This small OpenAI model has a 16,384 token max output and scored 87.2% on HumanEval at launch. It was retired from ChatGPT in February 2026 but remains live via the API and Azure, with a newer small model recommended for new projects.
GPT-4o mini is OpenAI's small multimodal model, scoring 82.0% on MMLU with vision input and tool-calling support. OpenAI retired it from ChatGPT in February 2026, though it remains available through the API and Azure for existing integrations, positioned as a legacy option behind OpenAI's newer small models.
Where it sits
- $0.262/M$ per 1M tokensBlended price (3:1)Lower is better#11 / 60peer median $1.86/Mvendor price, checked by HokAI
- 62 tok/stokens/sOutput speedHigher is better#25 / 34peer median 90 tok/scited: Artificial Analysis
- 7.8%% solvedSWE-bench VerifiedHigher is better#27 / 27peer median 78%per source, see benchmark scores
- --% correctGPQA DiamondHigher is better-- / 41peer median 86.9%per source, see benchmark scores
Cheaper than 83% of the 60 GA models with a published price, in the bottom third on SWE-bench Verified (rank 27 of 27), and one of 20 that document a zero-data-retention option. Ranked against GA models; this record is not GA.
Ranks are against GA models on HokAI that publish the same figure; ties share a rank.
Provider: OpenAI · Family: GPT-4o
Context window: 128,000 tokens · Max output: 16,384
Input modalities: text, image · Output: text, tool-calls
About GPT-4o mini
GPT-4o mini is OpenAI's small, cost-efficient multimodal model, announced on July 18, 2024, as the successor to GPT-3.5 Turbo in the API and ChatGPT free tier. It is a dense transformer in the GPT-4o family, distilled and optimized for low latency and low cost rather than peak capability. OpenAI has not disclosed an exact parameter count, but the model is widely understood to sit far below GPT-4o in size, trading raw reasoning power for a price point well below GPT-3.5 Turbo at launch. Its role in the lineup is the default "small" tier: the model developers reach for when a task does not need frontier reasoning but still benefits from instruction-following, tool use, and vision. On academic benchmarks, GPT-4o mini scores 82.0% on MMLU and 87.2% on HumanEval, which was competitive with mid-tier models at its 2024 launch but has since been surpassed by newer small models such as GPT-4.1 mini and Gemini 2.0 Flash. On agentic coding evaluations it scores far lower than frontier models: SWE-bench Verified sits around 7.8%, compared to 70%+ for 2026 frontier models like Claude Opus 4.7 or GPT-5.1. Its Chatbot Arena (LMArena) Elo is approximately 1274, well below the range occupied by 2026 frontier systems. GPQA Diamond and AIME 2025 scores have not been independently published for this model, reflecting that it was never positioned as a reasoning model. The honest read is that GPT-4o mini's benchmark profile reflects a 2024-era small model: solid at general knowledge and basic coding, weak at multi-step agentic and graduate-level reasoning tasks that define 2026 comparisons. The context window is 128,000 tokens with a maximum output of 16,384 tokens per request, matching the broader GPT-4o family's input ceiling. OpenAI has not published a dedicated long-context recall evaluation for the mini variant; independent needle-in-haystack tests of the GPT-4o family generally show solid retrieval up to the full context window but with some degradation in the middle of very long contexts, a pattern common to most 2024-generation models. There is no extended-context tier above the base window for this model. GPT-4o mini supports text and image inputs with text output in the API, plus function calling and structured outputs (JSON mode and JSON schema). Audio and video inputs were promised at launch as "coming in the future" but in practice ship as separate specialized variants (gpt-4o-mini-transcribe, gpt-4o-mini-tts, gpt-4o-mini-audio-preview) rather than as native modalities of the base chat model. OpenAI deprecated the gpt-4o-mini-realtime-preview and gpt-4o-mini-audio-preview variants on their own retirement track, with a shutdown date of May 7, 2026, and renamed their successors gpt-realtime-mini and gpt-audio-mini. Vision support covers document understanding, screenshot interpretation, and basic visual classification, and pairs with function calling so the model can reason over an image and then call a tool. There is no native computer-use or browsing capability in this model; those live in separate OpenAI products. Pricing has remained stable since the July 2024 launch, positioning GPT-4o mini as one of the cheapest models with vision and tool use still available from a frontier