This article was written with AI (Claude) and reviewed by an editor. Our company is a paying customer of both Anthropic and OpenAI, and we also use Google models in testing. Models from Anthropic, the company that built the AI that wrote this article, appear three times as comparison baselines in this story — the Hugging Face post set Kimi K3.1 Pro's score and cost alongside Claude Fable 5.2, Meta compared Muse Ember with Claude Haiku 5.5, and sixth place in the OpenRouter rankings is Claude Opus 5.5. For that reason we add no evaluation of any model and only carry over the published numbers and sentences. Figures a company announced are marked "company-reported," figures a platform counted from its own traffic are marked "the platform's own measurement," and anything from a single source is given together with that source. The acquisition news in this article relays only what was reported, and this article is not investment advice on any specific company.
In NO.48 we wrote "switching models takes weeks"; in NO.46, "read the price per task, not the price per token"; in NO.41, "under the one word 'open weights' there are three sets of license terms." This week's three stories show where those numbers have flowed — seven in ten tokens passing through the router go to open-weight models, a new player holding the same price sheet has entered the smallest-model tier, and the tool that runs those models on your own computer reached 1.0. We lay it out as facts, interpretation, and what we would do.
1. Hugging Face → OpenRouter — Handing Over Routing and Publishing "Open Weights 71.7%" (facts, company-reported / platform measurement)
On November 16, Hugging Face posted "Hugging Face and OpenRouter join forces to make open-weight models easy to run" on its official blog. We read it directly. There are two parts. The first is a product change — in Hugging Face's "Inference Providers," routing of language-model requests moves from its own router to OpenRouter's router. Whichever provider serves an open-weight model on the Hub, you call it with a single Hugging Face token; for each request OpenRouter picks the provider by price, speed, and reliability, and if one fails it resends to another (previously, the post says, that provider's error was returned as-is). You can append `:fastest` or `:cheapest` or a provider name after the model name to choose, and if you send the same ID per conversation in a header (`X-HF-Session-Id`), it keeps you pinned to the provider that holds your prompt cache. It accepts OpenAI's Chat Completions and Responses formats and Anthropic's Messages format, and setup guides are attached for coding tools such as Claude Code, Codex, and OpenCode. There is one breaking change — you can no longer use your own provider API keys for language models (image and video keys are unchanged). Prices are the providers' prices as-is with no additional Hugging Face fee, the post says, and if you turn off a specific provider in settings the router will not send there. Image, video, and audio generation are not part of this change.
The second part is the numbers. All of them are what OpenRouter counted from its own traffic. The share of tokens going to open-weight models went from 26.8% in October 2025 → 71.7% at the end of September 2026 (peaking at 77.5% in July); over the same period, weekly open-weight token volume grew more than 60x while closed models grew about 10x. The most-used model this week was DeepSeek V4.1 Flash 1017 at 59.2 trillion tokens. Average prices per 1M tokens were $0.28 input / $1.62 output for open weights and $1.27 / $6.36 for closed models — about a 4x difference. Citing the Artificial Analysis Intelligence Index, the post set the top open-weight model, Kimi K3.1 Pro, at 46 points and $3.68 per task next to the overall leader, Claude Fable 5.2, at 65 points and $26.54 (we opened the AA leaderboard ourselves on November 18 and confirmed the same numbers — a 19-point gap at about 7x the cost). OpenRouter's scale: more than 800 trillion tokens a month, more than 10 million users, more than 600 models, more than 80 providers, and after rebuilding the router, capacity errors down 46%, time to first token improved 16.9%, throughput up 16.3% (OpenRouter's own measurement). The post has no direct quotes and was published under the names of several Hugging Face staff.
We opened the OpenRouter rankings (openrouter.ai/rankings) ourselves on November 18. The top 10 by weekly tokens through November 16 were DeepSeek V4.1 Flash 1017 (61.2T), Tencent Hy4 (16T), DeepSeek V4.1 Flash (15.5T), Z.ai GLM 5.3 Flash (11.8T), GPT-6 Luna (8.08T), Claude Opus 5.5 (7.37T), StepFun Step 5 Preview (6.36T), Xiaomi MiMo-V2.6-Flash (4.16T), GLM 5.3 (3.74T), and GPT-6.1 Sol (3.71T). The top four are all models from Chinese companies, and US models sit at 5th, 6th, and 10th. The slight difference in the number-one figure between the blog post (59.2T) and the rankings (61.2T) appears to be a difference in tally date, but we could not confirm this. The rankings text had no total-volume or open/closed-share numbers. For background, a Bloomberg chart relayed by OfficeChai (June 26) (data from OpenRouter and Exponential View) put the combined token share of Google, OpenAI, and Anthropic at about 70% in June 2025 → about 30% in June 2026 (approximate, no methodology).
