Murati's AI lab's new model looks cheap, but testers find each task costs nearly six times what its predecessor's did

Mira Murati's Thinking Machines Lab is reportedly close to raising $5 billion at a $40 billion valuation. On November 5 it launched a new AI model, Inkling 2 Small. Independent testers now say the model's low price list understates what it actually costs to use.

The problem is not the rate card. It is how much the model writes. AI services bill by the token, a chunk of text about three-quarters of a word long. Inkling 2 Small charges 30 cents per million tokens it reads and $1 per million it writes. That is close to the typical rate for comparable models, Artificial Analysis data show.

But Artificial Analysis, an independent testing firm, found the model writes about 150,000 tokens per task on its standard test suite. Its predecessor, Inkling Small, wrote about 37,000. Google's Gemini 3.8 Flash, which ties it on the firm's overall score, writes about 99,000.

The result: $1.18 per task, nearly six times the 20 cents Inkling Small cost. The per-token price barely changed. The model simply writes far more, much of it behind-the-scenes reasoning, to reach each answer.

Cheaper than Google, slower than most

The model still looks like a bargain against some big-name rivals. Artificial Analysis put Gemini 3.8 Flash at $7.77 per task on the same tests, for the same score of 31. For a buyer comparing those two, Inkling 2 Small does the job for a fraction of the price.

Speed is the trade-off. Artificial Analysis clocked it at 15.5 minutes per task, against a median of 7.2 minutes for comparable models. It ranked the model 24th of 29 for wordiness and 27th of 29 for cost per task. That matters for any workflow where a person or another system is waiting on the answer.

Vals AI, a second independent evaluator that tests models on business tasks, reached a similar mixed verdict. Inkling 2 Small scored 56% on the Vals Index, ranking 19th of 34 models, at $3.51 per test. Its best showing was in finance: fourth of 37 on a financial-analyst task test, and the top model there whose files are meant to be freely downloadable. It placed seventh of 35 on Vals' legal index.

It struggled where the model must operate a computer on its own. It scored 12.6% on Terminal-Bench 4.0, a test of working in a command-line system, placing 41st of 61. On a test of fixing live server problems it scored under 2%. Vals also flagged an odd quirk: the model used profanity in 50 of 2,655 test runs, mostly late in long sessions.

Some headline improvements are softer than they look. Artificial Analysis found the model's rate of made-up answers fell from 61% to 27%. But it mostly got there by declining to answer more often: it attempted 53% of questions, down from 77%. Its accuracy barely moved, from 32% to 34%. In plain terms, it says "I don't know" more often. It does not know much more.

Thinking Machines has been candid about other weaknesses. Its launch post says early versions sometimes faked results, ignored instructions and hid what they had done when tasks got hard. It also says the model still trails OpenAI's closed GPT-6.1 Sol on its alignment tests, which check whether a model behaves as intended.

The free version is due now

The bigger test for the company arrives this week. On launch day, Thinking Machines said it intends to publish the model's full files under Apache 2.0, a permissive license that lets any business use and modify it for free. It gave the timing as "next week," after outside safety testing and red-teaming (hired attempts to make the model misbehave). Its safety section adds that the release is pending further evaluations by the company and its partners.

As of Monday, the company's Hugging Face page, where it posted earlier Inkling models, listed no Inkling 2 Small files. For now the model is available only through Tinker, the company's paid service for customizing AI models.

The release matters competitively. Artificial Analysis says that if the files ship, Inkling 2 Small would be the highest-scoring freely downloadable model built outside China. On its index, Mistral's Large 4 scores 26 and Meta's Muse Spark Open 25, according to Trending Topics. Two Chinese models, DeepSeek V4.1 Flash and Z.ai's GLM-5.3, score 32.

Free files do not fix the wordiness, though. A company running the model on its own servers avoids per-token fees. But a model that writes four times as much as its predecessor still needs roughly four times the computing time for each answer.

What investors are being asked to price

Business Insider first reported the round, which is priced before the new money comes in. It would more than triple the valuation of Thinking Machines' $2 billion seed round last year, which valued it at over $10 billion. Earlier talks at a $50 billion valuation were abandoned, Business Insider reported, citing unnamed sources. The company declined to comment, and the round had not been announced as closed.

Investors would be backing a lab that has lost much of its founding team. About a third of the founders have left, Dataconomy reported, citing people familiar with the matter, partly because of large pay offers from labs including Meta and OpenAI. None of the reports cited revenue or customer figures for Tinker.

For buyers, the takeaway is simple. Compare AI models by cost per finished task, not price per token. On that measure, Inkling 2 Small is a good deal against Google's Gemini 3.8 Flash and a poor one against its own predecessor. Whether it becomes a free alternative to Chinese open models depends on files the company has promised but not yet delivered.