OpenRouter · tokens vs dollars

WestvsChina

— — but tokens aren't dollars. Weekly (Mon–Sun, labeled by week-ending date), —.

SOURCE OpenRouter public rankings + model catalog list prices · updated 2026-10-07 · weekly Mon–Sun buckets labeled by ending date · value = token volume × estimated $/token (free tier = $0)

How to read this — coverage & bias

OpenRouter is a price-comparison API router for developers — a transparent but partial slice of demand, not the whole market. The coverage rows below are a qualitative reading of the venue, not measurements.

What it is

— — a small, developer- and agent-skewed slice of global token processing. For scale, Google reported ~3.2 quadrillion tokens/month of its own (Google I/O, May 2026).

Over-represents

price-sensitive developer / agent workloads, open-weight & Chinese models (DeepSeek, Qwen, Kimi, MiniMax) and free tiers.

Under-represents

first-party consumer apps (ChatGPT, Gemini, Claude.ai), hyperscaler enterprise (Bedrock / Azure / Vertex), regulated & gov — and Gemini's first-party volume especially; on the Chinese side, first-party apps and domestic clouds (e.g. Alibaba Cloud, Baidu Qianfan) that serve Chinese users directly.

So read it as

token share = a momentum signal on this venue, not a global share. West value share is not a measured floor: the Western first-party and hyperscaler spend this venue misses would raise it, the Chinese domestic spend it misses would lower it, and which is larger in dollars is not measured here. It is not a floor for any single vendor either: the missed spend includes ChatGPT and Gemini first-party use as well as enterprise Claude, so Anthropic's share could move either way. For a view of Western enterprise buyers, likely mark Gemini up and Chinese-open down (an inference).

Verbosity

tokens used and cost to run the same benchmark are compared in Panels 2–3 of the External cross-check below. Artificial Analysis benchmark, not OpenRouter traffic; see External cross-check below.

The inversion? Share of weekly tokens vs share of estimated value — West (blue), China (red), Other (grey)

Economic value over time Weekly estimated $ value by bloc (stacked) — toggle to token volume

Price per million tokens Estimated blended $/Mtok over time, on the selected price basis (see Methodology below). Indexed = 100 rebases to the first visible week — drag the slider to move the base.

Value attribution — indexed to 100 Estimated value = volume × price. Rebased to the first visible week (drag the slider to move the base): the gap between volume and value is the price contribution. Pick an aggregate or a single vendor.

Tracked vendors over time Eight individually tracked vendors: a fixed set, not a current ranking (blue = Western, red = Chinese). The $ value, value-share and $/Mtok views rest on per-vendor reference prices set when each vendor was first observed, so gaps between vendors in those views are not reliable. Switch metric, or filter West / China

How fast estimated value is changing Month-over-month change in estimated economic value — trailing-4-week mean vs the 4 weeks before. The venue total is the reference line: a vendor above it is gaining share of estimated value, below it is losing share.

Read: venue-wide growth peaked at — in the week ending — — — —. At the panel's latest week (—) the venue is — month-over-month; Anthropic — (—); Google — (—). DeepSeek —.

Price basis. As-of prices value each week at the catalog prices of the vendor's models visible in that week's mix; with none visible, the week falls back to a fixed reference price, and priced weeks are held within — of it. Current prices use the reference price every week (set when the vendor was first observed, never refreshed), so growth is pure volume. The gap between the two is the price/mix contribution, and it carries information only in weeks priced from the models visible that week. Among the four vendors plotted by default, the widest gap is —: — as-of vs — current, a — price/mix — — —. The narrowest is —, — apart: its tokens moved — against — of value—. Over the long window Google's share of estimated venue value — from — to — across — complete weeks — a window this 4-week panel cannot see.

Held out. The week ending — is excluded above: —. On this same 4-week basis it reads venue —, Anthropic —. Across that week Anthropic's single-week share of estimated value went — → — (—), while its 4-week mean share — — → —. Read this panel as momentum over a month, never as the latest week; for the latest week read the value-share panels above.

