Two new low-cost AI models arrived a day apart, and the cheapest price list does not buy the cheapest finished job.
Three budget models, one buying decision
Two new AI models reached the low-cost shelf one day apart. Thinking Machines released Inkling 2 Small, and Tencent released Hy4. Both now compete with DeepSeek V4.1 Flash 1017, a budget model from October. All three work through a problem step by step before they answer. That costs time and tokens, the small pieces of text that AI services bill.
Picture a team choosing a model for agents, programs that do multi-step work alone. Two questions decide it: what does one finished job cost, and can the team keep the model?
| Compared | Inkling 2 Small | Tencent Hy4 | DeepSeek V4.1 Flash 1017 |
|---|---|---|---|
| Price per million input tokens ($) | $0.30serverless inference; vendor | $0.834*listed API price, full release; vendor | $0.30max reasoning; vendor |
| Price per million output tokens ($) | $1.00serverless inference; vendor | $2.501*listed API price, full release; vendor | $1.20max reasoning; vendor |
| Cost per Intelligence Index task ($) | $1.18average per task; measured | $1.03average per task, reasoning version; measured | $0.60average per task, max reasoning; measured |
| Intelligence Index score | 31Intelligence Index; measured30.6index v5.0, xhigh effort; vendor | 35Intelligence Index; measured | 33max reasoning mode; measured32.6index v5.0, reported by a competitor; reported |
| Vals Index accuracy (%) | 56.1%index v2.2, xhigh effort; vendor56.08%measured | 59.35%accessed via OpenRouter; measured | 61.26%rank 10 of 34; measured61.3%index v2.2, reported by a competitor; reported |
| Time per Intelligence Index task (min) | 15.7 minvendor's own API; measured | 13.9 minutesvendor's own API; measured | 14.4 minutesvendor's own API, max reasoning; measured |
| Context window (tokens) | 1M tokensvendor256Kserverless offering; vendor | 1M tokensvendor1,048,576 tokenshosted deployment listing; vendor | 1M tokensvendor |
| Weights and license | Closed now; Apache 2.0 planned*weights due after safety testing; vendor | Weights planned, not yet released*release promised in coming days; vendor | Open weights, MIT licensecommercial use allowed; vendorOpen weightslicense not stated; vendor |
Data as of November 7, 2026. “Vendor” is what the maker states. “Measured” is an independent test. “Reported” is a third party’s figure that it did not measure itself.
* Tencent Hy4, Price per million input tokens ($): the published data conflict (approximately US$0.90; $0.83). I used $0.834, the exact listed API price, as the maker gives only a rounded one.
* Tencent Hy4, Price per million output tokens ($): the published data conflict ($2.70; $2.50). I used $2.501, the exact listed API price, as the maker gives only a rounded one.
* Inkling 2 Small, Weights and license: the published data conflict (“Proprietary, weights not public”; Open weights). I used “Closed now; Apache 2.0 planned”, the maker's own launch post.
* Tencent Hy4, Weights and license: the published data conflict (Weights not yet released; “Proprietary, weights not public”; Open weights). I used “Weights planned, not yet released”, the maker's own launch announcement.
The price list tells half the story
On the price list, Inkling 2 Small and DeepSeek charge almost the same. Hy4 asks more than twice as much. The cost of a finished task tells a different story. Artificial Analysis, an independent testing firm, measured it on its Intelligence Index, a score built from many tests. DeepSeek did a task for $0.60 on average, while Inkling needed $1.18. Hy4, despite its higher prices, landed in between.
Why does the order flip? The published numbers do not explain it. On that index, Artificial Analysis counted 154M output tokens for Inkling and 160M for Hy4. DeepSeek wrote the most, 174M, and still had the cheapest task. All three wrote far more than the median for comparable models. Price lists are like taxi rates per mile: each model picks its own route. Before committing, I would time a short test run on my own tasks.
“The model I would build on is the one whose weights I can already download, and this week that is still DeepSeek V4.1 Flash 1017.”
Close scores, different strengths
On broad scores, the three sit within a few points of each other. Hy4 leads the Intelligence Index. DeepSeek leads the Vals Index, a score from Vals AI built on professional work such as finance, legal and tax tasks. It scored 61.26%, Hy4 59.35% and Inkling 2 Small 56.08%. Inkling comes third on both.
The gaps open up on specific jobs. Inkling is strong at finance work: Vals AI measured 67.54% on its Finance Agent v2.1 test. On Terminal-Bench 4.0, a test of agents doing software and server work by typed commands, it scored just 12.63%. Hy4 managed 39.39% on that same test. Its weak spot is honesty about what it does not know. On a factual test by Artificial Analysis, Hy4 made up an answer 89 percent of the time when it did not know one. In a report, that means invented facts.
Can you trust these scores? Mostly, yes. Two independent testers produced them. The prices are softer, because each maker states its own. For Hy4, the listed price is $0.834 per million input tokens and $2.501 per million output tokens. Tencent says it offers half off its standard price on its own cloud until December 31, so check which price a quote uses, and plan on the full list price after that.
Which model you can keep
A model's weights are its learned numbers. Whoever holds them can run the model on their own servers. That keeps the service running if the vendor raises prices or changes terms. Today only DeepSeek offers that. Its weights are public under the MIT license, a permissive license that allows commercial use. 27 providers already host it. Its catch is a slow start: Artificial Analysis measured 53.56 seconds before the first word of an answer, against a median of 24.04 seconds. That hurts live chat, not background jobs.
Inkling is closed for now. Thinking Machines says weights under Apache 2.0, another permissive license, will follow next week, after outside safety testing. Artificial Analysis calls that date unconfirmed. Until then, one provider runs it, behind a waitlist. Tencent says the Hy4 weights will follow in the coming days. Today you reach it through Tencent's cloud and apps, or through OpenRouter, a marketplace for AI models.
Where each one fits
For coding agents on a tight budget, I would pick DeepSeek V4.1 Flash 1017. I would keep it away from a live chat window.
For finance and document work, Inkling 2 Small deserves a trial. I would run it on real company files and measure the cost of each finished task, not the price per token. For software and server agents, Hy4 is worth testing, but only behind checks that catch made-up facts.
Nov 7, 2026 · Sketchbook, p. 06
