Chinese AI models just completely overtaken US ones on OpenRouter (largest AI API aggregator). This shift happened in 2026 thanks to Kimi K2.5, MiniMax M2.5 & DeepSeek V4. For most tasks they are a no-brainer: 10-20x cheaper, +90% as good (coding, writing etc), and better privacy (open-source, self-hostable). Two types of users now: those already using Chinese models, and those who will. US/SK/Japan market collapse due to AI bubble burst is getting nearer xD
>>536635914Why do Chinese models require so many tokens? Why are US models so lightweight and efficient?
>>536636065Thats bullshit kek xD The graph shows _total_ tokens used, not per-query. Chinese models dominate because they are 10-30x cheaper with near-parity performance (MoE efficiency + lower inference costs), so devs run way more of them. US models arent lighter, they are actually far pricier. xD
>>536636065US models for public use have to be lightweight and efficient, otherwise they are not a good business model. We dont have access to thousands of workers, the ability to take land and build power sources/datacenters like China does. They will 100% overtake us in AI in the public model/general use. We have to specialize somehow
>>536636188Nobody's ever really seen this supposed "Chinese domination" beyond cope graphs though xD
Does this take in to account that most people aren't using API tokens but the 20 or 100 dollar monthly sub (on multiple accounts?)
>>536636305Nope. OpenRouter tracks API/pay-per-token usage only (credits bought for routing). It ignores direct consumer subs like ChatGPT Plus, Claude or whatever. The graph reflects dev-heavy traffic favoring cheap Chinese models.
>>536636065>Why do Chinese models require so many tokens?Are you retarded? Deepseek V4 is 70 cents per million token output, Opus is 22 euros.
Using open-source to destroy jewish western industry is 1000% based and redpilled
Chinese models are 90% as good at 1/20 of the price, it's a no brainer.
American AI
>>536636241>US models for public use have to be lightweight and efficient, otherwise they are not a good business modelThe only American model that is actually well optimized in how it handles context and memory is Gemini. Claude, ChatGPT are resource hogs and Grok is a joke. US tech companies and politicians are going all in with the financial bubble and throwing more computing power at transformer models trying to brute force YHWH into existence, because if the music stops the stock market charade collapses.>>536636483They're not open source. Actual open source models are Gemma and LLaMA (albeit with a restrictive license). Chinese models (Deepseek, GLM, Kimi etc.) are open weight, and anyone can download them and run them locally or sell metered API access, but the actual code is still closed source. Albeit Deepseek does release a ton of their research and tools as open source and they pioneer a lot of new techniques.
oh snap elon
>>536636065Tokens are a commodity. Commodities cost money. Much like electronics, China is pricing the US ventures out of the market.
True if big
BTW, i stopped using Grok for most tasks because using it without paying subscription became a pain in the ass. I moved to Deepseek app instead. Answers take few seconds more but the model is absolutely free of charge and as good as Grok. I use Grok now only to generate funny photos like adding sidelocks and kippahs to photos of famous Poles xD
>>536635914bovered ain't I.. fuckin tokens. couldn't give a shat.
>>536636065serious answer here:because chinese tokens are not as effective as us tokens.by not effective, I mean models like Deepseek have like 125k entries in the vocabulary only, while gemini or chatgpt is 250k.this means that for a similar piece of text, chinese models might need upto x2 tokens (input, output).the reason of that limitation is obviously because they don't have nvidia latest cards, thus limiting their computation power so they can not afford such fat vocabulary.
>>536635914American models are so restricted in their thinking they are a pain to use. USA is gay.
>>536636372yeah because fuck google.we had it good with gemni 2.5 models.now the fucker increase the price like crazy, of course we will pick cheaper one.though, usually for us models, it is easier just use the official services (vertex for gemini, azure for chatgpt), no need for 3rd party like openrouter.
>>536636984Grok is good at this shit, but if you guys know any Chinese free-of-charge photo-manipulating algorithm, i would gladly stop using Grok at all.
