OpenAI o1-mini
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,655 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Sources
10 citations
Review status
Source-backed
Revision
v2 · 1,655 words
Add missing citations, update stale details, or suggest a clearer explanation.
OpenAI o1-mini is a smaller, faster, and cheaper reasoning model released by OpenAI on September 12, 2024, alongside o1-preview, and optimized for science, technology, engineering, and mathematics (STEM) tasks such as coding and competition mathematics.[1][2] OpenAI reported that o1-mini was about 80% cheaper than o1-preview while remaining competitive with the larger o1 model on math and coding benchmarks, scoring 70.0% on the 2024 AIME mathematics competition and roughly 1650 Elo (about the 86th percentile) on Codeforces.[1][6] It was the entry-level member of OpenAI's first o-series of reasoning models (developed under the project codename "Strawberry") and was later superseded by o3-mini, which OpenAI released on January 31, 2025.[1][8]
OpenAI introduced o1-mini in a blog post titled "OpenAI o1-mini: Advancing cost-efficient reasoning," published the same day as the broader o1 announcement.[1] Like the larger models in the family, o1-mini is trained with large-scale reinforcement learning to produce an extended internal chain of thought before answering. The company positioned the model as a cost-efficient option for developers and users who need strong reasoning in technical domains but do not require the wide-ranging factual knowledge of larger general-purpose models such as GPT-4o.[1]
The model belongs to OpenAI's first generation of "reasoning" systems, which differ from earlier chat models in that they are trained to "think" before responding. According to OpenAI, the o1 models use reinforcement learning to learn to refine a long internal chain of thought, allowing them to break problems into steps, try different strategies, and recognize mistakes.[3] o1-mini applies the same reasoning-centric training recipe in a smaller package, trading breadth of knowledge for lower latency and cost while retaining competitive performance on STEM tasks.[1]
OpenAI noted that, because much of the cost of pretraining large models is spent acquiring broad world knowledge that is unnecessary for many STEM applications, a smaller model focused on reasoning could match larger models on technical benchmarks at a fraction of the price.[1] o1-mini was reported with a 128,000-token context window, a maximum output of 65,536 tokens, and a knowledge cutoff of October 2023.[4][5]
| Specification | Value |
|---|---|
| Developer | OpenAI |
| Released | September 12, 2024 |
| Model family | o1 (o-series reasoning models) |
| Context window | 128,000 tokens |
| Maximum output | 65,536 tokens |
| Knowledge cutoff | October 2023 |
| Focus | STEM reasoning (math, coding) |
| API shutdown | October 27, 2025 |
Source: OpenAI o1-mini announcement and model card.[1][4][5][10]
o1-mini is part of OpenAI's o1 series, all of which are trained to generate a long chain of thought before producing a final answer. OpenAI described this as the model learning, through reinforcement learning, to "hone its chain of thought and refine the strategies it uses," including recognizing and correcting its own errors and trying alternative approaches when one fails.[3]
In the application programming interface (API), this internal reasoning is billed as "reasoning tokens." These tokens are not returned in the visible API output but count toward usage and are priced at the same rate as output tokens.[2] Because o1-mini focuses its capacity on reasoning rather than encyclopedic knowledge, OpenAI recommended it for tasks such as mathematics and code generation, while cautioning that it is weaker on problems requiring non-STEM factual knowledge.[1]
OpenAI reported that o1-mini was competitive with the full o1 model on mathematics and coding benchmarks despite being substantially smaller and cheaper, while significantly outperforming o1-preview on those tasks.[1] On the American Invitational Mathematics Examination (AIME), a high-school mathematics competition, o1-mini scored 70.0%, close to o1's 74.4% and far ahead of o1-preview's 44.6%.[1][6] On the Codeforces competitive programming platform, o1-mini reached an Elo rating of about 1650, near o1's 1673 and well above o1-preview's 1258; OpenAI stated this placed the model at roughly the 86th percentile of programmers competing on the platform.[1][6] OpenAI also reported strong results on the HumanEval coding benchmark and on high-school-level cybersecurity capture-the-flag (CTF) challenges, while noting that o1-mini lagged behind o1-preview on tasks requiring broad world knowledge.[1]
| Benchmark | o1-mini | o1 | o1-preview |
|---|---|---|---|
| AIME (math, % solved) | 70.0% | 74.4% | 44.6% |
| Codeforces (Elo) | ~1650 | 1673 | 1258 |
| Codeforces (percentile) | ~86th | n/a | n/a |
Source: OpenAI o1-mini announcement.[1][6] "n/a" indicates a figure not specified by OpenAI for that model in the cited material.
