The mysterious Hy3 LLM is topping OpenRouter Mannequin Rankings by a big margin

OpenRouter is a service that gives entry to most LLMs with a singular API, which has turn out to be exceedingly helpful as of late given the speedy cadence of recent LLM releases. Because of the firm’s position as an middleman between customers and the LLM APIs, OpenRouter has strong, consultant information on how customers work together with LLMs and it publishes this information on the AI Mannequin Rankings web page: a welcome deviation from the labs themselves which usually preserve this information secret for aggressive causes. Just lately, I checked the OpenRouter rankings and observed one thing peculiar.

Two new fashions at the moment are beating LLM darling Claude by way of token utilization and by greater than 50%? I’ve heard of DeepSeek Flash V4: it’s an open-source launch from DeepSeek that isn’t solely quick/low-cost, but in addition performs nearer to the main LLM fashions at a really low value so it’s no shock that it’s extremely standard. However what the heck is Hy3 preview? I’ve by no means heard of Hy3 or anybody speaking about it. Googling it returns an announcement from Chinese language megacorp Tencent about Hy3’s open-source launch: the mannequin web page itself on Hugging Face is sparse and contains oddly sincere benchmark outcomes that aren’t favorable for the mannequin in comparison with different Chinese language open-source fashions.

A Hacker Information seek for Hy3 solely returned a single submission that isn’t about Hy3, and Reddit dialogue is extra in regards to the open-weights launch. One Reddit thread additionally famous the rise of Hy3 however from Could 6, when Hy3 was provided by OpenRouter totally free; that free endpoint is not accessible, and due to this fact Hy3’s utilization within the weekly rankings above is from paying customers.

Hy3 preview is outwardly standard in domains outdoors of agentic coding as properly.

Did I miss one thing? After some nonscientific testing, the mannequin high quality is certainly on par with the opposite Chinese language fashions indicated and never near fashions reminiscent of Claude Opus 4.7 and GPT 5.5. It’s not a magic ignored diamond-in-the-rough, so there needs to be one thing else at play. Fortuitously, OpenRouter has the information to slim down potential explanations, however after checking the information I turned extra confused.

Hy3 preview is accessible from the OpenRouter API at a acknowledged value of $0.066/1M tokens enter which is certainly cheaper than the present top-ranked mannequin DeepSeek V4 Flash with a acknowledged value of $0.10/1M tokens enter. Given the drastically rising value of LLMs and coding brokers, it is sensible {that a} cheaper mannequin would prevail, however provided that it provided comparable high quality and that doesn’t look like the case.

Right here’s the chart of Hy3 preview mannequin utilization over time on OpenRouter from the mannequin web page:

Hy3 preview has no utilization information earlier than Could 8, which suggests that’s the time the mannequin switched from the free SKU to the paid SKU. Utilization can be regular over time since then with the preliminary rankings proven on this put up being a number of weeks after launch, exhibiting that the utilization is not less than natural (or very costly to faux) and never a one-off outlier. Of be aware, for those who do the maths on the numbers offered right here, the input-token-to-output-token breakdown on LLM API calls is now 98% enter, 2% output in combination.

For the OpenRouter AI Mannequin Rankings, there have traditionally been spikes by particular apps switching their default to a specific LLM, reminiscent of when Kilo Code provided Grok Code Quick 1 totally free in September 2025, which rocketed it up in reputation. That doesn’t look like the case right here as a result of apps solely represent a really small a part of Hy3 preview’s exercise.

OpenRouter’s worth proposition is the power to robotically route a given API request to completely different suppliers: for open-weight fashions reminiscent of DeepSeek V4 Flash, OpenRouter lists 13 suppliers, however Hy3 preview solely has one supplier regardless of its open weights: the Singapore-based SiliconFlow. Their utilization web page on OpenRouter reveals that SiliconFlow had comparatively little utilization…till Hy3.

Coincidentially that information visualization reveals that utilization didn’t drop drastically when Hy3 preview moved from free to paid, which in itself is fascinating: if customers weren’t getting worth from the free mannequin, they possible would have stopped utilizing it as soon as the prices hit their pockets.

What am I lacking? Am I overthinking it and the reply is often because “it’s the most affordable” and it obtained ample loss chief traction from the free interval?

…however is Hy3 preview really the most affordable LLM backed by a serious firm on OpenRouter? Whereas I used to be double-checking some assumptions, I discovered that OpenRouter has information that reveals Hy3 preview is just not the most affordable well-performing LLM accessible: it’s really DeepSeek V4 Flash, however with fascinating caveats.