>>536636704I hate americans
>>536636984Grok isn't really a frontier model in any sense though. Really that's down to just openai and anthropic at this point. Even google is behind
>>536637321americans literally shot themselves in the head because they’re obsessed with mind control. you are the stupidest country on earth
>>536635914I've been using deepseek in my economics courses for my masters degree.Not a SINGLE hallucination.And I'm almost done with my degree
>>536636065U.S. models are now optimized to use the fewest tokens possible, giving you answers that an Amazon bot could provide, as seen with Gemini, ChatGPT, and even Claude Sonnet.Energy plays a fundamental role here, as China has cheaper energy and is reducing its cost every day, while all the graphs show that the U.S. has neither the capability, the flexibility, nor the infrastructure to support the evolution of new models and compete with Chinese open-source ones.This is why there is a craze to create new data centers; they need the infrastructure and the energy if they want to win the AI race. However, there is another critical detail: the rare earths used to make chips and how China has an almost complete monopoly over them. Even if one argues that ASML and TSMC chips are better, and yes, they are, this has impelled China to keep developing new chips, to the point that they are already releasing RAM memories and graphics cards approved to be sold in the rest of the world. It is not a matter of if they will succeed, but when.
>>536635914OK, Keep us posted.
>>536635914interesting graph, US models didn't decrease in numbers
>>536637272more like cope answer here lmao, reddit spacing glownigger
>>536637500>completely ignores the social credit system automatically baked in to any ccp modelsHow is china's not?? It was literally engineered from the get go at social conditioning??
>>536639179>muh sooocial credit mythlmao lof effort glownigger, here's your +100 palantir points
>>536639243If you have evidence disproving the claims then post it. Otherwise you just look like a chink shill.
>>536635914You don't go to OpenRouter if you want to use OpenAI/Anthropic/Google models. You go to the companies directly or to AWS/Azure/GCP. Therefore, the vast majority of US model usage isn't captured in that data.
>>536635914>>536636188ask the CHinkshit model what it is and it will always admit its claude
>>536639307you're not a serious person if you still trying to shill this "muh social credit system" bullshit cia talking point, but sure here's a video glowniggerhttps://www.youtube.com/watch?v=iXrx9ryVaeslet me guess, your response will be hurrhurrrr chink ccp shill blahblahblah clown ass palantir golem cia nigger
>>536639362Wrong. OpenRouter does route to official OpenAI/Anthropic/Google models via unified API. Many devs prefer it for easier switching, fallbacks, management etc. It captures significant US model traffic, especially price-sensitive high-volume use.
>>536635914From my experience, Chinese models are very wordy. Western models will write at least a page if you don't head them off up front, but Chinese models will write 2 or 3 pages for the same response to convey the same ideas
>>536635914>on OpenRouter (largest AI API aggregator)>Kimi K2.5, MiniMax M2.5 & DeepSeek V4Do you mind translating that into SOMETHING THAT IS EVEN REMOTELY INTELLIGIBLE, YOU FUCKING BOT-TEXT PASTING CUNT?!
>>536636241>We dont have access to thousands of workersgood morning saar. i will come and do the needful. you cannot let china win saar. they are communists saar. think of the uighurs saar.
>>536639537Even still, the vast majority of US model usage isn't captured in that data.
>>536639590I dont establish threads for retards who struggle to understand basic terms, sorry Winnetou. Copy-paste my thread to LLM and ask it to explain it like for an idiot xD
>>536639640OpenRouter is just one aggregator (1% of global inference spend) so it is obvious that it cant cover entire global output xD Same applies to Chinese models. Chinese models see huge direct usage via Alibaba, DeepSeek, Moonshot, Baidu APIs, self-hosting (open weights) and domestic enterprise deals - faaaaar beyond OpenRouters aggregator slice. OpenRouter captures mostly price-sensitive international traffic. OpenRouter is highly relevant gho. It processes tens of trillions of tokens weekly, serves millions of devs globally and offers clearest public view of cross-provider usage trends, price sensitivity and model adoption in the competitive API market. Researchers and analysts widely cite it.
