Inkling 2 Small charges $0.30 in and $1 out, and a task costs 5.8x what its predecessor's did. It rereads 29 million cached tokens per task, so Sonnet 5.5 matches its score for a third less.
Thinking Machines' new small model held or cut every line of its predecessor's serverless rate card and got much dearer per finished task. The output price has very little to do with why.
Image source: Thinking Machines Lab
Fifteen cents out of $1.18
Inkling 2 Small costs $0.30 in, $0.03 cached and $1 out. Artificial Analysis puts a task at $1.18 and only 15 cents of that is output. The model rereads about 29 million cached tokens per task, and those reads are 74% of the bill.
At its score of 30.6, Sonnet 5.5 on medium costs 34% less and Haiku 5.5 on high costs a fraction of that. The reason to care about this model is the Apache 2.0 weights Thinking Machines has promised, and those aren't out yet.
What it costs on Tinker
Inkling 2 Small launched on November 5 on Tinker, Thinking Machines' own platform, and nowhere else. The pricing page lists it next to its predecessor. Note the two different Inkling-Small cards: the serverless beta is what Artificial Analysis uses, and the higher one is what Tinker charges for sampling.
| Per 1M tokens | Input | Cached | Output | Train, 64K / 256K |
|---|---|---|---|---|
| Inkling 2 Small, sampling and serverless | $0.30 | $0.03 | $1.00 | $1.50 / $3.00 |
| Inkling-Small, serverless | $0.30 | $0.06 | $1.20 | |
| Inkling-Small, sampling | $0.58 | $0.116 | $1.44 | $1.73 / $3.47 |
Read November 7, 2026. Tinker labels the table "as of October 19, 2026" even though the new model is on it.
So the serverless card holds input at $0.30, halves the cache rate and takes 17% off output. Two limits come with it. Serverless is a beta, and for production use Tinker asks you to email and join a waitlist. And every Inkling 2 Small tier on the page stops at 256K of context, while the model card says up to 1M. If you need the full window, you can't buy it today.
We could not find a published API model ID or a maximum output length on any page Thinking Machines has up. The model has 594B parameters with 17B active, and takes text and images.
Where $1.18 comes from
The first thing you'll see quoted is about 150,000 output tokens per task. At $1 a million that is 15 cents, which doesn't get you near the $1.18 on Artificial Analysis. We pulled the cost lines from the page, and they add up to the cent once you divide by the rates.
| Line, per task | Cost | Rate per 1M | Tokens | Share |
|---|---|---|---|---|
| Cached reads | $0.873 | $0.03 | 29.09M | 74.1% |
| Cache writes | $0.154 | $0.30 | 514K | 13.1% |
| Output | $0.150 | $1.00 | 149,778 | 12.7% |
| Fresh input | $0.002 | $0.30 | 5,565 | 0.1% |
| Total | $1.178 |
Artificial Analysis Intelligence Index v5.0.1, read November 7. Costs are theirs; token counts are cost divided by rate. Tinker lists no separate cache write price, and Artificial Analysis's figures work out at the $0.30 input rate.
Agents resend their whole context on every turn, so long runs turn into enormous cache-read counts. This model takes long runs: 16.5 minutes per task, against a 7.3-minute median for comparable models, at 151 tokens a second. Without caching the same task would cost about $9.03, so the cache is doing its job. There are just a lot of reads.
Against Inkling Small, released in July at $0.203 a task, the new one is 5.8x dearer. Output tokens are up 4.1x. Cached reads are up 13.5x, from 2.15 million to 29.09 million, and halving the cache rate doesn't come close to paying for that. In exchange the score went from 12.9 to 30.6, so we wouldn't call the predecessor the better deal. It's just a warning about what happens when a rate cut meets a model that works longer.
Same score, different bill
Here are the bars for models near that score, split by the lines that make them up. Look at Sonnet. Its output tokens cost ten times as much, and output is the biggest part of its bill, but it reads 2.5 million cached tokens where Inkling reads 29 million.
Cost per task in USD on Artificial Analysis v5.0.1. Fresh input is 2 cents or less on every row and isn't drawn. The Haiku rows use the under-100K rate; see below.