Reading limits. (i) — OpenRouter can keep revising a week's totals for some time after the week ends, and a 4-week window is far more exposed to that than a 52-week one. (ii) Consecutive points share 3 of the 4 weeks in each window, so this curve is smoothed by construction and will look more trended than independent observations would. (iii) A vendor is drawn only where it held a top-9 slot across both windows — the weekly series is top-9 + others, so a vendor leaving the ranking would otherwise read as a −100% collapse and one entering as a spike. — (iv) Value here is an estimate: tokens × a per-vendor price. — Reference prices are set when a vendor is first observed and never refreshed, so they can be off in either direction. The default four are chosen on this estimate, so neither that selection nor any order among vendors is a reliable ranking. (v) OpenRouter is one venue and a small slice of global token demand; read this as momentum on that venue, not the market.

Where the Chinese tokens actually run Default serving host for each top Chinese model on OpenRouter (a fixed host snapshot from the week ending 2026-05-31, older than the weekly data elsewhere on this page) — a model's origin isn't the same as where it runs.

— Chinese infra · — Western-neutral

How OpenRouter routes: by default each call goes to the cheapest qualifying host (then load-balanced on uptime/latency, with automatic failover); a developer can override by pinning a provider or filtering on data policy / region. In this snapshot —; the open weights mean it can run anywhere, and — ran on Western-neutral clouds (Tencent & Zhipu defaulted to the US's GMICloud). Serving speed: host-level serving speeds are in the External cross-check below; for the models in this chart, which host serves them fastest is not shown on this page (Artificial Analysis provider benchmarks, not OpenRouter data; see External cross-check below). Trans-Pacific latency on agentic round-trips likely also penalises China-hosted serving for Western users — an inference, not measured here. Caveats: default/primary host only (not the full per-host split, which OpenRouter doesn't publish); OpenRouter is a small slice of demand and misses enterprise (Bedrock / Azure / Vertex), where Chinese open models plausibly run on Western clouds more often.

Deeper dive · tier structure inside one vendor

Inside a frontier vendor: Anthropic's demand by tier

The sections above compare vendors; this one looks inside one of them, Anthropic, by tier. In the month to the frontier tier (Opus/Fable) carried % of Anthropic's tokens on this venue, Sonnet % and the low-price Haiku tier %. Denominator throughout = Anthropic's own tokens only (not all-vendor token share).

A · Volume share within Anthropic, over time (Opus+Sonnet basis) Share of Anthropic's top-9-ranked Opus+Sonnet tokens on OpenRouter that are frontier (Opus/Fable) vs Sonnet, by week.

Frontier — Opus / FableMid — Sonnet

Read: frontier share from ~% of Anthropic's ranked tokens to by the week ending , the last week this top-9-constrained series resolves (see the series-basis note below). The weekly series is volatile (range ~–%, trailing-4-week mean ~%), so its endpoint is a single week, not a settled level. That last week reads % frontier and the month to reads % on Panel B's full per-model split — a pp gap. The two readings differ in period and in coverage: the weekly figure counts only the Anthropic models that cleared the global top-9 that week (), while the monthly figure counts every Opus, Fable and Sonnet model. Real change after that week, the one-week-versus-one-month comparison and the difference in coverage can each contribute to the gap. Not observed here: first-party and enterprise demand; whether its tier mix is more or less frontier-heavy than this venue's is not measured.

B · The full tier stack — latest month (Opus/Fable split out) Every Anthropic model's tokens for the month to , rolled to four tiers. Unlike Panel A, this per-model split is not limited by OpenRouter's top-9 weekly cutoff, so it resolves Haiku and Fable.

Read: the frontier tier (Opus+Fable) is of Anthropic's tokens on this venue in the latest month, and within it Fable carries % of all Anthropic tokens, vs Opus's % — . Caveat:

C · List price by tier, over time Each tier's successive models at their current listed OpenRouter price, placed at release date. A version's list price can change after launch, so this is not a launch-price history. Standard variants only (priority/“fast” and batch SKUs excluded). Log scale.