>>536637272Antidisestablishmentarianism = 1 token ins 250k vocab anti, dis, establish, ..., = 4 token in a 125k vocab. But I'm not sure if it's really important. unless you are seeking very long answer output like 8000 words
>>536636809kek, fair use is fair
>>536639172Nah he's right, it's different when prompted in Chinese vs. English. But the difference isn't that large.>>536639590Those are very understandable terms.
ching chong ping pong
>>536637272Thats bullshit. DeepSeek has 129k vocab (similar to Qwen 152k, Kimi 160k). GPT-4o 200k, Gemini 256k, so diferences are modest. Token usage on OpenRouter is driven by price + high-volume agentic/coding workloads, not vocab inefficiency. Modern Chinese tokenizers are efficient for multilingual text.
>>536639365Funny as fuck when it's been proven that American models will say they are actually using deepseek's models
>>536636065Why would somebody would just go on the internet and lie?
I wish this was true, since it would force frontier models into more defensive posturing where they would subsidize more and release better models to the public. Sadly, just like everything else from China, it’s just more fakeness and gayness.
>>536640827now try asking an american model about the holocaust, faggotry or shitskin issues
>>536644895>Sadly, just like everything else from China, it’s just more fakeness and gayness.Neo-Brazilians will repeat this cope till the end rather than face reality.
>>536636258I work for a medium-sized tech company and we use chinese models because:>a.) we can run them locally and not give our data to kikes (makes compliance very easy)>b.) much cheaper
I used a bunch of different Chinese models. They are great. Their large models really are a lot more efficient than western big models. They won't come close to Claude in coding, though. Ironically the best small-mid sized locally run model for roleplaying your waifu right now is Google's Gemma.
>>536645131>China never lies; it's just cope to say so
>>5366359145 American models4 chink modelsOP is a stupid laying faggot
>>536646525The graph shows Chinese models dominating: DeepSeek #1, Tencent #3, Xiaomi #5 , Minimax #6. US models: Anthropic #2, Google #4, OpenAI #7. Chinese hold top volume share thanks to price & efficiency. Fuck off, dumb Pole xD
>>536636984Grok just censored its video generator even for 3rd party API. It was fucking amazing until Friday.I can't help but feel we're losing functionality we'll never get back. The local models you can run aren't going to catch up to the current corporate ones any time soon, if ever.
>>536647144>Grok just censored its video generator even for 3rd party APISad
>>536647055the graph shows american models have higher market share than chink onesyou low IQ subhuman kike
>>536647295Your previous graph was a token share graph and it literally proved that Chinese models lead. The graph you just posted is a capability graph. US leads in frontier Elo, but Chinese models deliver near-parity at 10-30x lower price which results in massive volume advantage on OpenRouter. Fuck, dumb subhuman Pole xD
>>536647144>>536647181Yeah, it's funny because when it comes to AI video generation, the company that leaves their model uncensored WILL win. It's not even close.Yeah, they'll have to deal with ethical questions and a lot of condemnation. But they'll be the one everyone uses.The only way the censorship can continue is if they ALL agree, and there's nobody that wants to grab the huge bag of cash just sitting on the table.
>>536647523>FuckFuck off*
>>536647523chinese models lead what?number of retarded subhumans who use free deepseek tokens?
>>536647672You are too dumb to interpret graphs you upload, you retarded Pole.
>>536647732you have the same IQ as a pajeet>Based on publicly available OpenRouter usage data and industry trends (as of mid-2026), USA models currently hold a larger market share than Chinese models on the platform.kill yourself you low IQ kike
>>536648226False. As of June 8, 2026, the latest OpenRouter data shows Chinese models leading token market share: DeepSeek #1 (18.9%), Tencent #3 (13%), Xiaomi #5 (8.9%), Minimax #6 (8%). US models lag: Anthropic #2, Google #4, OpenAI #7.https://openrouter.ai/rankingsStop using your hallucinating US LLMs and read sources instead, if you can read with understanding, you fkn dumb subhuman Pole xD
>>536636809is this from the newly announced star wars xcom game?where can I preorder?
