hard one.

Long-context retrieval. OpenAI-MRCR (needle-in-haystack retrieval) drops from 95.2% at 128K context on full GPT-5 to 84.3% on mini, and falls further at 256K (58.8% versus 86.8%). Mini's 400K window is real, but retrieval fidelity degrades faster than the flagship's as documents fill more of it - worth testing before betting a RAG pipeline on the far end of that window.

Research-agent backbone. The autoresearch paper's choice of mini is instructive: when a harness is going to burn compute on tree search regardless of which model sits underneath it, a cheap, good-enough reasoning model lets researchers run more search branches per dollar. Its finding that sharing debug knowledge across branches nearly doubled medal counts on a fixed mini backbone is itself an argument that harness design matters more than model choice past a certain capability floor.

Pricing and Availability

GPT-5 mini is available through the OpenAI API on the Responses API, Chat Completions API, and Batch API, with three pricing tiers:

Within its own family at launch, GPT-5 mini sat between GPT-5 ($1.25/$10.00) and GPT-5 nano ($0.05/$0.40) - five times pricier than nano on input, a fifth the cost of the flagship. It's also the model ChatGPT quietly falls back to: free-tier users get roughly 10 GPT-5 messages every 5 hours before chats automatically route to mini for the rest of the window.

Against the family that shipped on top of it since, mini looks like the budget option: GPT-5.4 mini costs $0.75/$4.50 and closes much of the coding gap to its own flagship (54.4% vs 57.7% on SWE-bench Pro), while original GPT-5 mini stays three times cheaper on input. Against GPT-4o mini, the prior generation's budget workhorse at $0.15/$0.60, GPT-5 mini costs more but adds native reasoning, a 3x larger context window, and stronger benchmark scores across the board.

Strengths and Weaknesses

Strengths

• Cheapest entry point into GPT-5-class reasoning still live. $0.25/M input undercuts every newer GPT-5.x mini variant
• Full reasoning toolset carried over from flagship GPT-5. reasoning_effort, verbosity, custom tools, and parallel tool calling all work identically
• Solid coding baseline. 71.0% on SWE-bench Verified is workable for high-volume, well-scoped coding tasks
• 90% caching discount. $0.025/M cached input rewards repeated-context RAG and long system prompt workloads
• Proven at scale in research settings. A fixed backbone of choice in academic autoresearch benchmarking for its cost-to-capability ratio

Weaknesses

• Superseded on capability by every newer mini. GPT-5.4 mini and later variants close the gap to their own flagships more effectively than mini closed the gap to GPT-5
• Steep drops on multi-turn agentic tasks. A 22.6-point fall on Tau2-bench telecom versus flagship GPT-5 is the widest gap in OpenAI's published table
• Long-context retrieval degrades faster than the flagship's. OpenAI-MRCR at 256K falls to 58.8%, a bigger relative drop than GPT-5 sees over the same range
• No parameter disclosure. Architecture details remain unpublished, complicating any on-prem or distillation comparison
• Knowledge cutoff frozen at May 2024. Over two years stale as of mid-2026 for anything needing current-events awareness without web search

Related Coverage

• Recovering Wasted Compute in Autoresearch Agents - COLM 2026 paper using GPT-5 mini as a fixed backbone across nine ML competitions
• GPT-5.4 mini - The current-generation successor in the mini line
• OpenAI o4-mini - The o-series reasoning model GPT-5 mini effectively replaced
• GPT-4o mini - The prior-generation non-reasoning budget model
• Coding Benchmarks Leaderboard - Full cross-model coding rankings
• Reasoning Benchmarks Leaderboard - Cross-model reasoning rankings
• Cost Efficiency Leaderboard - Where mini ranks on price-to-performance

FAQ

What is GPT-5 mini best for?

High-volume, well-scoped tasks - coding assistance, classification, and research-agent backbones - where near-flagship reasoning at a fifth of GPT-5's price matters more than the last few benchmark points.

How does GPT-5 mini compare to GPT-5?

Mini trails GPT-5 by 3-4 points on most reasoning and coding benchmarks (71.0% vs 74.9% SWE-bench Verified, 82.3% vs 85.7% GPQA Diamond) while costing 80% less per token on both input and output.

Is GPT-5 mini still available in 2026?

Yes. It remains live in the OpenAI API at its original August 2025 pricing, even though newer mini variants like GPT-5.4 mini have since launched.

What context window does GPT-5 mini support?

400,000 tokens total, split into 272,000 max input tokens and 128,000 max output tokens.

How much does GPT-5 mini cost?

$0.25 per million input tokens and $2.00 per million output tokens, with a 90% discount to $0.025/M on cached input.

Why would a research paper choose GPT-5 mini over a newer model?

Cost predictability at scale. Tree-search research agents generate hundreds of API calls per run; a cheap, fixed backbone keeps the compute budget in the harness design rather than the model bill.

Sources

• Introducing GPT-5 for developers - OpenAI
• GPT-5 mini Model - OpenAI API Docs
• OpenAI API Pricing
• Recovering Wasted Compute in Autoresearch Agents (arXiv:2608.10424)