OpusFableMid — SonnetLow — Haiku

Read: Caveat:

… and the contrast · OpenAI

Inside OpenAI: — and the mix blends to $/M

Run the same cut on OpenAI: its routed mix reads — the frontier bucket is % of OpenAI's routed tokens against the cheap + open-weight (gpt-oss) tail at ~% — . The buckets follow the product line — not the generation number in the model name, and not the list price. A channel statement, not a firm one — the venue can't see first-party demand, so this is what routers see, not OpenAI's total mix. Denominator = each vendor's own tokens. Assigning lines to tiers is a judgment call.

OpenAI tier stack — latest month Every OpenAI model's tokens for the month to , rolled to tiers (text only), with the generation carrying the most frontier tokens split out of the frontier bucket. Chips show individual models (~% of OpenAI tokens).

Read: frontier is % of OpenAI's routed tokens against the cheap + open-weight tail at % — . (Anthropic's shape is in Panel B above, Google's in the three-way stack further down; the tiers are within-vendor and not equivalent, so each reads against its own denominator.)

The frontier bucket is split by generation — the one carrying the most frontier tokens. (, released ) puts % of OpenAI's routed tokens into that bucket — % of it. Read the two panels carefully though — Fable is a premium SKU added alongside Opus, whereas is a whole model generation, not one added SKU. A new-generation ramp and a new-SKU ramp are not the same quantity.

OpenAI list price by tier, over time Representative model per tier per generation, at current listed OpenRouter price, placed at release date. A version's list price can change after launch, so this is not a launch-price history. Log scale.

Read: Where its third-party volume sits is a separate question (the OpenAI tier stack above):

… and the capstone · Google & the three-way

Inside Google: the workhorse, not the frontier — three vendors,

Google is : . The Flash workhorse + Flash-Lite are ~% of its routed tokens, Gemma ~%, and the frontier Pro tier only ~%. Side by side, . On the narrower frontier-vs-tail rule, OpenAI's mix reads . Every share in this section comes from one developer- and agent-heavy third-party router. First-party demand is not visible here, so whether it skews more to the frontier tier is not measured.

D · Three vendors, one venue — tier stack by vendor, latest month Each vendor's models for the month to , rolled to tiers (text only). Five slots, so each bar sums to 100 — legacy is a live slot for OpenAI only. Read the shape; tiers are within-vendor, not equivalent across vendors.

Read: Google's frontier (Pro) is just of its tokens — vs Anthropic , OpenAI ; its Flash workhorse alone is . Priced out (each vendor's routed mix × current list input price): , so ’s mix lists at ~× ’s.

E · Google list price by tier, over time Representative model per tier per generation, at today's listed OpenRouter price (not necessarily its price at launch), placed at release date. Log scale.

Read: Blended prices for all three vendors are in Panel D above.

External cross-check · capability & efficiency

The other half of the picture: what the dollars buy

The panels above are OpenRouter traffic. Panels 1–5, 2b, 2c and 7–9 here are EXTERNAL — Artificial Analysis benchmark + METR data, not OpenRouter’s: capability, token efficiency, cost to run a benchmark, cost and time per task, serving speed, task length and capability over time. Panel 6 pairs Artificial Analysis scores with OpenRouter’s realized input prices. They test one reading of the token/dollar gap — that a low per-token price overstates the cost advantage once tokens per task and capability are counted. In the 25 Jun 2026 Artificial Analysis snapshot the Index band ; Panels 2b–2c and 6 use later snapshots, where index levels differ. US/West blue, China red throughout (Panel 7’s optional “Colour by lab” view keeps the bloc as a coloured ring). Panels 1–4: reasoning/max-effort variants unless noted (Tencent Hy3 is AA's non-reasoning measurement); Panels 2b–2c use Artificial Analysis’ default effort setting for each release; Panels 7–9 use whichever variant AA plots for each model (reasoning, effort-tiered and non-reasoning mixed), and Panels 8–9 compare the best point in the selection regardless of effort setting. Panels 1–4 are fixed at 25 Jun 2026; Panels 2b–2c are a separate manual Artificial Analysis snapshot of (cost and decode time per Index task); panel 5 is METR's published data; panel 6 pairs Artificial Analysis scores with OpenRouter realized input prices as of . Panels 7–9 are a third Artificial Analysis snapshot, of (Intelligence Index over model release date, and the Western-vs-Chinese gap and catch-up lag read off it); the series is the chart as plotted on that date, and Artificial Analysis’ leaderboard may already list newer releases.