OpenAI cautioned that, on tasks measuring broad factual knowledge, o1-mini scored lower than larger models, consistent with its design as a STEM-focused system rather than a general-knowledge model.[1] Summarizing the trade-off, OpenAI wrote that o1-mini "excels at STEM, especially math and coding, nearly matching the performance of OpenAI o1 on evaluation benchmarks such as AIME and Codeforces."[1]
o1-mini and o1-preview were released together on September 12, 2024, but they target different use cases. o1-preview was the broader, more knowledgeable preview of the full o1 model, while o1-mini was the cost-efficient, STEM-focused variant. The headline difference was price: OpenAI priced o1-mini at roughly 80% less than o1-preview, while o1-mini actually outscored o1-preview on the math and coding benchmarks above.[1][2] The main cost o1-mini pays for that efficiency is reduced general world knowledge.[1]
| Attribute | o1-mini | o1-preview |
|---|---|---|
| Input price (per 1M tokens) | $3.00 | $15.00 |
| Output price (per 1M tokens) | $12.00 | $60.00 |
| AIME (% solved) | 70.0% | 44.6% |
| Codeforces (Elo) | ~1650 | 1258 |
| Strength | Math, coding, STEM | Broader world knowledge |
Source: OpenAI / contemporaneous pricing reports.[1][2][4]
In the API, OpenAI priced o1-mini at $3.00 per million input tokens and $12.00 per million output tokens, compared with $15.00 per million input tokens and $60.00 per million output tokens for o1-preview, a roughly 80% reduction in price.[2][4] Reasoning tokens were billed at the output-token rate even though they were not shown in the response.[2]
At launch on September 12, 2024, o1-mini was available in ChatGPT to Plus and Team subscribers, initially with a weekly message limit.[1][3] On September 17, 2024, OpenAI increased the o1-mini rate limit for Plus and Team users roughly sevenfold, from 50 messages per week to 50 messages per day.[7] OpenAI said at launch that it planned to bring o1-mini to ChatGPT free-tier users as well; access for free users followed later.[1] In the API, access was initially offered to developers on the highest usage tier (Tier 5) before being broadened.[1][3]
The principal limitation of o1-mini stems from its design: as a smaller model, it holds less general world knowledge than o1-preview or GPT-4o and performs worse on questions that depend on broad factual recall rather than STEM reasoning.[1] OpenAI framed it as a specialized tool for technical work rather than a general-purpose assistant.
As with the rest of the early o1 series, o1-mini at launch lacked several features available in mainstream chat models. The reasoning chain of thought is hidden from users, and the early o1 models did not initially support tools such as web browsing, file uploads, or image inputs that were available with other ChatGPT models at the time.[3] Reasoning models also tend to take longer to respond than conventional chat models because they generate intermediate reasoning before answering.[3]
OpenAI released o3-mini on January 31, 2025, as the next-generation small reasoning model, and it replaced o1-mini in the ChatGPT model picker.[8][9] OpenAI made o3-mini available to ChatGPT free-tier users as well as Plus, Team, and Pro subscribers, the first time a reasoning model was offered to free users, and described it as offering higher rate limits and lower latency than o1-mini while delivering strong performance in coding, STEM, and logical problem-solving.[8] OpenAI characterized o3-mini as providing high intelligence at the same cost and latency targets as o1-mini.[8]
OpenAI reported that, in evaluations by expert testers, o3-mini's responses were preferred over o1-mini's 56% of the time, with a 39% reduction in major errors on difficult real-world questions.[8] The rate limit for ChatGPT Plus and Team users rose from 50 messages per day with o1-mini to 150 messages per day with o3-mini.[8] Positioning the two models, OpenAI wrote that "while o1 remains our broader general-knowledge reasoning model, o3-mini provides a specialized alternative for technical domains requiring precision and speed."[8]
For the API, OpenAI announced on April 28, 2025, that it was deprecating both o1-preview and o1-mini, with removal scheduled three months later for o1-preview and six months later for o1-mini.[10] OpenAI listed an o1-mini API shutdown date of October 27, 2025, and recommended o4-mini as the successor model for developers.[10]
o1-mini should not be confused with the other members of the o1 family: o1 (the full model that succeeded the preview), o1-preview (the initial public preview released the same day as o1-mini), and o1-pro (a higher-compute variant offered to ChatGPT Pro subscribers).