ModelsOpenAI o1 in 2026: Should You Still Be Calling It?
o1 was OpenAI's first full reasoning model in the API. What it costs, how its hidden reasoning tokens bill, and how to decide whether to keep it or migrate off.
Plenty of production code still sends requests to o1. It was wired in during early 2025, it worked, and nobody has had a reason to touch it since. The question worth asking in 2026 is not whether o1 is good. It is whether you are paying a reasoning-model premium for a model that newer, cheaper ones have overtaken on your actual workload.
This guide covers what o1 is, how it bills, and a practical way to decide whether to keep it.
Where o1 came from
o1 was OpenAI's first model trained with reinforcement learning to reason before answering. o1-preview and o1-mini reached ChatGPT on 12 September 2024. The full o1 followed in ChatGPT on 5 December 2024, and the API snapshot is o1-2024-12-17, which is still the only snapshot OpenAI lists.
The launch claims were about maths, competitive programming and science questions. o1-preview solved 83% of AIME problems against 13% for GPT-4o, and placed in the 89th percentile on Codeforces. Those figures belong to the preview model and to 2024; they establish why the family mattered, not how o1 compares with anything released since.
The successor line is o3. OpenAI skipped the name o2 entirely. If you are looking at o1 today, the comparison you actually need is against the o3 guide and the smaller o-series models that followed.
The spec sheet
From OpenAI's model page:
- Context window — 200,000 tokens, with up to 100,000 output tokens.
- Input — text and images. Output is text only.
- Knowledge cutoff — October 2023.
- Endpoints — Chat Completions, Responses, Assistants and Batch.
- Features — streaming, structured outputs, function calling, file search and image input.
The endpoint list is one of o1's quiet advantages. It works on Chat Completions, so existing integrations built for GPT-4-era models can call it without moving to the Responses API. Its more expensive sibling, o1-pro, cannot say the same.
The knowledge cutoff is the quiet disadvantage. October 2023 is a long time ago for anything involving libraries, APIs or current events. A reasoning model reasons well over what you give it; it cannot reason its way to an API that did not exist when it was trained.
How the bill actually works
OpenAI lists o1 at $15 per million input tokens, $7.50 for cached input and $60 per million output tokens. Through the Batch API the rates halve to $7.50 and $30. OpenRouter lists the same standard rates at the time of writing.
The output figure is the one that surprises people, because reasoning models generate tokens you never see. OpenAI's reasoning guide is explicit: reasoning tokens are not visible through the API, they occupy context, and they are billed as output tokens. A request that returns a three-line answer can bill thousands of output tokens of hidden thinking.
Two practical consequences:
- Read the usage object. The reasoning token count is reported separately from visible output, and it is the only way to know what a request really cost.
- Do not set the output limit tight. The cap covers reasoning plus visible output. If reasoning exhausts it, you get an incomplete response, sometimes with no visible text at all. OpenAI recommends reserving at least 25,000 tokens for reasoning and output when you start experimenting.
The token cost of reasoning models goes further into why hidden output makes these models hard to budget.
What o1 is still reasonable for
If o1 is already integrated and passing your checks, there are cases where leaving it alone is defensible:
- Low-volume, high-stakes analysis where you validated its output carefully and a change would mean revalidating.
- Workflows on Chat Completions or Assistants that you are not ready to port.
- Image-plus-text reasoning tasks you have already tuned prompts for.
Outside those, o1 is hard to justify for new work. It is expensive per output token, it has an old knowledge cutoff, and OpenAI's own comparison table on the o1-pro page lists o3-mini at $1.10 input and $4.40 output, more than ten times cheaper than o1 in both directions.
o1 or o1-pro
o1-pro is the same model family given more compute per request, at ten times the price: $150 input and $600 output per million tokens. It is also Responses-only. If o1 is getting a hard problem nearly right, o1-pro may get it right; if o1 is getting it wrong in a consistent way, more compute rarely fixes that. The o1-pro guide covers when the premium is worth paying.
How to decide whether to migrate
Treat this as a regression test, not a judgement call.
- Pull thirty real requests that o1 currently handles, including the ones that matter most.
- Run them against one or two newer reasoning models with the same prompts and tools.
- Score the outputs blind, against the criteria your users care about.
- Record total output tokens per request, reasoning included, and compute cost per acceptable answer.
If a newer model matches o1 on your set at a fraction of the cost, migrate and pin the new snapshot. If it does not, you have a documented reason to stay. Regression testing a model upgrade covers how to build the set so the result means something.
Common questions
How much does OpenAI o1 cost?
OpenAI lists o1 at $15 per million input tokens, $7.50 for cached input and $60 per million output tokens, with Batch API rates of $7.50 and $30. Hidden reasoning tokens are billed as output, so real costs exceed what the visible answer suggests.
What is the context window of o1?
o1 has a 200,000-token context window and can produce up to 100,000 output tokens, which includes its hidden reasoning tokens. Its knowledge cutoff is October 2023.
Can I call o1 through the Chat Completions API?
Yes. OpenAI lists Chat Completions, Responses, Assistants and Batch as supported endpoints for o1, with streaming, function calling and structured outputs. o1-pro, by contrast, is available only through the Responses and Batch endpoints.