One thing to note here. Both companies had acquisition announcements this year. Reuters (August 19) reported that Stripe had agreed to acquire OpenRouter (price undisclosed; a source said "a little over $8 billion"; Wikipedia relays Bloomberg's and the WSJ's "more than $7 billion"), and Stripe's CEO said "tokens are the central currency of companies building with AI." Nvidia announced on September 4 that it would acquire Hugging Face for $12.9 billion (Fortune, Advanced Television; reported as an "agreement," and we could not confirm whether it has closed), and Jensen Huang wrote on his blog that "Hugging Face will remain an open platform for the entire AI ecosystem, and using Nvidia compute is not required" (as quoted by Advanced Television). The November 16 post mentions neither acquisition. The OpenRouter scale figures in the Reuters article (10 trillion tokens a day, 10 million developers) and the company claim on Wikipedia (400 trillion tokens a month) are from different dates than the November post (800 trillion a month), so we left them as they are.
2. Meta Muse Ember — "Same Price, One-Third per Task" in the Smallest Tier (facts, company-reported / independent evaluation)
We read "Introducing Our Smallest, Most Efficient Model: Muse Ember" on the Meta AI research blog (November 9) directly. The intended use is "executing complex reasoning and the work agents repeat all day" — tool calls, coding subtasks, routing, extraction, and parsing long documents. The price is $0.10 input / $0.004 cached input / $0.50 output per 1M tokens, and $0.30 / $0.012 / $1.50 for prompts over 200,000 tokens. Context is 1 million tokens, with a maximum output of 512,000 tokens. It is available through the Meta Model API (`muse-ember-1.0`), Muse Code, Google Cloud, and Oracle Cloud. The company's table compares it with GPT-6 Luna and Claude Haiku 5.5 at the maximum reasoning level — AutomationBench-AA 37.8% vs 20.2%/13.9%, Terminal-Bench 4.0 40.9% vs 16.4%/39.2%, OSWorld 2.1 69.5% vs 50.5%/68.9%, and on FrontierCode 2.0-AA it is 27.2% vs 25.2%/34.9%, where Haiku leads. An Artificial Analysis Intelligence Index of 36 is "tied with Haiku 5.5, above Luna (28)," and "matched Haiku 5.5's top score at about one-third the cost" is the company's summary sentence. Against its own higher-tier model Muse Spark 1.3 (index 37, about $2.39 per task), it is 1 point lower at about one-seventeenth the cost. The post itself notes that several benchmarks are internal measurements and that some use a specific harness or partial scoring. Before release, it says, the model was evaluated under the "Meta Superintelligence Scaling Framework" for chemical, biological, cyber, and loss-of-control risks and came in below the high-risk threshold, and went through agentic-behavior evaluations (refusing instructions hidden in web pages and tool outputs, asking a human before high-impact actions). There is no mention of open weights.
We opened the independent evaluation side directly. Artificial Analysis's Muse Ember page lists an Intelligence Index (v5.0.1) of 36 at the max level (2nd of 36 in its class) and 34 at xhigh, $0.14 per task, output speed of 449 tokens/second, and 3.1 minutes per task (the class median is intelligence 24 and prices of $0.24/$0.83). The AA overall leaderboard (November 18) shows Claude Haiku 5.5 Max at 36 points and $0.41 per task, GPT-6 Luna Max at 28 points and $0.15, DeepSeek V4.1 Flash 1017 Max at 33 points and $0.60, and Muse Ember Max at 36 points and $0.14 — Meta's "one-third" matches the $0.14 vs $0.41 in this table. What Simon Willison (November 10) pointed out is the price sheet — Ember, Luna, and Haiku 5.5 all have the same short-prompt price of $0.10/$0.50, and what differs is the "long prompt" threshold (200,000 / 272,000 / 100,000 tokens) and the price beyond it (Haiku is $0.50/$2.50). So he wrote that "price per token is a weak comparison — the number of tokens spent on the same task differs by model," and charted AA's cost per task (Ember xhigh about $0.12, max $0.14; Haiku xhigh about $0.25; Luna xhigh about $0.10). In his own tests, he wrote, "high" among the six reasoning levels failed to finish an SVG, and the level did not cleanly predict output quality. Whether Ember's weights are released appears nowhere — not in Meta's post, the AA page, or Willison's post.
