Tencent Hy4 kept the preview's price and took half off until December 31. An Artificial Analysis task drops from $1.03 to about 52 cents, cheaper than Sonnet 5.5 medium, until January 1 flips it back.
The finished Hy4 kept the preview's rate card to the fen and then took half off it for eight weeks. That makes it one of the cheapest ways to buy a mid-30s index score in November, and a much less interesting one in January.
Photo by Joshua Aragon on Unsplash
Fifty-two cents, for eight weeks
Hy4 went GA on November 4 with image input and computer use added. The price didn't move: ¥6 in, ¥18 out, ¥0.30 for a cache hit, or $0.834, $2.501 and $0.042 on the dollar card. What moved is a 50% discount that runs to December 31 on Tencent's own TokenHub and, with no end date shown, on OpenRouter.
Artificial Analysis scores it 34.6 and bills a task at $1.03 at list. Halved, that's about 52 cents, which undercuts DeepSeek V4.1 Flash at peak and Sonnet 5.5 on medium while scoring higher than both. On January 1 it goes back to $1.03, and Sonnet medium is cheaper again.
Same card as the preview, half off until New Year's Eve
Tencent's TokenHub billing page, updated November 5, lists the GA model as hy4 right next to hy4-preview, at the same price. The Shanghai STAR Market Daily reported it the same way: pricing unchanged. Under the table is a notice that Hy4 calls bill at 50% "until 2026-12-31 23:59:59" Beijing time. That's 15:59:59 UTC on the 31st, or 8 a.m. Pacific, so a US team pays list for most of December 31.
| Per 1M tokens | Input | Cache hit | Output |
|---|---|---|---|
| TokenHub list (yuan) | ¥6 | ¥0.30 | ¥18 |
| TokenHub to Dec 31 (yuan) | ¥3 | ¥0.15 | ¥9 |
| TokenHub batch, list (yuan) | ¥3 | ¥0.15 | ¥9 |
| Dollar list card | $0.834 | $0.042 | $2.501 |
| OpenRouter tencent/hy4 now | $0.417 | $0.021 | $1.25 |
TokenHub read November 8, 2026; the Guangzhou and Singapore tables match. No cache write price and no length tiers are listed. OpenRouter output is $1.2505 before rounding.
OpenRouter's endpoint data shows one provider, Tencent, serving at fp4 with a discount: 0.5 field and implicit caching. So the international route gets the same deal. What it doesn't get is a date. If you budget off OpenRouter, plan for the Beijing deadline anyway. A discount with no end date can end any day.
Batch on TokenHub is already half of list. The notice doesn't say whether the two discounts stack, so we wouldn't plan on ¥1.50 input until someone sees it on an invoice.
Pay in dollars, not yuan
We wrote in August that Tencent's dollar card is the yuan card converted at about 7.196, while the market sat near 6.73. That gap is still there. The yuan traded at 6.69 to the dollar this morning, so ¥6 is really $0.897. The $0.834 card is 7% under that.
The discount keeps the gap. Discounted yuan works out to $0.448 input and $1.345 output at spot, against OpenRouter's $0.417 and $1.25. Put the other way, paying in yuan costs 7.5% more on every line, for the same model from the same provider. If you can buy in dollars, do.
Where a $1.03 task goes
Artificial Analysis prices its index run at the list card, not the discount. We pulled the cost lines and divided by the rates to get token counts. They add up to $1.031.
| Line, per task | Tokens | Rate | List | To Dec 31 |
|---|---|---|---|---|
| Uncached prompt (AA's cache-write line) | 584K | $0.834 | $0.487 | $0.243 |
| Cached reads | 8.16M | $0.042 | $0.343 | $0.172 |
| Output, incl. reasoning | 78,058 | $2.501 | $0.195 | $0.098 |
| Fresh input | ~7K | $0.834 | $0.006 | $0.003 |
| Total | $1.031 | $0.516 |
Artificial Analysis Intelligence Index, read November 8, 2026. Costs are theirs; token counts and the discounted column are our arithmetic.
