OpenAI launches GPT-4o mini, cheaper and more capable than GPT-3.5

On July 18, 2024, OpenAI launched GPT-4o mini as its most economical small model and GPT-3.5’s replacement in ChatGPT. The release announcement publishes prices, context, modalities and the company’s own results; reading it well requires separating cost per token from the cost of completing a task at acceptable quality.

The launch matters for a very practical reason: much of the business use of artificial intelligence does not require the most powerful model available. Classifying documents, summarizing conversations, extracting data from invoices, answering frequently asked questions and assisting human agents are repetitive, high-volume tasks. For these use cases, the price of each query often matters more than the ability to solve an exceptionally complex problem.

A much lower cost for products handling millions of queries

GPT-4o mini will cost 15 cents per million input tokens and 60 cents per million output tokens. A token is a unit of text that models use to read and produce language; in Spanish, it does not correspond exactly to a word, but it provides a way to measure the amount of text processed. Primary source

That is 70% cheaper for input and 60% cheaper for output than GPT-3.5 Turbo, which cost 50 cents and $1.50 per million tokens, respectively. Primary source

The comparison helps put the move in context. The GPT-3.5 family was the technological foundation of ChatGPT when it launched in late 2022, but its performance had fallen behind later models. OpenAI is now trying to ensure customers do not have to choose between that inexpensive but aging model and GPT-4o, which is more capable but also more expensive.

Better performance, with a 128,000-token context window

GPT-4o mini is designed for lower-cost, high-volume tasks.

OpenAI says it scores 82% on MMLU, a widely used test of knowledge and reasoning across multiple subjects; 87% on MGSM, which focuses on mathematical problems written in different languages; and 87.2% on HumanEval, which measures the ability to write code. As with any set of benchmarks, these figures are a useful reference, but they do not replace testing with each company’s real-world data and tasks. Primary source

The model supports a 128,000-token context window and can generate up to 16,000 tokens in a single response. In simple terms, it can process lengthy documents, extensive conversation histories or large blocks of code without requiring developers to split them across many smaller requests. Primary source

It also accepts text and images as input and responds with text. This makes it possible, for example, to build applications that read a screenshot, interpret a chart or extract information from a scanned document. OpenAI plans to add audio and video inputs and outputs later, the modalities that already define GPT-4o.

More useful for automating specific tasks

The company has added support for function calling, which allows the model to ask an application to query a database, send an email or carry out an action defined by the developer. Combined with JSON output generation, this makes it easier to integrate the model into automated workflows without relying on text taking an unpredictable form.

That detail matters to businesses. A conversational assistant can tolerate a somewhat imprecise answer or an awkwardly worded sentence; a system that fills out a product listing or routes a support ticket needs structured, verifiable results. Small models are becoming the standard building blocks for these workflows, while larger models are reserved for difficult queries or for reviewing uncertain cases.

GPT-4o mini is OpenAI’s first model to apply an instruction hierarchy, designed to help it distinguish between system commands, user instructions and the external content it analyzes. It does not eliminate errors or attempts to manipulate the model, but it is an important layer when the model is connected to tools or handling third-party information.

Available in ChatGPT and for developers

Starting today, GPT-4o mini replaces GPT-3.5 Turbo for ChatGPT Free, Plus and Team users. ChatGPT Enterprise users will receive it next week. In the API, it is available as gpt-4o-mini for those building their own applications.

OpenAI is keeping GPT-3.5 Turbo on its developer platform, so the change does not immediately require existing services to migrate. However, the difference in price and capabilities makes it likely that GPT-4o mini will become the default choice for new low-cost projects.

The release points to a new phase in the competition between models: delivering the best result on an academic test is no longer enough. For AI to become part of everyday products, it must handle millions of requests with sufficient quality and a manageable bill. GPT-4o mini is OpenAI’s answer to that less visible but decisive part of the market.

Price per token is not cost per task

An application pays for input, output, retries, tools, storage and human review. A cheap model can cost more if it needs extra context, writes longer answers or forces repetition. The correct comparison fixes a task, an acceptance threshold and a case set, then divides total spending by accepted outcomes.

Vendor benchmarks are useful for forming hypotheses, not making the choice alone. An internal test should represent real documents, languages, formats and errors. For extraction, measure each field and run validation; for classification, inspect every class; for code, execute tests. The failure type that reaches production matters more than an average.

Routing by difficulty can reduce cost without hiding errors

A workflow may send routine cases to the small model and escalate doubtful ones to another system or person. It needs a verifiable signal: invalid schema, missing source, low confidence or disagreement between runs. The model should not decide alone that its own answer is sufficient.

Pinning versions and logging changes matter too. A remote alias may be updated, so prompt, date, parameters and exact model belong to the result. Before migration, run the same evaluation set and inspect regressions rather than average improvements alone.

The transferable skill is to calculate cost per accepted result instead of cost per unit sold. Keep quality, latency, retries and human correction in the same record. A low price then becomes a measurable advantage rather than an automatic promise of savings.

Correct format does not guarantee correct data

Structured output lets an application reject missing fields or invalid types, but perfect JSON can still contain an invented amount. Validation must combine a schema with domain rules and, where possible, check the answer against a source. Automating format reduces one error class; it does not replace verification.

Before connecting tools, test inputs that try to override instructions or embed actions in documents. Restrict permissions, validate arguments and require confirmation for irreversible operations. The savings from a small model disappear if a cheap answer can trigger an expensive action without control.

The decision record can be brief: version, case set, acceptance rate, critical errors, latency, total cost and share escalated for review. Repeat the measurement with the alternative model. Without that denominator, “cheaper” describes only the vendor’s tariff, not the economics of your product.

The metric must follow the task through to the result accepted by a person or system.