OpenAI unveils dirt-cheap GPT-4o mini

While everyone was waiting for GPT-5, OpenAI just dropped something arguably more practical.

OpenAI officially launched GPT-4o mini, a compact, cost-efficient model built to take over high-volume everyday tasks. It immediately replaces GPT-3.5 Turbo across ChatGPT for Free, Plus, and Team tiers, with Enterprise access rolling out next week.

Slashing costs by 60%

The headline here is the price tag.

GPT-4o mini costs just $0.15 per million input tokens and $0.60 per million output tokens. That makes it over 60% cheaper than GPT-3.5 Turbo, and 99% cheaper per token than 2022's text-davinci-003.

Despite the budget pricing, the benchmarks pack a serious punch. GPT-4o mini scored 82.0% on MMLU, beating Google Gemini Flash (77.9%) and Claude Haiku (73.8%). In fact, OpenAI claims it even outperforms the original GPT-4 on the LMSYS chat preference leaderboard.

On technical tasks, it hit 87.2% on HumanEval coding evaluations and 87.0% on MGSM math reasoning. It features a 128K context window, supports up to 16K output tokens per request, and carries knowledge through October 2023. While it currently accepts text and vision in the API, audio and video support are planned down the road.

Defending against prompt injections

Smaller models often stumble when it comes to guardrails, but OpenAI gave mini a new layer of armor.

It is the first model in OpenAI's API to implement an "instruction hierarchy" method. This technique helps the model separate system instructions from untrusted user inputs, improving resistance to jailbreaks and prompt injection attacks.

Early enterprise partners are already putting it through its paces. Ramp reported significantly better structured data extraction from receipts, while Superhuman integrated it to draft email replies based on long conversation threads.

Why it matters

We think GPT-4o mini signals a critical pivot in the AI race: brute-force intelligence is taking a backseat to operational efficiency.

For builders, this drastically lowers the financial risk of deploying AI agents. Workflows that require chaining dozens of model calls, parsing full codebases, or running high-throughput customer support bots are now genuinely economical. If you have been relying on GPT-3.5 Turbo or juggling budget open-source endpoints, upgrading today is practically a no-brainer.

For the broader industry, OpenAI just fired a direct shot at Google's Gemini Flash and Anthropic's Claude Haiku. It proves the frontier labs won't let rivals dominate the low-cost tier without a fight.

The catch? It looks like mini models will still hit a ceiling on nuanced, multi-step creative work. It won't replace full-fat frontier models for delicate tasks, but for the vast majority of programmatic plumbing, mini is about to do most of the heavy lifting.