lab; exact current input, output, cached, batch and fine-tuned rates plus worked cost examples are detailed in the pricing section below rather than repeated here. Its low per-token cost is the main reason teams keep it in production for high-volume, low-complexity workloads even as newer, more capable small models have emerged. The model is available through the direct OpenAI API and through Azure OpenAI Service (as gpt-4o-mini), where it sits alongside transcription and audio-preview variants. OpenAI has not published a dedicated AWS Bedrock or Google Vertex AI listing for GPT-4o mini, since those platforms primarily host first-party Amazon, Anthropic, and Google models alongside select third-party open-weight models rather than OpenAI's proprietary lineup. SDKs are available for Python, Node.js/TypeScript, and via community libraries for Java, Go, and Ruby through the standard OpenAI-compatible API surface. Safety follows the GPT-4o system card: OpenAI applies its Moderation API and safety classifiers during training data curation to filter CSAM, hateful content, violence, and CBRN-related material, alongside human preference alignment (RLHF) and red-teaming before release. Users can opt images out of training, with fingerprinting used to remove opted-out images from future training runs. The model defaults to standard refusal behavior for clear policy violations, consistent with OpenAI's usage policies, and supports the same moderation endpoint and system-prompt-based steering as other GPT-4o family models. The model's training data stops in October 2023, the same cutoff as the rest of the original GPT-4o family. OpenAI's default API data retention is 30 days for abuse monitoring, with zero-data-retention available to approved enterprise customers. OpenAI maintains SOC 2 Type II compliance for the API platform and offers a Business Associate Agreement for HIPAA-eligible workloads on enterprise plans, alongside GDPR-aligned data processing terms. As of 2026, GPT-4o mini has been retired from the ChatGPT consumer product (alongside GPT-4o, GPT-4.1, GPT-4.1 mini, and o4-mini, retired February 13, 2026) but remains available via the API and Azure OpenAI for existing integrations. OpenAI is steering new users toward GPT-5.1 mini for general small-model tasks and o3-mini or GPT-5.1 mini for reasoning-heavy workloads. Teams should treat GPT-4o mini as a legacy-but-supported option: fine for stable production pipelines already built on it, but new projects should default to GPT-5.1 mini, which offers materially better reasoning and agentic benchmarks. Anyone needing native audio or video input should look at GPT-4o (full) realtime variants or GPT-5.1, not the mini text/vision tier.
Pricing
$0.15 per 1M input tokens, $0.60 per 1M output tokens, unchanged since launch. Batch API: $0.075/$0.30 (50% off). Fine-tuned inference: $0.30 input / $1.20 output per 1M tokens; fine-tuning training itself costs $0.30 per 1M training tokens.
What a real job costs
| Job | Input | Output | Total |
|---|---|---|---|
| Summarise a 20-page PDF | $0.0045 | $0.0006 | $0.0051 |
| Support reply | $0.0003 | $0.0002 | $0.0005 |
| One coding agent run | $0.030 | $0.012 | $0.042 |
Budgets: 20-page PDF = 30k in / 1k out · Support reply = 2k in / 300 out · Coding agent run = 200k in / 20k out. Computed from the vendor's per-token prices at render time; cached-input discounts are not applied.
Key Features
- 128K context window: Matches the broader GPT-4o family's 128,000 token input capacity for document and long-prompt workloads.
- Native vision input: Accepts images alongside text for document understanding, screenshot interpretation, and visual classification.
- Function calling and structured outputs: Supports JSON mode and JSON schema structured outputs, making it usable as a low-cost tool-calling layer.
- Batch API discount: Async Batch API processes requests at a discount versus the standard synchronous price, with results returned within 24 hours.
- Fine-tuning support: Can be fine-tuned on custom data, with inference on the fine-tuned model priced higher per token than the base model.
Pros
- Among the cheapest vision-and-tool-use models available from a frontier lab, with per-token pricing unchanged since launch.
- A six-figure context window with a correspondingly large max-output ceiling is generous for a small-tier model.
- Still accessible via API and Azure OpenAI after the February 2026 ChatGPT retirement, so existing integrations keep working.
Cons
- SWE-bench Verified around 7.8%, unsuitable for agentic coding compared to 2026 frontier models.
- Training data cutoff of October 2023 is dated relative to 2025-2026 models.