1 · The frontier ceiling — as of the 25 Jun 2026 snapshot, no Chinese model reached it Artificial Analysis Intelligence Index (composite, ~0–70). Higher = more capable. Artificial Analysis snapshot, 25 Jun 2026 — Panels 2b–2c and 6 below use later snapshots, where index levels differ.

Read: in this snapshot China tops out at Index 51 (GLM-5.2, Jun 2026; next Qwen3.7 Max 46); the US holds the entire 54–60 tier (Fable 5 60, Opus 4.8 56, GPT-5.5 55, Opus 4.7 54). As of 25 Jun no Chinese model had crossed into that tier, so the cost and price comparisons below are undefined at the top.

2 · Cost to finish the work vs capability $ to run the full AA Intelligence Index (log) vs the index score. Down-and-right = cheap and capable.

Read: In the 25 Jun 2026 snapshot, China was cheaper on list price but topped out below the US band; the US listed higher prices and reached it. On one consistent basis — reasoning-model medians — . It inverts at matched capability (Index 46: Qwen $25/index-pt > Gemini $18). Scope: those medians are reasoning-only — the plotted Tencent Hy3 point is AA's non-reasoning measurement and is excluded from them; and GLM-5.2 (Index 51, China's ceiling in Panel 1) has no AA cost-to-run in the 25 Jun 2026 snapshot, so the China set plotted here tops out at Index 46. Vintage: cost-to-run was mid-flux at the snapshot (DeepSeek price cut + AA re-score moved V4 Pro $1,071→$176 and Index 52→44 at unchanged 180M tokens) — the stable metric is Panel 3's tokens-to-run.

3 · Verbosity — tokens to complete the same task Output tokens (millions) generated to run the Intelligence Index, reasoning models only. Dashed line = M, the median of the models shown. GLM-5.2 (Index 51) is absent — AA published no tokens-to-run for it in the 25 Jun 2026 snapshot.

Read:

2b · Cost per task vs capability Artificial Analysis’ weighted-average cost per Intelligence Index task (USD, log scale) vs Intelligence Index. Releases: . Up-and-left is better. The Pareto line and shaded band () are this page’s definitions, not Artificial Analysis’. Different basis from Panel 2 (cost to run the whole Index) and Panel 6 (OpenRouter realized input price per 1M tokens, as of ).

Colour by lab

2c · Decode time per task vs capability Artificial Analysis’ weighted-average decode time per Intelligence Index task (minutes, log scale, excluding time-to-first-token) vs Intelligence Index. Releases: . Up-and-left is better. The Pareto line and shaded band () are this page’s definitions, not Artificial Analysis’. Decode time mixes verbosity with the speed of the host measured — a different basis from Panel 4 (per-host speed) and Panel 5 (METR task length).

Colour by lab

4 · Serving speed by provider — median vs fastest host, per model Output tokens/sec per model: median provider → fastest provider. Bar length = the spread between the two. Colour = model origin, not host origin.

median providerfastest provider model's own first-party API

5 · Mission length — how long a task the model can carry METR 50% time horizon: the length of task, in human-expert time, that the model completes with 50% success, on METR's mostly software, ML and cybersecurity tasks. Endurance, not speed — panel 4 is speed. Source: METR time horizons, not Artificial Analysis (). Each group is marked with the METR task suite it was measured on; the two suites are not directly comparable.

Read:

6 · The current frontier — price vs capability, Pareto tradeoff AS OF · OpenRouter benchmarks Weighted-avg realized input price (OpenRouter's own traffic, $/1M tokens) vs Artificial Analysis Intelligence Index, for every AA-scored model OpenRouter prices — models in this snapshot, . The universe is set by AA's scoring coverage, not OpenRouter's traffic ranking — read it as a capability cut, not a traffic cut. Every model scoring Index or better is named, plus every Pareto point. The view crops to Index ≥ to keep that band legible — weaker models fall below the floor and are off-view, . Score is AA's external methodology (same family as Panel 1 above); the price axis here is OpenRouter traffic data, not AA's synthetic cost-to-run. Taken from OpenRouter's benchmark rankings as of , not the 25 Jun 2026 (Panels 1–4) or (Panels 2b–2c) snapshots.

