OpenAI rolls out API access for o1 to high-paying developers

• Some developers can now access the o1 model via OpenAI's application programming interface (API).

• OpenAI will eventually expand o1 access to lower-tier users.

• OpenAI also announced WebRTC support for its Realtime API.

Starting today, OpenAI is giving high-spend API users access to its "o1" reasoning model.

Developers paying the firm at least $1,000 per month through the company's API will be able to wrap their own products and services around the advanced model.

o1 spends more time processing answers and mulling over its responses step-by-step, so it can spot errors and make improvements before responding. This result is more accurate answers and better solutions to more complex problems.

It offers much better results for math, coding, and PhD-level scientific queries than its predecessors (and many of its rivals), according to standard industry benchmarking.

The "reasoning" method proved painfully slow in its earlier phases, with a preview version released this fall taking over a minute to answer some questions. But the full version of o1 gave lightning fast answers when it was launched via the ChatGPT platform earlier this month, and benchmarking shows it's also smarter than the preview:

The latest version of o1

OpenAI also announced that developers at the highest usage tier will get access to an updated version of the model, released today under the name o1-2024-12-17. It's also been added to the ChatGPT platform, and will eventually be expanded to API users at lower tiers.

o1-2024-12-17 offers many of the features developers have come to see as standard, including vision capabilities, function calling and structured outputs, whereby a user can make sure text responses follow a specific format.

Benchmarking shows it's better than OpenAI's most advanced non-reasoning model, GPT-4o, at some of these capabilities:

The model's reasoning capabilities can also be tailored to a user's needs with a new reasoning_effort parameter. This allows developers to choose how long the model spends "thinking" before it returns a response.

OpenAI has bumped up o1's context window and maximum output token limits considerably to 200,000 and 100,000 tokens, respectively. A "context window" essentially dictates how much information a large language model can consider while coming up with a response.

The preview version of o1 had a context window limit of 128,000 tokens and a maximum output limit of 32,768 tokens, while a lighter version (o1-mini) had limits of 128,000 and 65,536 tokens, respectively.

For context, 128,000 tokens is roughly the length of a 300-page book.

Realtime API updates

If that wasn't enough to satisfy developers, the firm also announced updates to its AI voice generation platform Realtime API.

The product now supports popular open source real-time communication software WebRTC, meaning software developers can weave OpenAI's Realtime into their web and mobile applications with "just a few lines of code."

async function createRealtimeSession(inStream, outEl, token) {
  const pc = new RTCPeerConnection();
  pc.ontrack = e => outEl.srcObject = e.streams[0];
  pc.addTrack(inStream.getTracks()[0]);

  const offer = await pc.createOffer();
  await pc.setLocalDescription(offer);

  const headers = { Authorization: `Bearer ${token}`, 'Content-Type': 'application/sdp' };
  const opts = { method: 'POST', body: offer.sdp, headers };
  const resp = await fetch('https://api.openai.com/v1/realtime', opts);
  await pc.setRemoteDescription({ type: 'answer', sdp: await resp.text() });

  return pc;
}

OpenAI has slashed Realtime's costs by 60%, added support for LLM GPT-4o-mini and improved the product's quality.

12 Days of "Shipmas"

The o1 model released via ChatGPT earlier this month was the first of 12 new features and updates OpenAI promised to "ship" in the days before Christmas.

The company nicknamed Tuesday, which was day #9, the "DevDay Holiday Edition" because of its focus on developers.

Other days have seen the release of a $200 subscription tier, and the much-anticipated video generation model Sora.

But it hasn't been plain sailing for the company. Users have reported some service delays and outages as OpenAI has rushed to get its developments out.

With three days to go, there's still plenty of opportunity for other big releases. Could GPT-4.5 be on the way? Or perhaps a long-rumored autonomous AI agent with the codename Operator? Only time will tell.