- Retired from ChatGPT consumer product in February 2026; OpenAI is steering new projects toward its newer small models instead.
Benchmarks
- MMLU: 82% vendor-reported · 18 Jul 2024 — General-knowledge exam across 57 subjects, % correct.
- HumanEval: 87.2% vendor-reported · 18 Jul 2024 — Small programs that must pass hidden tests, % passing.
- LMArena Elo: 1274 vendor-reported · 18 Jul 2024 — Rating from blind human votes on which answer is better.
- SWE-bench Verified: 7.8% vendor-reported · 18 Jul 2024 — Real GitHub issues fixed end to end, % solved.
- AA blended price: $0.26/M cited: Artificial Analysis · 14 Jun 2026 — Price per 1M tokens at a 3:1 input to output blend, as listed by Artificial Analysis.
- Output speed: 62 tok/s cited: Artificial Analysis · 14 Jun 2026 — Median tokens written per second as measured by Artificial Analysis.
A benchmark is an exam, not the job. Scores transfer unevenly between tasks, so weigh the one closest to your workload and read every figure with its source.
Frequently Asked Questions
How much does GPT-4o mini cost in 2026?
GPT-4o mini costs $0.15 per 1M input tokens and $0.60 per 1M output tokens, unchanged since launch, with cached input at $0.075 per 1M. The Batch API offers a further 50% discount at $0.075 input / $0.30 output per 1M tokens, with results returned within 24 hours. Fine-tuning costs $0.30 per 1M training tokens, after which fine-tuned inference rises to $0.30 input / $1.20 output per 1M tokens. OpenAI's newer small models cost more per token but score higher on reasoning benchmarks, so GPT-4o mini remains attractive mainly for high-volume, low-complexity workloads.
Does GPT-4o mini have a free plan?
GPT-4o mini was available to free ChatGPT users with usage caps before OpenAI retired it from the ChatGPT consumer app in February 2026, so that free consumer access no longer applies. API and Azure OpenAI access have never included a free tier: both require a paid account with prepaid credits at the rates listed above. Anyone who needs a genuinely no-cost API option should look elsewhere, since GPT-4o mini has none.
Which models compete with GPT-4o mini in 2026?
The closest budget-tier competitors are Google's Gemini 3.1 Flash-Lite, which pairs a much larger context window with current 2026 benchmark scores; Mistral's open-weight Ministral 3 8B, which can be self-hosted or fine-tuned outside OpenAI's infrastructure; and MiniMax M3, an open-weight model built for stronger agentic coding than GPT-4o mini's dated benchmark profile. OpenAI's own recommended successor, GPT-5.1 mini, is the direct in-ecosystem upgrade for new projects.
What separates GPT-4o mini from GPT-5.1 mini?
GPT-5.1 mini is OpenAI's current small-model recommendation and clears agentic and reasoning benchmarks that GPT-4o mini was never tuned for, since GPT-4o mini predates OpenAI's reasoning-model line entirely. GPT-4o mini's main advantage is inertia: it is a known quantity already running in production pipelines built before its ChatGPT retirement, with per-token pricing untouched since its original launch. Teams starting a new project have little reason to pick GPT-4o mini over GPT-5.1 mini unless they specifically need to match existing infrastructure behavior or cost.
How do you get started with GPT-4o mini?
Getting started requires an OpenAI developer account with billing enabled, since there is no free API tier for it. From there, request an API key and call the model by name through the Chat Completions or Responses endpoint. Teams already on Microsoft's cloud can instead reach the same model through Azure OpenAI Service under the identical gpt-4o-mini name, authenticating with Azure AD or an API key rather than an OpenAI key.
Top Alternatives
- Gemini 3.1 Flash-Lite: Pick Gemini 3.1 Flash-Lite for its much larger context window and current 2026 benchmark scores; pick GPT-4o mini if you already have an OpenAI-based pipeline you don't want to migrate.
- Ministral 3 8B: Pick Ministral 3 8B if you need open weights you can self-host or fine-tune; pick GPT-4o mini for a fully managed API with nothing to deploy or maintain.
- MiniMax M3: Pick MiniMax M3 for stronger agentic coding on a much larger context window; pick GPT-4o mini for a stable, well-documented pipeline already running on OpenAI or Azure.