Show Pareto frontier & names

7 · Intelligence over time — each model, by release date Artificial Analysis Intelligence Index (y) against model release date (x), for . Bold lines are each bloc’s frontier; the larger dots are the models that set it.

Colour by lab

8 · The gap — top Western model vs top Chinese model, monthly

Index points

Definitions and caveats

9 · The time gap — how long ago the West first reached each Chinese frontier score

Look forward from Western frontier points

Definitions and caveats

SOURCE Panels 1–4: Artificial Analysis (Intelligence Index, tokens and cost to run the Index, per-provider output speed; model and provider pages, 25 Jun 2026) · Panel 5: METR time horizons () · Panel 6: Intelligence Index scores by Artificial Analysis with OpenRouter realized prices, via OpenRouter rankings, as of · Panels 2b–2c: Artificial Analysis (artificialanalysis.ai), manual snapshot · Panels 7–9: Artificial Analysis (artificialanalysis.ai), snapshot of , Intelligence Index

External data · compute & token prices

What a GPU-hour and a token cost: compute and token prices

The series here () are third-party indexes published by Silicon Data · SemiAnalysis · Ornn AI, not OpenRouter’s data: what it costs to rent a GPU for an hour, and what a million tokens cost. Panels A and B show each kind on its own axis; Panel C sets a GPU rental family against the token price on one rebased chart, and Panel D divides one by the other. They measure different things and start on different dates, so the table below says what each index is and each panel says what it can and cannot support. The series are a manual snapshot pulled , not a feed. Colour here means the data provider only (olive Silicon Data, grey SemiAnalysis, violet Ornn) and has no link to the vendor or region colours used elsewhere on this page.

How these indexes differ One row per index, as each provider describes its own method. The dates and latest values are read from the data on this page.

Index	What it measures	Method	Clouds covered	Unit	First date here · public launch	Latest value

A · GPU rental prices What it costs to rent one GPU for an hour, USD per GPU-hour, one axis. Colour = provider; dashed = A100. SemiAnalysis’s one-year contract price is published at the start of some months only, so it is drawn as a step held until the next print.

Read:

Definitions and caveats

B · LLM token prices USD per 1M tokens, one axis, for the two token price indexes. Thin line = daily value, bold = trailing 7-day median. They are different baskets (a usage-weighted blend of a defined set of models and what Ornn labels a realised Anthropic price), so read each against its own history, not against the other’s level or against list prices.

Unit check: Silicon Data’s unit is checked against its own published level; Ornn’s is taken from Ornn’s documentation and the series name, with no published level to check it against.

Read:

What the unit rests on, and caveats

C · GPU rental family against the token price, rebased to 100 The bold black line is the equal-weight average of a GPU rental family, each index a trailing 7-day median divided by its own value on the base date. The bold olive line is the Silicon Data token index (USD per 1M tokens, market-wide) on the same rebasing; the thinner violet line is Ornn’s realised Anthropic token price. The shaded area is the gap between the GPU family and the Silicon Data token line. Rebasing removes units, not differences in what is measured.

Base date GPU family

Read:

Definitions and caveats

D · GPU-hour rent divided by token price derived, illustrative How many million tokens, sold at the token price, it would take to cover one GPU-hour of H100 rent: H100 rent (USD per GPU-hour) ÷ token price (USD per million tokens), each H100 index paired with the same provider’s token index, both as trailing 7-day medians. It is a ratio of two indexes, not a serving throughput, a margin or a profit measure.

Read:

Definitions and caveats

SOURCE Silicon Data (GPU rental indexes, LLM token expenditure index) · SemiAnalysis (H100 GPU price indexes) · Ornn AI (GPU compute and token price indices). Daily provider indexes, a manual snapshot pulled , not a live feed; each series ends on its own last observation. The 7-day medians, rebasing and the rent-to-token-price ratio are this page’s calculations, not the providers’.

Methodology, assumptions & caveats

Estimates built from public data; dollar figures and value shares are modelled and should be read as indicative, not exact — the pricing rules described above mean this page cannot size their uncertainty. Not investment advice.