OpenAI's reinforcement-learning reasoning model for multi-step math, science, and coding problems needing deliberate step-by-step thinking.
Output Speed *
Intelligence Index *
Context Window *
Input price
Output price
Performance
Where o1's Test-Time Reasoning Actually Pays Off
Benchmarks
o1 on Independent Reasoning and Science Benchmarks
Output Speed
Intelligence Index
MMLU *
GPQA *
HLE *
LiveCodeBench *
Technical Specifications
What the model supports
Quickstart
import requestsurl = "https://api.anyapi.ai/v1/chat/completions"
payload = {
"model": "o1",
"messages": [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Text prompt"
},
{
"image_url": {
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
]
}
]
}
headers = {
"Authorization": "Bearer AnyAPI_API_KEY",
"Content-Type": "application/json"
}
response = requests.post(url, json=payload, headers=headers)
print(response.json())
const url = 'https://api.anyapi.ai/v1/chat/completions';
const options = {
method: 'POST',
headers: {
Authorization: 'Bearer AnyAPI_API_KEY',
'Content-Type': 'application/json'
},
body: '{
"model":"o1",
"messages": [
{
"role":"user",
"content":[
{
"type":"text",
"text":"Text prompt"
},
{
"image_url":{
"url":"https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type":"image_url"
}
]
}
]
}'
};
try {
const response = await fetch(url, options);
const data = await response.json();
console.log(data);}
catch (error) {
console.error(error);
}curl --request POST \
--url https://api.anyapi.ai/v1/chat/completions \
--header 'Authorization: Bearer AnyAPI_API_KEY' \
--header 'Content-Type: application/json' \
--data '{
"model": "o1",
"messages": [
{
"role": "user",
"content": [
{
"type": "text",
"text": "Text prompt"
},
{
"image_url": {
"url": "https://upload.wikimedia.org/wikipedia/commons/thumb/d/dd/Gfp-wisconsin-madison-the-nature-boardwalk.jpg/2560px-Gfp-wisconsin-madison-the-nature-boardwalk.jpg"
},
"type": "image_url"
}
]
}
]
}'
Comparison
o1 vs GPT-4o: Reasoning Depth or Speed and Cost?
o1 and GPT-4o are both OpenAI flagship-class models, but they solve different problems. GPT-4o is a fast, general-purpose multimodal model tuned for low latency and high throughput. o1 is a reasoning model that trades speed for deliberate chain-of-thought accuracy. The gap is stark on hard reasoning: OpenAI reports o1 solving around 74% of 2024 AIME problems versus roughly 12% for GPT-4o. The practical decision is whether your workload needs step-by-step correctness on math, science, and complex code, or fast, cost-efficient responses at scale.
Choose o1 when accuracy on multi-step reasoning, competition math, graduate-level science, or algorithmically hard coding justifies higher latency and reasoning-token cost, and when responses can take seconds rather than being instant. Choose GPT-4o when you need low latency, high-volume serving, cheaper output, audio input/output, or general conversational and multimodal tasks where deep reasoning is unnecessary. Many production systems route only hard subtasks to o1 while GPT-4o handles the routine, latency-sensitive majority.
Limitations & Trade-offs
Best-Fit Workloads
Where this model earns its place
Competition-Grade Math and Quantitative Reasoning
o1 excels at problems that reward stepwise deduction. OpenAI reports it averaged 74% on the 2024 AIME with a single sample, far above GPT-4o. This makes it well suited to quantitative analysis, algorithmic derivations, and financial or scientific modeling where a wrong intermediate step invalidates the answer. Because each request is slow and reasoning-heavy, run these as asynchronous jobs rather than interactive endpoints.
Graduate-Level Scientific Reasoning
o1 exceeds human PhD-level accuracy on the GPQA benchmark across physics, biology, and chemistry, and its image input lets it reason over diagrams and scientific figures. This suits research assistance, technical literature analysis, and complex domain question answering. Note the October 2023 knowledge cutoff: pair o1 with retrieval for current findings, and verify outputs in high-stakes scientific contexts.
Complex, Well-Specified Coding Tasks
o1 ranks near the 89th percentile on Codeforces and produces strong results on multi-step implementation tasks such as planning a project structure then generating full files. Function calling and structured outputs let it slot into agentic developer tooling. It performs best when requirements are explicit; ambiguous or heavily format-constrained tasks can produce inconsistent results, so keep specifications tight.
Reasoning Steps Inside Agentic Pipelines
With function calling and structured outputs, o1 fits as the deliberate reasoning node in an agent that also uses faster models for routine steps. Developers have used the o1 line for agentic applications in customer support, supply chain, and financial forecasting. Route only hard planning or verification subtasks to o1 to contain latency and cost while preserving accuracy where it matters.