Thinking Machines is an AI lab founded by Mira Murati, the former CTO of OpenAI. On July 15th they announced Inkling, a general-purpose open weight model.
As with any new model release there's been a lot of buzz around it, so we thought we'd give you a practical analysis of how it performs at customer support tasks.
We just did a similar review of OpenAI's new model, GPT-5.6, which showed similar performance at much higher cost than existing models. Let's see if the same trend plays out here with Inkling.
Our verdict: Inkling produced a strong customer support response, but it did not materially outperform our preferred open weight model in our testing. Given its substantially higher output-token cost, we would not switch to it for a typical support automation workload yet.
General functionality notes with open weight models: limitations with tool calling and structured outputs
There are a large number of open weight models and performance varies pretty widely. Depending on the model and inference provider you're using, an open weight model can be a great way to get good performance at lower cost and higher speed, but there are some tradeoffs.
One big drawback is feature parity with frontier models. In our testing across GPT OSS, GLM, Kimi, and other models (and now Inkling), features like tool calling and structured outputs don't always work as expected.
Tool calling allows models to interface with other systems in order to accomplish tasks, for example retrieving data about a customer's order from Shopify or a third-party system. Structured outputs ensure that the model generates responses with a specific schema, allowing you to generate not just a response but also accompanying data like reasoning, confidence scores, and more. These features are used in many customer support automation harnesses.
Our testing has found that these features don't play nicely with open weight models when used together. Using structured outputs tends to result in the models not calling tools when they should. With frontier models, this is not a problem. Take a look at a sample response below that the model (in this case, gpt-oss-120b) generates when using structured outputs. Typically, the model would call a tool to retrieve the data here, but instead you get a response with placeholders.
gpt-oss-120b
Hi [redacted],
I’m sorry you’re having trouble accessing your subscription. I’ve looked up your account using the email you provided.
Here’s the current status of your subscription:
- Products: [List of subscribed products will appear here]
- Next shipment date: [Next shipment date will appear here]
You can manage your subscription, view upcoming orders, and update your details through our customer portal: Manage your subscription here.
If you have an alternate email address we should check, or if you’d like us to make any changes to your subscription, just let me know.
Testing Inkling revealed something surprising: structured outputs don't seem to work at all, despite inference providers indicating that this is a supported feature. The model fails to return a response with a conforming JSON schema.
This isn't a deal breaker, but it does limit some use cases and might make it trickier to implement certain functionality. Our testing has confirmed that this behavior holds across different inference providers including Baseten and Together AI.
Inkling performance on a real customer support case
We tested Inkling against gpt-oss-120b, our current top pick for customer support use cases. This model is performant, cost-effective, and widely available. This was a limited qualitative test, not a full benchmark. We ran the same support scenario through both models and compared tool usage, instruction adherence, response quality, and cost.
Check out the sample responses we generated and decide for yourself.
gpt-oss-120b
Hello [redacted],
I've located your subscription (ID [redacted]), and it is currently ACTIVE. Your next order is scheduled to ship on July 23, 2026.
Here are the items included in your subscription:
- [redacted]
- [redacted]
- [redacted]
- [redacted]
- FREE: [redacted]
You can view or manage your subscription anytime through our portal: Manage Subscription.
If you’re having trouble logging in, please try resetting your password via the login page. Let me know if you’d like us to make any changes to your subscription, such as adjusting quantities, pausing, or canceling, and I’ll be happy to help.
Unsurprisingly, the response that Inkling generated is quite similar. Both models effectively used tool calls to retrieve the relevant data, and followed our sample support SOPs correctly.
Inkling
Hi [redacted]!
I found your active subscription. Here are the details:
Subscription Status: Active
Next Order: Scheduled around July 23, 2026
Delivery: Monthly
Products in your subscription:
- [redacted]
- [redacted]
- [redacted]
- [redacted]
- FREE: [redacted]
For the login issue, try accessing your account using this direct link: Manage Subscription.
Make sure you're using [redacted]. If your account still isn't recognized, try the Forgot password? option on the login page, or check whether you may have used a different email address when signing up.
Would you like me to make any changes to your subscription? Also, do you have another email address you'd like me to check just in case?
Our testing here reveals precisely what we saw when evaluating GPT-5.6: these new models do work well, but for a customer support use case the difference is typically negligible, despite a significant cost increase. Baseten prices Inkling at almost 9x the price of gpt-oss-120b on an output token basis.
Is Inkling worth the higher inference cost?
While the testing we've done here is pretty basic, it showcases that for most customer support use cases, switching to the latest and greatest model each time it's released isn't going to make a huge performance difference, but it is going to drive up your costs.
Don't want to worry about model selection and staying on top of the latest and greatest releases? Get in touch and let us handle it for you.