Model Insights

Arrivals

A model that was not carrying traffic a week ago and is now near the top of the board. This section renders only while that is true of something: on an ordinary day it is not here at all, and its absence is the reading. The rule that admits a model is fixed in the collector and travels with every row it writes, so a finding stays interpretable after the rule changes.

NOV 10, 2026 · ENTERED THE TOP 20 · AGAINST 7 DAYS EARLIER · UNDER 1% OF TODAY · RULE V1

Where the tokens go

One day of traffic across every model the gateway served, re-drawn from each sweep. Concentration is the standing story of this chart: attention is rarely evenly spread, and how top-heavy the day is tells you as much as who is on top. Every model name links to its page on the gateway.

NOV 11, 2026 · 33.2T TOKENS · TOP MODEL 29% · TOP FIVE 56.1% · TOP TWENTY 81.7%

THE SAME DAY, ORDERED BY REQUESTS

The same gateway, the same day, counted by how often each model was asked instead of how much it wrote. The orderings are not the same list in a different sequence: a model answering short questions constantly can sit near the top here and well down the chart above, and only one of those two positions describes how many decisions it was actually trusted with.

One model on the gateway is a System One model — whose product is a typed decision with a probability rather than prose. It took 9.19% of the day’s requests and 1.21% of its tokens, which is the whole point of the class: the work is in the answer, not the wordcount. One model carries the judgment; it is declared from each model’s own description, never inferred from its traffic.

Rising and falling

A model’s share of the latest day against its share of the trailing month, both measured by this instrument in the same sweep. A line sloping up means the model took a bigger cut of the day than its month average — attention arriving, not just volume. Models below 100B tokens for the month are excluded; a multiple on a tiny base is noise wearing a percent sign.

The agentic tell

Tokens divided by requests. A person typing questions spends a few thousand tokens a turn; an agent working through a codebase spends tens of thousands. The ratio separates the two without anyone declaring which they are.

Who is actually spending it

The gateway names 20 applications behind its traffic, ranked by its own numbering — which runs to 29: where the ranks skip a number, either the gateway published no application for that slot or this site chose not to list the one it named, so a gap is not a missing row. A marks an application whose repository this site’s Repos instrument follows — one instrument watches who builds a thing, the other watches whose tokens it spends, and this column is where they meet. Both are read on the same daily cadence, so a repository that starts drawing tokens today is marked today — and a is one followed daily that has not yet cleared the floors the Repos panel publishes, so there is no page to send you to for it yet.

In the current reading, 16 of the 19 are repositories the Repos instrument follows — claude-code, hermes-agent, codex, kilocode among them, and 11 of the 19 are command-line or IDE agents rather than chat interfaces. 7 of those 16 appear on the published panel today; the rest are followed daily but sit below the floors that panel publishes, and reach it at its next monthly rebuild.

The same lab at two doors

Red is the second gateway’s share of tokens; gray is the share computed from this instrument’s own daily sweep of the first. A row reading no join is a name that does not match across the two sources, shown rather than hidden — dropping it would quietly flatter the overlap.

The widest disagreement in the current reading is xiaomi: 0.5% of one gateway’s tokens against 2.5% of the other’s, a 5× difference in how big the same lab looks depending on where you stand.

THE SAME GATEWAY, MODEL BY MODEL, ON BOTH AXES

One level down from the labs above, and counted two ways: how often each model was asked, and how much it wrote. A lab-level chart cannot express this, and neither can a token-only one — a model answering a constant stream of short questions takes a large share of the first bar and a small share of the second. The distance between a model’s two bars is what the pair is for.

The widest gap is Jev: 17.6% of this gateway’s requests against 2.1% of its tokens, an 8× difference in how big it looks depending on which axis you count. Measured on one axis it is a rounding error; measured on the other it is near the top of the board.

Neither gateway is the market. Each is a customer base, and the divergence between them measures who routes where — it is not an error to be reconciled away. Neither sees traffic sent straight to a lab’s own API, which is where most frontier revenue lands. The two shares also cut differently: the second gateway’s is its text traffic, while this instrument’s sweep counts everything the first gateway served that day. And a corporate umbrella is folded to the lab that owns it — a model line published under its own name counts toward its parent on both sides.

The tape, day by day

A second gateway publishes a fixed range of daily history, which is the one thing this instrument can go back and get rather than wait for. Share of tokens above, share of spend below, per lab, every day of the range. The lab leading the latest day is drawn in red so the crossings stay readable; the rest are gray on purpose.

