What changes now that OpenAI's o1 reasoning model is in the API?

Yesterday, December 17, 2024, OpenAI made its full o1 model available in its API, the service programmers use to build OpenAI's models into their own apps. Until now developers could only use o1-preview, an early version with few of the features real products need. The new o1, which ChatGPT subscribers received two weeks ago, now comes with those features.

The same announcement bundled several other updates for builders: cheaper and easier voice apps with the Realtime API, a new way to fine-tune models, and official toolkits for two more programming languages.

What developers get with o1

o1 is a reasoning model: before answering, it works through a problem in steps and checks itself, which makes it stronger at math, code and multi-step logic but slower and more expensive than a regular chat model. In the API, it now supports:

Feature	What it lets a developer do
Function calling	Let o1 ask the app to run a task, such as looking up an order or a stock level
Structured Outputs	Get answers that always match a fixed data format (a JSON Schema), so software can read them reliably
Developer messages	Give o1 standing instructions on tone, style or rules, the way a system prompt works
Vision	Send images, such as a chart, a diagram or a photo of a part, for o1 to reason about
`reasoning_effort`	Choose how long the model thinks before replying, trading depth for speed and cost

According to OpenAI, requests now take 60% fewer reasoning tokens on average than with o1-preview. Tokens are the small pieces of text a model reads and writes; reasoning tokens are the ones it spends thinking before it answers, and you pay for them too.

The version released is named o1-2024-12-17. OpenAI describes it as a further-trained version of the o1 in ChatGPT, adjusted based on feedback, and says ChatGPT is moving to it as well.

How much better is it?

OpenAI published test results comparing the new o1 with o1-preview:

Test	What it measures	o1-preview	o1 (Dec 17)
AIME 2024	A difficult US high school math competition	42.0	79.2
MATH	Competition-style math problems	85.5	96.4
LiveCodeBench	Programming problems	52.3	76.6
SWE-bench Verified	Fixing real bugs in software projects	41.3	48.9
GPQA diamond	Graduate-level science questions	73.3	75.7

The jump in competition math is the largest. o1 can now also be scored on image tests, where o1-preview couldn't take part, and OpenAI says it clearly beats GPT-4o in its own tests of function calling and structured outputs. According to TechCrunch, OpenAI also says the new o1 refuses fewer reasonable requests by mistake.

Who can use it, and what it costs

Access starts small. o1 is rolling out to developers in usage tier 5, OpenAI's highest spending level. TechCrunch explains that this means having spent at least $1,000 with OpenAI and having had an account for more than 30 days after the first payment. OpenAI says it will open access to other tiers and raise usage limits gradually.

The price is high. TechCrunch puts it at $15 for roughly 750,000 words of input and $60 for roughly 750,000 words of output, six times the cost of GPT-4o. Because reasoning tokens count as output, a hard question can cost noticeably more than its short answer suggests.

The other announcements

Realtime API. This is OpenAI's tool for live voice conversations with a model. It now supports WebRTC, an open standard browsers and phones already use for calls, so a voice app can be connected with a few lines of code. GPT-4o audio gets 60% cheaper, the smaller GPT-4o mini joins at a fraction of that price, and sessions can last 30 minutes instead of 15.
Preference Fine-Tuning. Fine-tuning means further training a model on your own examples. The usual method shows the model ideal answers. The new one shows it pairs of answers, one preferred and one not, so it learns what "better" looks like. OpenAI says this suits subjective work such as summaries or creative writing.
Go and Java SDKs. SDKs are ready-made code libraries that make the API easier to use. These two join the existing ones for Python, Node.js and .NET, in beta.

Why it matters

Until now, a company that wanted o1's careful reasoning inside its product had to work around missing basics. With function calling, structured outputs and images, o1 can sit inside real workflows: reading an invoice image, checking it against a database through a function, and returning the result in a format the accounting system accepts. OpenAI says developers have already used o1-preview to build customer support, supply chain and financial forecasting tools.

The cost and tier limits mean most small developers will wait or test sparingly for now. The reasoning_effort setting matters here: it lets builders save deep thinking for the requests that need it.

Try it: Even without API access, you can practice the key skill. Take one task you might automate, such as sorting customer emails, and write three lines: the instructions you'd put in a developer message, the exact fields you want back (for example category, urgency, summary), and which requests deserve slow, careful reasoning versus a quick answer.

How should a developer start building with o1?

First, check which usage tier your OpenAI account is on. If you're not on tier 5, keep building with GPT-4o or o1-mini and design your app so the model name is easy to switch later.

When you do have access, start with a developer message that sets the rules and a structured output that defines the answer's shape:

Prompt

Developer message: You review expense claims for a small company. Follow the policy text provided. Reply only in the JSON format defined by the schema, with fields "decision" (approve, reject or ask), "reason" (one sentence) and "policy_section".

Test the same set of real examples at different reasoning_effort settings and compare accuracy, response time and cost. Use function calling for anything that needs fresh data instead of asking the model to guess. And keep a human check on decisions that affect money or people until your tests show the results are dependable.