Four times the tokens for an 18-point jump
8 Nov 2026 · News
Thinking Machines shipped a sharper Inkling. The same eval that raised its score also raised the token bill.
Inkling 2 Small scored 31 on Artificial Analysis's Intelligence Index, 18 points above Inkling Small. It also used about four times as many output tokens per Index task, so the average cost rose from $0.20 to $1.18, almost six times as much, even as the price per million tokens fell. A higher chart score can raise the bill for each scored task. By 31 March 2027, another major open-weights lab will publish its Index gain next to the output-token count or dollar-per-task cost for the same suite, not the score alone.
What 18 Index points cost
On 5 November, Thinking Machines Lab introduced Inkling 2 Small, a mixture-of-experts model with 594 billion total parameters and 17 billion active ones. It takes images as well as text, and it can spend more or less effort on a problem. For now you can use it on the company's Tinker platform with closed weights. The lab says it intends to release the full weights under an Apache 2.0 license this coming week, after outside safety testing. As of this morning, those weights are not on the company's Hugging Face page, which still lists the first Inkling and Inkling Small.
Artificial Analysis ran the model on its Intelligence Index and put the score at 31, up 18 points from Inkling Small's 13. The biggest advertised gains are agentic. AutomationBench-AA, a test of automated business workflows, rose from 5% to 46%. FrontierCode 2.0, a software engineering suite, rose from 5% to 25%. The same pass found roughly 150,000 output tokens per Index task, against about 37,000 for Inkling Small.
That is why the bill moved. Inkling 2 Small is priced at $0.30 per million input tokens and $1.00 per million output tokens, a bit cheaper per token than Inkling Small. Cost per Index task still jumped from $0.20 to $1.18. The tokens got cheaper and each scored task still cost more, because each task used many more tokens.
The model is not expensive next to peers with similar Index scores. Gemini 3.8 Flash, at its high setting, costs $7.77 per Index task on the same board, and Claude Sonnet 5.5 at medium effort costs $10.02. Inkling also splits those comparisons. It scored 25% on FrontierCode 2.0 against 11% for Claude Sonnet 5.5 at medium effort, and 13% on Terminal-Bench 4.0 against 30%. The chart improved. It did not improve evenly, and running the new model against its predecessor got more expensive, not less.
Thinking Machines ran its own published benches at maximum effort. The lab calls that setting 0.99, or xhigh. It trained the model to trade speed for thoroughness by sampling different effort levels. The number on the launch page is the highest-effort setting. Artificial Analysis independently measured the token count that comes with that kind of run.
Then and now
On 2 October we said that when a lab brags about a leaderboard, the habit worth stealing is asking what still fails on real work. On 14 October, Meta's Muse Wonder made the other half plain: the leaderboard said frontier, and the meter said stop. Wonder Full cost $4.52 per Index task. Users found the quota before the chart had time to settle.
Yesterday we wrote that companies budgeted for seats and got tokens. That was the enterprise forecast breaking when agents scaled spend with work. Today's story is the model-card version of the same meter. The eval that celebrates the 18-point jump also lists what the jump consumed: about four times the output tokens, almost six times the dollars per scored task.
Our call
By 31 March 2027, at least one other major open-weights lab, not Thinking Machines, will publish an Intelligence Index score, or an equivalent public suite score, next to the output-token count or dollar-per-task cost for that same suite, rather than the score alone. We will be wrong if no other lab does that by the date.