Output is less than a fifth of it. The biggest line is what AA calls cache writes, $0.487, even though Tencent lists no cache write price. Divide it by $0.834 and you get 584,000 tokens at the plain input rate. That's prompt text the first time the cache sees it, billed as ordinary input, which is correct. It's also the line you control least, since every new file or tool result an agent reads lands there.
Tencent told TechNode on November 6 that Hy4 costs about $0.78 a task on Artificial Analysis. We couldn't match that to anything on AA's page, which says $1.03 at list and works out to $0.52 at the discount. Neither number is $0.78.
November against January
Here's the per-task cost for models near Hy4 on the index, plus a few above it. Hy4 appears twice.
Cost per Artificial Analysis index task in USD, read November 8, 2026. AA prices DeepSeek at its peak card; off-peak is half. The discounted Hy4 bar is AA's list figure halved.
Until December 31, Hy4 beats DeepSeek V4.1 Flash at peak by 14% and scores 2 points more. It beats Sonnet 5.5 medium by a third and scores almost 4 more. GLM-5.3 on max costs nearly ten times as much for a lower score. Those are good trades for eight weeks.
From January 1 the picture flips. Hy4 at $1.03 costs 32% more than Sonnet medium. Sonnet high is $1.64 for 41.5, so for 61 cents more a task you get 7 more points and a price with no expiry on it. DeepSeek off-peak does the same work as Hy4 for roughly $0.30.
And if you just want the cheapest decent score, GPT-6 Luna on max gets 27.9 for 15 cents. Hy4 doesn't touch that on cost per point, discount or not.
What GA added
Per TechNode's launch report, Hy4 now reads charts, documents and several images in one prompt, and does computer use across the command line and a GUI, with specific training on Blender, CAD and long tasks that cross apps. The window is still 1,048,576 tokens, and OpenRouter lists 65,536 of output, up from the preview's 64,000. Reasoning runs at high, low or none, default high.
We found no separate image price on TokenHub or OpenRouter, so images seem to bill as input tokens. Tencent hasn't said how many tokens a screenshot costs, and that number matters more than the rate for computer use, where an agent sends a screenshot every step. Without it, you can't price a computer-use run in advance.
The benchmarks Tencent published for GA are mostly its own: WorkBuddy, a finance blind test and a hallucination rate it says fell from 12.95% to 7.2%. AA's own run puts Terminal-Bench 4.0 at 37.4% and AutomationBench at 65.6%. No vision or OSWorld numbers from anyone yet.
Free quota, weights and Flash
WorkBuddy and CodeBuddy give Hy4 away for free, in an unstated amount, from November 4 to 15. Tencent says it raised those limits on November 5 and expanded inference capacity tenfold after people hit them on day one. That's Tencent's own app, not the API.
The preview's weights are on Hugging Face under Apache 2.0. The GA weights aren't. Tencent says they come "soon" with a system card covering cyber and bio risk, and the tencent/Hy4 repo that OpenRouter already links to returns an auth error. Artificial Analysis lists the model as proprietary for now. Hy4 Flash is announced with no date and no price.
Our take: run your evals on Hy4 now, while a mistake costs half. If the numbers hold up, check two things before January. Does Tencent extend the discount, and do the weights ship, so another host can undercut the list card? If neither happens, Sonnet medium or DeepSeek off-peak is the cheaper place for that traffic.
Hy4 has its own model page, and our catalogue switches it to list price on January 1. Compare it with everything else on the pricing table, or put your own cache-hit rate into the calculator. For Hy4, the first-read prompt line moves the bill more than output does.
Sources
- Tencent Cloud TokenHub billing - yuan card, batch rate, discount notice and end time
- OpenRouter endpoints for tencent/hy4 - dollar prices, discount field, fp4, context and max output
- Artificial Analysis: Hy4 - index score, per-task cost lines, tokens, evals; comparison rows from each model's page on the same site
- TechNode, November 4 and November 6 - launch, vision and computer use, capacity, the $0.78 claim
- STAR Market Daily (cls.cn) - unchanged pricing, Tencent benchmarks, weights and Flash timing
- Hugging Face: Hy4-preview - preview weights and license
- ExchangeRate-API - USD/CNY 6.6908, November 8, 2026