LLM Economics in 2026#

So listed below are a number of extra notes about how LLM APIs work that aren’t usually mentioned. LLM calls are nonetheless stateless, which implies that after each flip (together with consumer messages to the LLM asking questions), the entire tokens within the present dialog thread are reprocessed, that means that within the case of brokers, the depend of enter tokens will increase cumulatively with every successive message and is one motive why beginning new threads regularly as context fills up is inspired for efficient agent use.

However even earlier than agentic workflows, massive inputs reminiscent of full PDFs bloated context equally. Consequently, most LLM suppliers applied immediate caching, which reuses enter tokens processed earlier within the dialog: this can be a win-win that saves time/compute for the LLM supplier and the financial savings are handed to the client. Most LLM suppliers cache inputs robotically, together with when accessed by way of OpenRouter: the disk-lightning-bolt image subsequent to the fee signifies tokens have been cached and the cache could not at all times be hit, particularly if OpenRouter switches suppliers mid-thread. The odd API supplier out is the Anthropic (Claude) API which requires paying for a cache write first for some motive.

Sometimes, cache learn prices are 10% of the enter prices: that is the case for the most recent fashions from OpenAI API, Anthropic API, and Google Gemini API. For the 13 suppliers that serve DeepSeek V4 Flash, cache learn prices are between 20% and 50% of enter value, which is sensible as they could not have the identical economies of scale. There’s one DeepSeek V4 Flash supplier that’s an exception, although:

That’s a 2% cache learn value! (multiply by 2, transfer decimal left 2 locations) How are DeepSeek’s cache learn costs so low? DeepSeek has applied a brand new method to KV caching beginning with V4 and because the mannequin’s creator it’s positioned to greatest leverage its personal improvements, which as talked about the advantages are handed to the client. The DeepSeek V4 Professional variant mannequin, when served by DeepSeek, has a cache learn value of 0.83%! (use a calculator for that one)

Bear in mind how I confirmed that 98% of LLM API prices at the moment are enter tokens, that are aggressively cached? Which means the “acknowledged” costs of LLMs at the moment are deceptive, however unusually in a pro-customer method as a result of the efficient value shall be less expensive! To counter this ambiguity, OpenRouter now has a desk for efficient costs on the mannequin web page, which accounts for the fee financial savings from cache hits. Right here’s the efficient pricing for DeepSeek V4 Flash through OpenRouter by supplier, which is completely different for every supplier as they’ve completely different cache learn prices and cache hit charges:

The costs are all over, however discover the second row the place DeepSeek itself is the supplier, which is priced at a whopping $0.018/1M enter tokens! That 2% cache learn actually pays off. Evaluating apples to apples with Hy3 preview, the efficient pricing for Hy3 preview as famous on its mannequin web page from SiliconFlow (a whopping 44% cache learn value) is $0.034/1M: almost double DeepSeek V4 Flash from DeepSeek! After all, that is solely relevant if DeepSeek is explicitly used because the supplier, which some downstream OpenRouter purchasers/brokers could not assist: the OpenRouter costs match the costs immediately from DeepSeek, so utilizing a direct DeepSeek API key will work the identical.

There may be additionally an elephant within the room: DeepSeek is a China-based firm and a few could not need—or could not legally give you the option—to present their fee processing info or LLM enter information to a Chinese language firm who has set immediate coaching = true on their OpenRouter information coverage info, which is a reliable concern.

Sure, subscription-based LLM providers reminiscent of Claude Code and Codex are nonetheless one of the best bang on your buck for those who’re in a position to constantly exhaust the utilization limits. However the super-cheap DeepSeek V4 Flash through the API doesn’t lock you right into a subscription, and for those who want a bit extra agentic compute to complete a venture, it’s cheaper than paying for additional utilization from the subscription providers. At least, it’s a microeconomic examine in opposition to further pricing shenanigans that may possible proceed by way of 2026 as competitors in agentic AI heats up.

General, I nonetheless don’t perceive the recognition of Hy3 preview on OpenRouter. Given the accessible information and evaluation above, my guess is {that a} single massive app not affiliated with Tencent is certainly utilizing Hy3 as its data-processing spine, and this app isn’t solely an agentic coding app. However one of many benefits of OpenRouter is that it’s low-lift to modify fashions and suppliers: it wouldn’t shock me if DeepSeek V4 Flash will get a spike in a number of weeks as soon as individuals catch on to its pricing.