3. llama.cpp 1.0.0 — The Local Runtime Promised "If It Breaks, the Number Goes Up" (facts, project announcement)
We read "Hot llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem" directly, posted by ggml-org on the Hugging Face blog (November 13). llama.cpp is a C/C++ runtime for running LLMs on your own computer (laptop, server, mobile device), and it put 1.0.0 on two libraries (`libllama` and `libggml`) about three and a half years after the first commit. The promise is semantic versioning — a change that breaks the public C API bumps the major number, compatible features and new model support bump the minor, bug fixes bump the patch. The post lists more than 150 supported model architectures, support for the mainstream speculative decoding methods, distribution via Docker, Debian, Homebrew, conda-forge, MacPorts, Winget, and XCFramework, and CI running on Hugging Face Buckets and Jobs plus self-hosted runners provided by Nvidia, AMD, Qualcomm, Arm, IBM, and Intel. The post has no GitHub star count, download count, contributor count, or license (we did not open the repository), and the comments include a note that processing speed on very long prompts is still a challenge and a question asking "is there no Linux desktop app?" (the desktop apps the post lists are macOS and Windows).
Reading the Three Together (interpretation)
First, tokens flow along the price sheet, not the scoreboard — but we have to write down "tokens passing through which door." The 71.7% is the share of tokens passing through one router, OpenRouter. Tokens from companies contracting directly with model makers, and tokens running inside clouds, are not in this number. In NO.16 we wrote "which AI is used most — why the numbers disagree"; Ramp's corporate-card payment data (NO.47: by US corporate payments, Anthropic 44.8%, OpenAI 40.6%) and OpenRouter's token data (top four are Chinese models) are the same market in the same month seen from different doors. Both can be right — one is "who pays," the other is "which model eats the tokens." And token volume is not money (NO.43's "token volume up 37%, spending up 9%"). Taking the average prices from the Hugging Face post as-is (open $0.28/$1.62, closed $1.27/$6.36) and doing a very rough calculation, even with 71.7% of tokens going to open weights, in money terms it most likely does not exceed half (our calculation; since we do not know the input/output ratio, this is only order-of-magnitude — do not carry the number itself anywhere). In other words, "seven-tenths" is seven-tenths of volume, not of revenue.
Second, the smallest tier has become one where the price sheet is identical, so you cannot choose by price sheet. Ember, Luna, and Haiku 5.5 all have a short-prompt price of $0.10/$0.50. What differs is cost per task ($0.14, $0.15, $0.41, per AA), the long-prompt threshold (200,000, 272,000, 100,000 tokens), and the score (36, 28, 36). In NO.46 we wrote "price per token ≠ price per task" using Hydrogen and Flash; this time, three models with the same price split 3x on cost per task — because the number of tokens they spend on the same job differs, and that number comes only from independent evaluations and our own records, not from the manufacturer. That Meta's "one-third" matches the AA table is a good sign, but AA's cost per task is the value on AA's own evaluation set, which does not mean the same ratio will appear in our work (NO.21). We also note that the company's table left in the losing cell, "Haiku leads on FrontierCode" — as with Anthropic's table in NO.44, a table with a losing cell is more worth reading than one without (interpretation).
Third, the owners of the platforms producing the numbers have changed. OpenRouter, which counted the 71.7% open-weight figure, and Hugging Face, which published it, are companies that announced agreements to be acquired by a payments company and a chip company, respectively. There is no evidence that either company distorted that number, and we do not write that — but the explanation "a clean signal because it is a neutral marketplace" (the phrasing in OfficeChai's June article) now needs one more line: "a marketplace owned by a payments company." Huang's "will remain an open platform; Nvidia compute not required" is an answer to that question, and the Hugging Face post's breaking change ("you can no longer use your own provider keys") and the "you can turn providers off" toggle are a small version of the same question — when the router chooses it is convenient, but in exchange I cannot see who is choosing. On the opposite side stands llama.cpp 1.0. In NO.48 we wrote "only what you keep outside the model carries over to the next model"; the runtime that runs the model is also "something kept outside," and 1.0 is a promise that the API of that outside box will not break, at least within a major version. The three stories come down to the same three options — leave it to a router, call the manufacturer's API directly, or run it on your own computer — and the difference between the three is not price but "who chooses, where the data goes, and who fixes it when it breaks."
