Free ChatGPT users are getting another way to try a newer reasoning model. OpenAI’s March 17 launch of GPT-5.4 mini makes it available through the Thinking option, while a separate research roadmap points toward a much more ambitious kind of AI assistant. One is a product release; the other is a goal. The distinction is doing quite a lot of work here.
Meet GPT-5.4 Mini and Nano
OpenAI released GPT-5.4 mini and nano on March 17 as smaller models aimed at faster, lower-cost work. Mini is available in ChatGPT, Codex and the API. Nano is API-only; it is not a second new model selectable in free ChatGPT.
The launch announcement gives Free and Go users access to mini through Thinking in the + menu. OpenAI reports that mini runs more than twice as fast as GPT-5 mini and improves on it across several evaluations. Those are the developer’s results, rather than a guarantee of identical speed or success on every prompt.
Here are three examples from OpenAI’s launch table. The GPT-5.4 models use the xhigh reasoning setting, while GPT-5 mini uses its highest available setting, high:
- SWE-Bench Pro (coding): GPT-5.4 mini scores 54.4% vs. 57.7% for the full GPT-5.4 — and 45.7% for the old GPT-5 mini. That’s a meaningful jump.
- OSWorld-Verified (computer use): 72.1% for mini vs. 75.0% for the full model. Basically within spitting distance.
- GPQA Diamond (hard science reasoning): 88.0% vs. 93.0%. Still very solid.
Nano is the smaller option for tasks such as classification, extraction and supporting agent work. Its launch API prices are $0.20 per million input tokens and $1.25 per million output tokens. Small enough for a postage-stamp joke, perhaps, but input and output still have separate prices.
Why This Actually Matters
OpenAI has been pretty explicit about what these models are designed for: speed-sensitive workflows where you can’t wait around for a big model to think. Coding assistants that need to feel snappy. Multimodal apps that interpret screenshots in real time. Subagent systems where a large model plans and smaller models execute.
That last use case is worth pausing on. A larger model can coordinate a job while smaller models take on narrower tasks in parallel. It is less “one AI doing everything” and more “a team of AIs with different job descriptions”. This is a way to build an agent system, not a new guarantee that every ChatGPT conversation automatically runs a team behind the scenes.
At launch, mini uses 30% of the GPT-5.4 quota in Codex. Its standard API prices are $0.75 per million input tokens and $4.50 per million output tokens, with a 400k context window. That context specification belongs to the API; it should not be treated as a promise of the same allowance in the free ChatGPT interface.
The appeal for coding is quick iteration: find the relevant code, make a focused edit, test it and try again. That fits the broader interest in building software with natural-language assistance. It does not make benchmark speed equivalent to a guaranteed autocomplete response time. For another approach to that workflow, see our look at Cursor’s own coding model.
But Wait, There’s a Bigger Story
The mini and nano launch is the present-tense news. The future-tense news is arguably more interesting.
In MIT Technology Review’s March 20 reporting on an interview with chief scientist Jakub Pachocki, OpenAI’s ambition is an AI research intern by September 2026: a system able to take on a limited set of research problems with greater independence. That is a research milestone, not an announced September feature for every ChatGPT account.
The longer-term goal described in that reporting is a fuller automated, multi-agent research system in 2028. The ambition extends beyond drafting a literature summary toward sustained work on difficult research problems. “Planned” is the key word: this is not evidence that the complete system already exists or that human judgement will become unnecessary.
It is a significant ambition. Whether it sounds visionary or wildly optimistic may depend on the last AI answer you had to correct. Improvements in coding and tool use make the direction interesting; they do not establish that every release is better at everything, or that the research schedule will hold.
What Free Users Should Actually Do Right Now
If you are on the free tier, the Thinking option is worth exploring on a task where you can check the answer. The reported improvements are especially relevant to technical work, but a better benchmark score is not a substitute for checking a calculation or running the generated code. Even a clever intern benefits from supervision.
For developers, mini offers another candidate for frequent, narrower tasks. The 400k API context window can accommodate substantial input, but fitting a document is not the same as using every part of it reliably. Longer prompts also cost more. Test a realistic example before deciding that chunking, retrieval or a larger model is no longer needed.
Two Different Kinds of Progress
The immediate news is broader access to a smaller reasoning model. The longer story is OpenAI’s ambition to make research systems more independent. Faster supporting models could be useful building blocks, but the mini launch does not prove that its architecture will deliver the research roadmap on schedule.
Whether the September target holds is still an open question from this March vantage point. In the meantime, there is something concrete to try—and no need to wait for a digital researcher to arrive wearing a visitor badge.