SHARE OF TOKENS · 408 DAYS · OCT 1, 2025 — NOV 12, 2026

SHARE OF SPEND · SAME RANGE

The daily sweep the rest of these readings rest on is forward-only: its source serves a rolling window and sells no history at any price, so its record begins the day collection began. This range is re-read whole every morning and arrives in a single request. The two are not interchangeable, and that difference is why both are here. Each chart names the seven largest labs by its own metric on the latest day, so the two lists can differ.

Volume against money

Every lab on one line: where its dots sit is its share of tokens against its share of spend, and the bar between them is the finding. A lab whose red dot sits far to the right of its gray one is being paid a premium for every token it serves.

On Nov 12, 2026 the biggest source of tokens was deepseek at 25.0% of volume, while the biggest source of money was anthropic at 35.5% of spend on 11.3% of the tokens. Per token routed, anthropic collects 3.16× its share of the bill and deepseek collects 0.22× — a 14× difference in what a token is worth depending on whose it is.

The same model is not one price

A model is served by many providers at once, and they do not agree. Each bar spans the cheapest and dearest dollars-per-million-prompt-tokens across every provider serving that exact model. Free and disabled routes are excluded — a $0 route is a promotion, not a price.

Widest spread in the current reading: z-ai/glm-5.2 spans $0.01 to $2.31 across 24 providers — 350× for identical weights, relace cheapest and alibaba dearest.

And it is not only which company — which tier

openai/gpt-6-astra · every openai offer on the gateway, Nov 12, 2026

One vendor, identical model — and a 12× spread on service tier alone in the current reading. This distinction is the reason the sensor records every offer rather than a headline price. The exemplar above is picked by rule: whichever single provider shows the widest tier spread on the day, not an editor’s favorite.

The price of privacy

Suppose your prompts must not be kept. A provider counts as clean here only if it publishes a data policy declaring it neither trains on nor retains what you send — so each model’s real price becomes the cheapest clean offer, not the cheapest offer. For the highest-scoring models: the filled dot is the cheapest route from anyone, the hollow dot the cheapest clean route, and the gap between them is what privacy costs that day.

In the current reading, privacy costs nothing for 27 of the 42 scored models with a clean route — the clean provider already is the cheapest.

What a point of intelligence costs

Price alone says nothing without a denominator. Each model’s cheapest real price is plotted against an independent capability score (Artificial Analysis’s intelligence index) — the down-and-right edge is the efficiency frontier, and over a series this chart becomes the answer to the only question that matters about inference cost: is that frontier moving, and how fast.

The cheapest point of capability in the current reading is deepseek/deepseek-v4.1-flash-1017 at $0.0008 per index point; the highest-scoring model, anthropic/claude-fable-5.2, costs $0.124 per point — a 155× difference in what a unit of measured capability costs.

The same chart, by the job

“Smartest per dollar” is rarely the actual question — a developer wants the smartest at their job per dollar. The same frontier, re-plotted against the independent coding and agentic scores. The solid mark is the best value on each.

In the current reading, one model is the value pick for both jobs: qwen/qwen3.8-27b.

The speed frontier

For anything interactive, tokens per second matters as much as the score. Best measured throughput against intelligence; the labeled edge is the set of models where nothing faster is smarter. Throughput is measured on each model’s default route only, so a model can run faster than shown from a different provider — never slower than a measurement that exists.

The fastest model in the smarter half of the field, in the current reading: meta/muse-spark-1.3 at 180 tokens per second.

What a window costs

The cheapest priced route at each context-window tier — for when the question is not how smart the model is, but how much it can hold.

The free tier

Free is technically the best price, so price stops being an axis. What replaces it is what the zero costs: every scored model the gateway serves at $0, plotted by intelligence against the worst measured string its serving providers carry — from no declared strings to training on the prompts you send. The solid mark is the highest score in the cleanest occupied column.

“Free” has a price, and the sensor measures it. In the current reading, 2 of the 2 scored free models are served by a provider that trains on your prompts, and 2 of the 2 by one that retains them. Free routes are promotions, not commitments — they can be withdrawn, rate-limited, or deprioritized without notice. And a pre-release running under a placeholder label carries no independent score, so it cannot appear here no matter how capable it looks.

