Thinking Machines’ Tinker Adds OpenAI-Compatible Inference

Thinking Machines, the AI lab led by Mira Murati, said on 15 October 2026 that customers can now use Tinker, its model-training service, for “production inference via sampling API that’s compatible with the interfaces of OpenAI and Anthropic (in beta).”

In the same post, marking Tinker’s first year, the company said it has begun offering full fine-tuning as a beta feature, has added over 20 models since launch, and passes efficiency gains on to customers with price cuts.

Business Pill · OPEN SOURCE BUSINESS MODELS

An explainer on how open-source business models work. It teaches the general idea only and says nothing about any organisation in this story.

What Thinking Machines Announced

Tinker is an API for post-training: customers fine-tune open-weight models on their own data while Thinking Machines runs the GPUs. “Every checkpoint saved on Tinker is a model you can sample from,” the post says.

The inference option is the new step. The company also says it is working on “a more fully featured offering with higher throughput and interactivity,” with a waitlist for interested customers.

On training, “we have begun offering full fine-tuning as a beta feature and are considering a broader release,” the post says. Until now the company has promoted LoRA, a lighter method that trains a small set of added weights.

For Tinker’s second year, Thinking Machines says it plans to make it “a platform for both training and serving custom models at commercial scale.”

What Was Posted on X

Mira Murati, who leads Thinking Machines, shared the One Year of Tinker post on X on 15 October 2026. Her comment is about users training models whose weights they own; the product details in this piece come from the company’s post.

Models and Prices

“We launched with 14 models from Qwen and Llama, all of them text-only,” the company writes. Since then it has added over 20 models from Qwen, Nemotron, Kimi, GLM, DeepSeek, GPT-OSS and its own Inkling, it says.

The lineup now runs from 4B to 1T parameters, according to the post. Most models support image input, and Inkling, released in July, added audio, it says.

The company says efficiency work lets it run Tinker more cheaply, and that it passes these savings on to customers with price cuts. The post does not list the new prices.

The Results It Cites

The post cites customer results against frontier models. Bridgewater’s text-to-SQL model, it says, “beat Fable 5 and GPT-5.6 Sol on accuracy at 15% of their cost.”

That result first appeared in a Thinking Machines post of 27 August 2026 by researchers from UIUC and Bridgewater AIA Labs. That post put the cost at 12% to 15% of Fable 5 and GPT-5.6 Sol Ultra, at $0.56 per task.

Sundial’s model, the post says, “beat every frontier model but GPT-6 Astra on the share of LaTeX errors fixed, at 7% of Astra’s cost.”

Glean’s custom search agent “is twice as fast and uses 25% fewer tokens with no quality regression,” according to the post. ReViSQL-K2.6, trained by UIUC and Bridgewater, reaches 92.97% accuracy on a clean text-to-SQL benchmark when picking from 16 samples, it says.

Enterprise Terms

Thinking Machines says enterprise demand shifted its focus to speed, security and inference. It lists single sign-on, role-based access control, an audit-log API, encryption and private VPC connectivity.

“All customer content on Tinker belongs to the customer, and we never train on it,” the post says.

Spending on training, sampling and storage can now be broken down by session, and users can monitor and cancel sessions from the console, the company says.

What Murati Said

Murati shared the post on X on 15 October. “One of the best parts of this year, inside Thinky and out, has been watching people follow their own ideas all the way into weights they own,” she wrote.

The Structural Read

As we read it, compatibility with the OpenAI and Anthropic interfaces lowers the cost of switching an application from a general model to a custom one, because the calling code can stay the same. The post does not say how many customers use the inference beta.

As we read it, the company is moving along the chain from training to serving: the second-year plan in the post is a platform “for both training and serving custom models at commercial scale.”

As we read it, the case Thinking Machines makes is cost per task, not leaderboard rank: its cited results put custom models at 15% and 7% of the frontier models’ cost, as the company reports them.

Three Implications

TEAMS WITH IN-HOUSE DATA Thinking Machines says customers’ content belongs to them and it never trains on it, and that a custom model can be served through an OpenAI- and Anthropic-compatible API, in beta.

TEAMS USING LORA TODAY The post says full fine-tuning is now offered as a beta feature, and that the company is considering a broader release.

BUDGET OWNERS The company says it passes efficiency gains on to customers with price cuts, and spending can now be broken down by session; the post does not give the new prices.

The Business Engineer Lens

This story maps onto the Business Engineer framework The Open vs Closed Meta-Framework.

The framework’s rule is: “Close the SCARCE layer (keep proprietary) + Open the ABUNDANT layer (commoditize).”

As we read it, Tinker is built on the open layer, open-weight models from Qwen, Llama, Kimi, DeepSeek and others, while the customer keeps the scarce layer, its own data and the weights it trains. The framework predates the post and does not discuss Thinking Machines.

What Is Not Established

In what we read, the post gives no count of Tinker users or customers and no revenue figure. It does not list the new prices or say when production inference leaves beta.

The customer results are as Thinking Machines reports them. The Bridgewater figure was first published in August, and the August post compared against GPT-5.6 Sol Ultra, while the new post names GPT-5.6 Sol.

The Bottom Line

A year after launch, Thinking Machines says Tinker customers can serve their trained models through an inference API compatible with OpenAI and Anthropic, in beta, and can try full fine-tuning, also in beta.

The company says it has added over 20 models to the 14 it launched with, cut prices, and plans to make Tinker a platform for training and serving custom models at commercial scale.