What We Would Do — Three Things a Team Running Commerce, Apps, and an AI Influencer Should Do This Week (practice)
- Add a "route" column to the model cost table. Next to the "cost per task" column we built in NO.46, write down which door each model is called through — manufacturer API direct / router (OpenRouter, Hugging Face Inference Providers, etc.) / local (llama.cpp) — and attach three cells to each route: who picks the provider (us or the router), which company's servers the prompt and response pass through (for routers, record the provider list), and whether we can use our own key (Hugging Face: not for language models as of November 16). For any job that uses a router, actually save the "providers turned off" list in settings and record the date. For requests containing customer data, such as ingredient-label judgments (NO.03), make it a one-line rule: do not send them through "a route where we do not choose the provider."
- Measure the three smallest-tier models on the same 20 cases. From the high-repetition jobs — short corrections to Lia's (our AI influencer's) scripts, classifying commerce inquiries, extracting ingredient-label items — take 20 cases and run them on Muse Ember, Claude Haiku 5.5, and GPT-6 Luna at the same reasoning level, recording completed cases, actual input/output token counts, cost per task, and time taken (the same table as the four cells of NO.48's "model-swap rehearsal"). If you write the three models' long-prompt thresholds (200,000 / 100,000 / 272,000) in the table header, you can see at a glance which threshold our prompts cross. AA's $0.14/$0.41/$0.15 goes in a "reference value" column — and once our numbers come in, they sit next to it.
- Pick one local 1.0 test. Spin up one small open-weight model with llama.cpp 1.0.0 on an office computer, and see whether 20 ingredient-label judgment cases involving questions mixed with personal information (allergies, medications being taken) can be processed without leaving the building — recording judgment quality, response time, and the hardware specs we can afford. Add a runtime name and version cell (llama.cpp 1.0.x) to NO.48's "box outside the model." If the results are bad, write that they are bad — the purpose of this test is not to move to local, but to have the boundary between "work that can move and work that cannot" in numbers.
What We Did Not Pick This Time
The three posts OpenAI published on November 16 — an analysis of its academic-researcher pilot (1,819 people, 580,186 messages, self-reported "would have taken 1.9x longer without AI"), expansion of its researcher program to 200,000 people (50,000 grants on November 23), and ChatGPT for Healthcare (more than 1,300 blinded physician evaluations) — we saw only through the AI-Weekly (November 17) summary, and since openai.com would not open from our side we did not use them (that the same "1.9x" figure as METR's in NO.40 appeared is a follow-up candidate). Nvidia's Q3 earnings (November 17) and reports about a CEO change are investment and personal territory, so we did not use them. The November 17 item in Anthropic's newsroom did not show a title, and the same-day news visible in search is about a listing, so we did not use it (conflict of interest, investment). Simon Willison's llm 0.37.2 (fixing a flaw where API keys leaked to other providers) is a security matter, so we leave only the title. The Fireworks AI and Notion security incidents (AI-Weekly) also title only. The City of Oakland union clause "retrain and reassign instead of layoffs when AI changes a position" (AI-Weekly) — we did not open the original. Hugging Face's addition of a "Decision Models" category (November 10) we keep as a follow-up candidate to NO.47's Decisions API. The OpenAI news list again would not open this time, and Google AI Blog and Axios continue to show stale listings.
What We Could Not Confirm in This Story
- 71.7%, 4x, 800 trillion, 10 million, 46%, 16.9%, and 16.3% are all figures OpenRouter counted from its own traffic, with no independent verification. We could not confirm why "this week's number one at 59.2T" (blog) and the rankings' 61.2T differ. The rankings' total-volume and open/closed-share numbers were not in the text.
- The Stripe–OpenRouter acquisition price is undisclosed; "a little over $8 billion" is from a Reuters source, and "more than $7 billion" is Bloomberg and the WSJ (via Wikipedia). Nvidia–Hugging Face's $12.9 billion is the announced figure, and we could not confirm whether either deal has closed. Huang's "open platform" sentence is Advanced Television quoting his blog; we did not open the original blog post.
- The Bloomberg chart's "70%→30%" is an approximation relayed by OfficeChai, and we did not see the methodology.
- Meta's benchmark numbers are the company's table, and some are internal benchmarks, specific harnesses, or partial scores. The "one-third" matches the AA table, but AA's cost per task is based on AA's evaluation set. Whether Ember's weights are released appeared nowhere. Willison's Haiku xhigh $0.25 and the AA leaderboard's Haiku Max $0.41 are at different reasoning levels, so we left both as they are.
- The llama.cpp post has no star, download, or contributor counts and no license, and we did not open the repository.
- The interpretation's "most likely does not exceed half in money terms" is our rough calculation from the average prices in the Hugging Face post, an estimate made without knowing the input/output ratio.