The full field, in score order. All of these are on the same index version, so the scores compare directly.
| Model and effort | Index | Input / cached / output | Per task |
|---|---|---|---|
| Claude Haiku 5.5 xhigh | 34.6 | $0.10 / $0.01 / $0.50 | $0.248 |
| DeepSeek V4.1 Flash 1017 max | 32.6 | $0.30 / $0.006 / $1.20 | $0.602 |
| Gemini 3.8 Flash high | 30.8 | $0.75 / $0.075 / $3.75 | $7.766 |
| Claude Sonnet 5.5 medium | 30.8 | $2.00 / $0.10 / $10.00 | $0.781 |
| Inkling 2 Small | 30.6 | $0.30 / $0.03 / $1.00 | $1.178 |
| Claude Haiku 5.5 high | 29.9 | $0.10 / $0.01 / $0.50 | $0.159 |
| GPT-6 Luna max | 27.9 | $0.10 / $0.01 / $0.50 | $0.152 |
| GLM-5.3 Flash | 27.8 | $0.15 / $0.03 / $0.50 | $0.583 |
| Mistral Large 4 high | 26.1 | $0.70 / $0.07 / $2.10 | $2.062 |
| Muse Spark Open xhigh | 24.9 | $1.25 / $0.15 / $4.25 | $2.018 |
| Inkling Small | 12.9 | $0.30 / $0.06 / $1.20 | $0.203 |
Rates as listed by Artificial Analysis, which prices DeepSeek at its peak card. DeepSeek's off-peak rates are half that.
Inkling beats Mistral Large 4 and Muse Spark Open on both score and cost, which matters if you're shopping for open weights from a Western lab. It also costs far less than Gemini 3.8 Flash on high, whose problem is $7.03 of cache writes per task, the opposite line of the cache. DeepSeek V4.1 Flash on max scores 2 points more for 49% less, at the peak price.
One caveat on the Haiku rows, and it's a big one. Haiku 5.5 bills every line at five times the rate once a prompt passes 100,000 tokens. The Artificial Analysis figures don't include that: Haiku max's output line is 225,966 tokens for $0.113, exactly $0.50 a million. We can't tell how many of those requests crossed the line, so here is the worst case. If all of them did, Haiku high costs $0.79 a task and xhigh $1.24. So Haiku high stays about 33% under Inkling even then, for under a point less. Xhigh lands 5% over Inkling, with a 4-point lead. Sonnet 5.5 bills the same rate all the way to 1M, so its $0.78 is the cleanest number on the table.
What would change the answer
Every cent off the cached rate takes 29 cents off a task. That is the number to watch. Thinking Machines says other platforms get the model "in the coming weeks", and a host charging Haiku's $0.01 for cached reads would bring this same run to about $0.60, under Sonnet. That's our arithmetic, not an announced price. We found no third-party listing yet, and OpenRouter's API still has only the older Inkling models.
On Tinker's own card, the lever is your agent setup. Everything other than cached reads comes to $0.31, nearly double Haiku high's whole listed bill, so you won't catch that one. To match Sonnet at $0.78 you would need about 46% fewer cached reads, and that assumes the score holds when you compact harder. We have no data on whether it does.
Your workload matters too. One evaluation, Terminal-Bench Science, contributes $0.55 of the $1.18 at $3.68 per task with a 15% weight. If your agents do short chores, every model on the table costs less per task than this, and how they rank could move.
The weights are the point
As an API purchase today, we can't find a reason to pick this over Sonnet medium or Haiku high unless you have a specific fit. Thinking Machines claims the best results among open models on several finance benchmarks, though its own tables put Muse Spark Open ahead on Diligence Bench and GLM-5.3-Flash ahead on Tau 3 Banking. Its launch post shows third place on DesignArena's website leaderboard at 1,340 on November 5. By November 7 the live board had it fourth at 1,334, listed under the ID "california".
The real offer is the one that hasn't shipped. The launch post says Thinking Machines intends to publish the full weights under Apache 2.0 "next week", after external safety testing and red-teaming. As of November 7 we found no public repository, and Artificial Analysis still lists the model as proprietary. The Decoder's headline about the top spot among Western open-weight models is therefore a forecast. It would hold against Mistral Large 4 at 26.1 and Muse Spark Open at 24.9, the best Western open models we found on Artificial Analysis, which files Apache 2.0 as permissive and Muse's license as commercial.
Tinker also lets you fine-tune it, at $1.50 per million tokens with 64K of context and $3 with 256K. If you were going to train on an open model anyway, that is a real reason to look at it. If you just want to pay for tasks, wait for the first third-party host and read its cached rate before anything else.
We've added it to its own model page. The original 975B Inkling is covered in our earlier post, though the index score in it predates the current version. Our Haiku 5.5 post walks through the 100,000-token step. Put your own token counts in the calculator, and include the cached reads this time.
Sources
- Thinking Machines: Inkling 2 Small - launch date, weights plan, availability, DesignArena and finance claims, hero image
- Tinker pricing - sampling, serverless and training rates, context tiers, waitlist
- Inkling 2 Small model card - 1M context claim, parameters, inputs
- Artificial Analysis: Inkling 2 Small and Inkling Small - index, per-task cost lines, speed and time per task
- Artificial Analysis: Claude Haiku 5.5 and Claude Sonnet 5.5 - comparison scores and costs; the other models in the table come from their own pages on the same site
- The Decoder - launch coverage and the open-weight framing