Can you actually download the weights?

Every model carrying traffic on the daily sweep is probed against the public model hub. The test is not whether a repository exists with a matching name. It is whether that repository contains downloadable weight files, published by the model’s own author. A model the probe has not yet reached is counted in its own row, never assumed.

TRAFFIC DAY NOV 11, 2026 · WEIGHTS PROBED NOV 12, 2026

60.2% of the day’s tokens ran on models whose authors publish downloadable weights, across 157 models. Only that top tier is a claim. The rest are degrees of not-knowing, kept visible rather than rounded into the headline — a model with no weights repository found is recorded as not found, never as closed.

The first version of this test matched on name alone and reported more than double whatever the top tier reads. Matching by name declared closed frontier models open, because someone had registered a repository wearing each name with nothing inside it.

Repositories wearing a name

Some probed models match a public repository belonging to someone other than the model’s author, whose name normalizes to theirs, and which contains no weights at all. They are kept rather than discarded: a repository carrying a closed frontier model’s name is a finding in its own right.

In the current reading that is 26 models, carrying 3.1% of the day’s tokens by the traffic of the models being impersonated; the table shows the 14 busiest by that traffic. A model’s own author publishing a card-only repository is a different thing entirely, and is counted in its own tier above rather than here.

And the harder case — 23 models whose weights were found, but not published by the model’s author; the table shows the 8 most-downloaded of them. Some are genuine, an organization whose name simply differs. Some are unrelated models that happen to share a word. This is why author-verification, and not the mere presence of weights, is what the headline is allowed to rest on.

Where open-weight models come from

Open-weight releases per year among notable models, US-affiliated against China-affiliated. Affiliation is recorded per author, so a model with authors in both places counts for both — these are two counts, not a split of one.

Across the complete years, US-affiliated open-weight releases peaked at 37 in 2023 and stood at 11 in 2025, the last complete year. Chinese-affiliated releases went from 13 to 29 across the same span. China first published more open-weight notable models than the US in 2025 (29 to 11).

From a curated registry of 1,092 notable models, 838 of which carry an explicit open-weights verdict. It is deep on provenance and thin on the live commercial frontier — a model absent from it is outside its notability criteria, not unreal. The final year on the chart is partial by definition: it is still being published, which is why every sentence above stops at the last complete year.

What a parameter costs

Only author-published open-weight models appear here: they are the only ones whose true size is knowable. Circle area is that model’s measured token volume and both axes are logarithmic. Labels go to the largest by traffic, and a label that would collide is dropped rather than moved.

The cheapest route on the board in the current reading is openai/gpt-oss-20b at $0.03 per million output tokens for 20.9B parameters. The largest is moonshotai/kimi-k3.1-pro at 2780B and $8.41. Size and price are related but not tightly, and the spread across the middle of this chart is the part a headline price cannot show.

The parameter counts come from each repository’s own weight index. The gateway catalog carries a field that looks like a parameter count and is not one — its values sit between 6 and 19 across nearly every model, and anything sized from it would be wrong by nine orders of magnitude.

Where the compute sits

What these readings cannot tell you

Source: OpenRouter’s public catalog and rankings, read once a day by a collector this instrument runs on its own hardware. The gateway publishes no history — every number here exists because it was recorded on the day it was true, and a missed day cannot be bought back at any price. Figures describe traffic routed through this one gateway; traffic sent directly to a vendor never appears in it.

Coverage, measured

Coverage differs by dataset and by metric, so it is counted from the stored rows rather than asserted. Two limits are worth stating plainly next to it: the second gateway publishes shares of its own traffic and never volumes, so a share here cannot be multiplied by a token count from anywhere else and two gateways’ shares cannot be added; and the join between sources is a lossy name normalization, so where it fails the row is shown unjoined rather than dropped.

Sources and licenses

• epoch-notable-ai-models — 1,092 rows, read through Nov 12, 2026, CC-BY-4.0
• huggingface-hub-api — 591 rows, read through Nov 12, 2026, no separate data license asserted
• vercel-leaderboard-export — 153,932 rows, read through Nov 12, 2026, CC-BY-4.0

AI Gateway Leaderboard Data © Vercel, used under CC BY 4.0. Data on Notable AI Models © Epoch AI, used under CC BY 4.0. Model metadata from the Hugging Face Hub. Model volume, pricing, and endpoint data collected by this instrument.