Data

Proprietary models converge. Open-weights models don't.

When a model's maker sets a list price, the hosts that resell it follow that price and the market converges: 6 of the 8 models served by two or more hosts in this dataset cost the same wherever you run them (within 1.15×). When there is no list price to follow, because the weights are open and the maker does not sell them itself, each host charges what it likes and there is no pattern: 2 models show a real spread, the widest being Llama 3.3 70B Instruct at 10.4×, and 2 have no reference price at all. The list prices were verified on October 1, 2026 against the official pages of DeepSeek, Moonshot AI, Z.ai, Alibaba and Minimax. The comparison below covers 26 verified offerings of 13 models across 5 hosts.

Offerings verified between September 30, 2026 and October 1, 2026 against each host's official pricing page. Prices are USD per million tokens. How this is checked

Looking for first-party API prices (OpenAI, Anthropic, Google…)? See the main pricing table →

Where the price is the same

6 models where every host charges within 1.15× of the cheapest. Pick on latency, region or tooling; the price will not decide it.

• DeepSeek V4 Pro 0813 $1.32 in · $3.96 out
  Together AI · OpenRouter

  OpenRouter's $1.32 / $3.96 is the maker's list rate, which applies only during the maker's peak hours (Monday to Friday 01:00-04:00 and 06:00-10:00 UTC: 35 of the 168 hours in a week, 21%); the rest of the time that host, like the maker's own API, charges half, $0.66 / $1.98. Together AI charge $1.32 / $3.96 at all hours.

• GLM-5.3 $1.40 in · $4.40 out
  Together AI · Fireworks AI · OpenRouter
• gpt-oss-120b $0.15 in · $0.60 out
  Together AI · Fireworks AI · Groq
• Kimi K3 $2.85–$3.00 in · $14.25–$15.00 out 1.1×
  DeepInfra · Together AI · Fireworks AI · OpenRouter
• MiniMax M3 $0.30 in · $1.20 out
  Together AI · Fireworks AI · OpenRouter
• Qwen3.8 Max $2.00 in · $6.00 out
  Together AI · Fireworks AI

Where it still pays to shop

2 models where the most expensive host charges more than 1.15× the cheapest for identical weights. Sorted by input price, cheapest first.

Models with a single host

7 models below have a single host; for 2 of them (DeepSeek V4 Flash 0731 and Qwen3.8 27B) there is no list price to show. There is nothing to compare here; the question is whether you are comfortable depending on a single vendor for that model.

DeepSeek V4 Flash 0731 is open-weights and has no model maker's list price on OpenRouter: 27 resellers served it on September 15, 2026, at prices that neither converge nor carry a struck-through list price, so there is no reference figure to publish. Ask for :floor to get today's cheapest. Qwen3.8 27B is open-weights and has no model maker's list price on OpenRouter: 16 resellers served it on September 15, 2026, at prices that neither converge nor carry a struck-through list price, so there is no reference figure to publish. Ask for :floor to get today's cheapest.

Recently retired

10 offerings that were in this table and no longer can be: delisted, moved behind an enterprise contract, or no longer on the host's official price list. Hosts remove these quietly; we keep the record.

• DeepSeek V4 Flash 0731 on Fireworks AI last $0.22 in · $0.66 out

  Retired September 30, 2026 — No longer listed on Fireworks' pricing page (docs.fireworks.ai/serverless/pricing), read on 2026-09-30. The page says models not listed individually are priced by size; this site does not know whether the model is still served.

• DeepSeek V4 Pro 0813 on Fireworks AI last $1.32 in · $3.96 out

  Retired September 30, 2026 — No longer listed on Fireworks' pricing page (docs.fireworks.ai/serverless/pricing), read on 2026-09-30. The page says models not listed individually are priced by size; this site does not know whether the model is still served.

• Kimi K2.6 on Fireworks AI last $0.95 in · $4.00 out

  Retired September 30, 2026 — No longer listed on Fireworks' pricing page (docs.fireworks.ai/serverless/pricing), read on 2026-09-30. The page says models not listed individually are priced by size; this site does not know whether the model is still served.

• gpt-oss-120b on DeepInfra last $0.037 in · $0.17 out

  Retired September 6, 2026 — Not listed on the official DeepInfra pricing page; previous figure came from OpenRouter.

• gpt-oss-20b on DeepInfra last $0.03 in · $0.14 out

  Retired September 6, 2026 — Not listed on the official DeepInfra pricing page; previous figure came from OpenRouter.

• MiniMax M3 on DeepInfra last $0.28 in · $1.10 out

  Retired September 6, 2026 — Not listed on the official DeepInfra pricing page; previous figure came from OpenRouter.

• Llama 3.3 70B Instruct on Groq last $0.59 in · $0.79 out

  Retired September 6, 2026 — Moved to Groq Enterprise; no public per-token price on console.groq.com/docs/models.

• Llama 3.1 8B Instruct on Groq last $0.05 in · $0.08 out

  Retired September 6, 2026 — Moved to Groq Enterprise; no public per-token price on console.groq.com/docs/models.

• Qwen3.8 Max on OpenRouter last $2.00 in · $6.00 out

  Retired September 6, 2026 — No longer listed in the OpenRouter model catalog (checked via /api/v1/models).

• DeepSeek R1 on Replicate last $3.75 in · $10.00 out

  Retired September 6, 2026 — Only per-token LLM on Replicate; a single offering has no comparative value in this table.

Verification

Every offering above was read on its host's official pricing page; dates range from September 30, 2026 to October 1, 2026. Every price on this page is a link to the page it was read from. For OpenRouter, the price shown is the model maker's own list price on OpenRouter's providers page for that model, without promotional discounts (OpenRouter shows a discounted price with the list price struck through). It is neither what a default request pays (by default OpenRouter load-balances each request across several hosts, weighted towards the cheaper ones, so the effective price varies per request) nor the lowest price available through OpenRouter: resellers often undercut the maker, and the :floor suffix or sort: "price" gets you the cheapest of them at that moment, a figure that moves on its own as promotions end and hosts come and go (why the catalog number is not what you pay). Open-weights models whose maker does not serve them on OpenRouter have no list price; they are listed with "no list price" instead of a number. Offerings that could not be verified on an official page are not shown as prices: they are either excluded or listed under "Recently retired" (last review September 30, 2026). Grouping threshold: hosts within 1.15× of the cheapest input price count as "the same price".