OpenAI o1-pro
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,376 words
Improve this article
Add missing citations, update stale details, or suggest a clearer explanation.
Suggest edit
CtrlK
Last reviewed
Jun 3, 2026
Sources
8 citations
Review status
Source-backed
Revision
v1 · 1,376 words
Add missing citations, update stale details, or suggest a clearer explanation.
OpenAI o1-pro is the highest-compute variant of OpenAI's o1 reasoning model. It exists in two closely related forms: o1 pro mode, a feature of the ChatGPT Pro subscription that OpenAI introduced on December 5, 2024, and o1-pro, a model offered through the OpenAI developer API beginning March 19, 2025. Both spend substantially more inference-time compute than standard o1 in order to "think harder" and return more reliable answers on the hardest mathematics, science, and coding problems. [1][2][3]
OpenAI's o-series models are trained with reinforcement learning to produce a long internal chain of thought before answering. A central finding behind the series is that accuracy on hard reasoning problems improves not only with more training compute but also with more compute spent at inference time, when the model is allowed to "think" longer before responding. Within that family, o1-pro represents the high-compute end: it allocates more test-time computation to a single query than o1, which OpenAI says yields more accurate and more consistent responses on difficult tasks at the cost of higher latency and price. [1][3]
It is important to distinguish the variants. o1 pro mode is a user-facing feature inside ChatGPT, available only to ChatGPT Pro subscribers. o1-pro is the corresponding model exposed to developers through the API. Both differ from standard o1, the full reasoning model released the same day to lower ChatGPT tiers and to the API, and from o1-mini, a smaller, faster, lower-cost reasoning model. OpenAI reported that the full o1 itself makes 34% fewer major errors on difficult real-world questions than the earlier o1-preview while also responding faster. [1][4]
OpenAI unveiled o1 pro mode on December 5, 2024, as part of Day 1 of its "12 Days of OpenAI" announcement series, simultaneously launching the ChatGPT Pro plan at $200 per month. ChatGPT Pro provides what OpenAI described as scaled access to its best models and tools, including unlimited use of o1, o1-mini, GPT-4o, and Advanced Voice, plus exclusive access to o1 pro mode. [1][2][4]
OpenAI described o1 pro mode as "a version of o1 that uses more compute to think harder and provide even better answers to the hardest problems." In evaluations by external expert testers, OpenAI said the mode produced more reliably accurate and comprehensive responses, particularly in data science, programming, and case law analysis. Because answers can take longer to generate, the ChatGPT interface shows a progress indicator while the model reasons, and OpenAI noted that users can keep working in other conversations while waiting. [1][2]
Alongside the launch, OpenAI announced a ChatGPT Pro Grant program, initially funding ten grants of ChatGPT Pro access to medical researchers at leading U.S. institutions. [1]
On March 19, 2025, OpenAI released o1-pro to developers through its API. An OpenAI spokesperson described it as "a version of o1 that uses more computing to think harder and provide even better answers to the hardest problems." Access was initially limited to selected developers: at launch, those who had spent at least $5 on OpenAI API services, and OpenAI later made it available to usage tiers 1 through 5. [3][5]
A defining technical constraint is that o1-pro is offered only through OpenAI's newer Responses API and is not available on the older Chat Completions endpoint, which required existing integrations to be reworked. OpenAI stated that o1-pro was made available in the Responses API in order to support multi-turn model interactions before responding to a request and to enable other advanced API features in the future. The Responses API uses an item-based request and response format, in which messages, function calls, and function-call outputs are represented as discrete items rather than a flat list of chat messages. [5][6]
According to OpenAI's model documentation, o1-pro has a 200,000-token context window, a maximum of 100,000 output tokens, and a knowledge cutoff of October 1, 2023. It accepts text and image input and produces text output. It supports function calling, Structured Outputs, and a configurable reasoning-effort setting, and it generates internal reasoning tokens that are billed as output. It does not support streaming responses and is not available for fine-tuning. [5]
For the December 2024 launch, OpenAI published benchmark results comparing o1-preview, o1, and o1 pro mode. On a standard single-attempt (pass@1) basis, o1 pro mode improved on o1 in competition mathematics and matched or exceeded it elsewhere, though the gains over o1 were modest on coding and science. [1][7]
| Benchmark (pass@1) | o1-preview | o1 | o1 pro mode |
|---|---|---|---|
| AIME 2024 (competition math) | 50% | 78% | 86% |
| GPQA Diamond (PhD-level science) | 74% | 75% | 79% |
| Codeforces (competition code, percentile) | 62% | 89% | 90% |
To highlight consistency rather than single-shot luck, OpenAI also reported results under a stricter "4/4 reliability" metric, in which a model is counted as solving a question only if it answers correctly on all four of four attempts. Under this measure the advantage of o1 pro mode over o1 widened across all three areas. [1][7][8]
| Benchmark (4/4 reliability) | o1-preview | o1 | o1 pro mode |
|---|---|---|---|
| AIME 2024 | 37% | 67% | 80% |
| GPQA Diamond | 58% | 67% | 74% |
| Codeforces (percentile) | 26% | 64% | 75% |
Independent and OpenAI-internal assessments were more mixed. OpenAI's own benchmarks reportedly showed o1-pro performing only marginally better than standard o1 on math and coding tasks, while still delivering more reliable answers, and early testers found it could still stumble on puzzles such as Sudoku and visual jokes. On the third-party Artificial Analysis index, o1-pro's measured intelligence score was not markedly higher than less expensive models, prompting commentary that itsvalue lay in reliability on hard problems rather than across-the-board capability. [3][8]
o1 pro mode is included in the $200-per-month ChatGPT Pro subscription and is not metered per token for end users. The API model, o1-pro, was priced as OpenAI's most expensive model at its launch. [1][3]
| Item | Price |
|---|---|
| ChatGPT Pro (includes o1 pro mode) | $200 per month |
| o1-pro API input | $150 per 1M tokens |
| o1-pro API output | $600 per 1M tokens |
At those rates, o1-pro's input cost was twice that of GPT-4.5 and roughly ten times the price of standard o1, whose API pricing was $15 per 1M input tokens and $60 per 1M output tokens. Reasoning tokens generated internally by the model are billed at the output rate, so practical per-query costs could be high; one early test reported costs near 89 to 94 cents per query. [3][8]
The launch drew wide attention as one of the first mainstream $200-per-month consumer AI subscriptions. Coverage from outlets including TechCrunch, Axios, and VentureBeat framed ChatGPT Pro and o1 pro mode as products aimed at researchers, engineers, and other power users who routinely tackle the hardest problems, where greater reliability could justify the price. [2][3][4]
When o1-pro reached the API in March 2025, reporting focused heavily on its cost. TechCrunch described it as OpenAI's "most expensive AI model yet," and developer commentary noted that the high price, combined with the Responses-API-only requirement and only incremental benchmark gains over o1, made it a niche option to reserve for problems where standard reasoning models fall short. By that point OpenAI had also released newer reasoning models in the o-series, and o1-pro was positioned as a specialized high-reliability tool rather than a general default. [3][5][8]