- We did not test or count any company's claims ourselves.
Sources We Read
We opened and checked sources of different kinds directly. We list them by company announcements and official documents, independent evaluations and data, media reports, personal blogs, wikis, and our previous articles.
Company announcements and official documents:
- huggingface.coHugging Face and OpenRouter join forces to make open-weight models easy to run (Hugging Face blog, November 16)English; routing handover, `:fastest`/`:cheapest`, cache header, own keys not allowed, open weights 26.8%→71.7%, 4x price gap, Kimi 46/3.68 vs Fable 65/26.54, 800 trillion a month, 10 million users (OpenRouter measurement)
- research.meta.aiIntroducing Our Smallest, Most Efficient Model: Muse Ember (Meta AI, November 9)English; $0.10/$0.50, 1M context, company benchmark table (Haiku leads on FrontierCode), "one-third the cost of Haiku," safety evaluation, no mention of open weights
- huggingface.coHot llama.cpp v1.0.0: A stronger foundation for the local AI ecosystem (ggml-org, Hugging Face blog, November 13)English; semantic versioning, 150+ architectures, distribution channels, CI runners from six hardware companies
Independent evaluations and data:
- openrouter.aiOpenRouter Rankings — weekly token usage (through November 16)English; top 10 models and token counts (checked directly on November 18)
- artificialanalysis.aiMuse Ember — Artificial Analysis model pageEnglish; Intelligence Index v5.0.1 36 (max) / 34 (xhigh), $0.14 per task, 449 tokens/second, class median
- artificialanalysis.aiArtificial Analysis Intelligence Index leaderboardEnglish; Fable 5.2 65/26.54, Kimi K3.1 Pro 46/3.68, Haiku 5.5 36/0.41, Ember 36/0.14, Luna 28/0.15, DeepSeek V4.1 Flash 1017 33/0.60 (checked directly on November 18)
Media reports:
- thestar.com.myPayments firm Stripe to buy marketplace OpenRouter in AI push (Reuters, published by The Star, August 20)English; August 19 agreement, price undisclosed (source: $8 billion+), 10 trillion tokens a day, Collison: "tokens are the central currency"
- fortune.comIt's official: Nvidia to acquire Hugging Face for $12.9 billion (Fortune, September 4)English; announcement stage, Delangue remarks, 18 million users, 200,000 companies (company figures)
- advanced-television.comNvidia acquires Hugging Face for $12.9bn (Advanced Television, September 4)English; $11.9 billion plus $1 billion in employee incentives, Huang: "will remain an open platform; Nvidia compute not required" (quoting the blog)
- officechai.comShare Of US Models Being Used On OpenRouter Has Collapsed From 70% To 30% (OfficeChai, June 26)English; cites the Bloomberg chart (OpenRouter, Exponential View), approximate, "neutral marketplace" phrasing (background)
Personal blogs:
Wikis:
- en.wikipedia.orgOpenRouter (Wikipedia)English; founded 2023, owner Stripe, Bloomberg/WSJ "$7 billion+," company claim of 400 trillion tokens a month, 10 million MAU
- en.wikipedia.orgHugging Face (Wikipedia)English; founded 2016, Nvidia acquisition agreement of $12.9 billion reported 2026-08-26 (closing unconfirmed), declined Nvidia's $500 million in 2025
Our previous articles:
- apholdings.krThree Documents That Write Down "What Is Allowed" Before the Agent Arrives (AP News NO.50)
- apholdings.krSwitching Models Takes Weeks — Only What You Keep Outside the Model Carries Over (AP News NO.48)
- apholdings.krWeekly AI News — Google Changed Three Models and Search in Five Days, and OpenAI Handed Its Agents a Wallet (AP News NO.47)
- apholdings.krA "Daily Driver Model" Tier Has Appeared — How to Read It by Cost per Task (AP News NO.46)
- apholdings.krTwo Doors to the Same Model — Fable 5.2 and Mythos 5.2 (AP News NO.44)
- apholdings.krCustomers Cannot Bring Their Own Agent (AP News NO.43)
- apholdings.krWeekly AI News — Three Models of 2.8 Trillion, 280 Billion, and 594 Billion Parameters Released or Announced Their Weights (AP News NO.41)
- apholdings.krCheck the Counting Method Before Comparing Numbers (AP News NO.21)
- apholdings.krWhich AI Is Used Most — Why the Numbers Disagree (AP News NO.16)
- apholdings.krWhen AI Reads the Ingredient Label for You (AP News NO.03)