I asked ChatGPT to edit a document for work and it returned portions of it in Hindi.I will never use it again.
>>536648429look top leftimagine being so much of a subhuman like youi can't
>>536648815DeepSeek hallucinates too when it tries to read dynamic data.
>>536648910>DeepSeek hallucinates too when it tries to read dynamic data.my Opus 4.8 does notyou chink shit is beyond embarrassing
>>536649282Opus 4.8 frequently hallucinates on charts/dynamic data (misreads bars, invents numbers, fabricates trends). Users report it in agentic workflows.https://www.reddit.com/r/ClaudeCode/comments/1tsp6h0/opus_48_is_constantly_hallucinating_and_is_stuck/
>>536649505>redditLOOK AT THIS FAGGOTAHAHAHAHHAHAHAHAHHAHAHHAHAHAHAHHAHA
>>536637272AI slop is just words words words the answer to your question that you were too lazy to search words words words here is some cool behind the scenes facts words words words happy cake day kind sir words words words
>>536649715I'm convinced that guy is a Chinese shill, and not even legitimately Polish, or in Poland
>>536649282>my Opus 4.8 does notIf you know anything about the technology you'll realize it's impossible to stop hallucinations. It's also why the LLM -> AGI idea is unworkable bullshit sold by retards for retards.
>>536649715Official Anthropic Claude API Docs (Vision section) explicitly warn:"Claude may hallucinate or make mistakes when interpreting low-quality, rotated, or very small images under 200 pixels" and dynamic/complex charts often trigger this. Link: https://docs.anthropic.com/en/docs/build-with-claude/visionAll modern LLMs read static data from static good resolution photos all right but they usually get fucked when reading dynamic charts.
>>536637590>win the ai raceKeep telling yourself thatEveryones busy building infrastructure for the antichrist so he literally walks into a environment ready to adopt the mark of the beast currency and financial singularity magnum opusTheres zero competition here dude, lmao, coke vs pepsi illusion of choiceEverythings a fake and gay hegelian dialectic at this point
>>536650089I am being a Chinese agent only in free time between being a Russian Olgino agent and a tasked/hired jeet.
>>536650190i made 200 thousand USD from NVIDIA the retard might be YOU
Remember when cursor released an "AI model" that was just a wrapper around Chinese Kimi k2.5?
>>536650717Japan's rakuten did the same thing trying to pass off deepseek v3 as their own modelRemember when Japan led everyone else in tech?
>>536650717it was not a "wraper" the cursor model is superior in every way to kimi k2.5
>>536635914It's funny because they all know how to make them this cheap, but they need to charge the premium to have any hope of recouping all the investor money they duped people out of.
>>536650476Typical Pole boasting about imagined things that exist only in his mind. Fun fact: the most frequent/popular social media trend in Bolanda for decades was/is taking selfies/photos in front of somebody elses luxury cars. Pollacks also use to make discount last-minute holidays photos only to upload them on Facebook and pretend being "rich". An average Pole driving 10 yo trash-car is also buttblasted about Ukros driving here in expensive cars that they bought for busification bribes and aid-money defraudation xD
>>536650881cope and seethe russian mongoloid pole can buy your whole village with an month pay
What's a good chink model? Deepseek can't analyze images
>>536650829But it is literally Kimi k2.5https://xcancel.com/elonmusk/status/2034941631871455262#m
>>536635914Mimo is what you want, clears those and is cheaper than actual dirt as well. But US models still rape everything. Sadly the gap hasn't closed, everything is just better, second tier intelligence is getting commodified but the best is still a duopoly. If US model lead continues that difference will always be worth a huge amount and they will accrue the most value.
