provider price index

Kimi K3.1 Proprovider price index

Effective input, cache, and output economics for Kimi K3.1 Pro across 5 gateways. Compare the best current price from every inference provider, with ZDR status, gateway fees, discounts, and the evidence behind every number.

24 hours5-minute bucketsself-hostedOpenRouterVercel AI GatewayRequestyOpperTrustedRouter

Last update Nov 21 2026, 18:40 UTC/data age 2m ago/tracking since Nov 21 2026, 18:28 UTC

Cheapest standard$0.1764per 1M · default mixDecart · OpenRouter
Offers tracked87across 5 gateways84 Standard
Strict ZDR offers5867% of tracked offersprovider-reported
24 hours coverage1.4%4/288 buckets10m fill · gaps stay gaps
Range
Tier
Display
Coverage
1.4%4/288 · 5m bucketstracking since Nov 21 2026, 18:28 UTC
Gateways
ZDR
Usage mix always sums to 100.0%
2.5%
%
97.0%
%
0.5%
%
Total 100.0% — effective mix

Effective blend = input × 2.5% + cache × 97.0% + output × 0.5%. Fees are PAYG estimates; BYOK and enterprise contracts may differ.

offers 57

Providers by price

Best current price

5/12 selected
29 / 29 providers

Each line is the selected provider’s current best offer. Use the gateway menu to pin a different route.

Effective blended price

Price over 24 hours

5 lines1.4% coverageUSD / 1M
Updated Nov 21 2026, 18:40 UTC · data age 2m agoLive

Full snapshot · server-rendered

Every offer, with its evidence

57 offersUSD / 1M

Rows are sorted by effective blended price using the current mix (2.5% Input · 97.0% Cache · 0.5% Output). List price is reconstructed from the gateway discount; peak badges use the UTC window at collection time. TPS is shown only when the source reports a positive p50.

Every offer, with its evidence · USD/1M
#PlotProviderGatewayAttributesListDiscountQuotedPeakEffective blendInputCacheOutputTPS p50Latency p50Uptime
01★DecartOpenRouterZDRmxfp4$0.294-40%$0.1764—$0.1764per 1M tokens$1.2$0.12$658Model page · 30mn/a30m100.00%30m · 1d 100.00%
02★WaferOpenRouterZDRUSmxfp4$0.2249—$0.2249—$0.2249per 1M tokens$0.253$0.18$8.884Model page · 30mn/a30m100.00%30m · 1d 99.93%
03★MorphOpenRouterZDRfp8$0.2335—$0.2335—$0.2335per 1M tokens$1.306$0.163$8.5576Model page · 30mn/a30m100.00%30m · 1d 100.00%
04★RelaceOpenRouterZDRfp4$0.242—$0.242—$0.242per 1M tokens$0.2403$0.2$8.40236Model page · 30mn/a30m100.00%30m · 1d 100.00%
05RelaceVercel AI GatewayZDR$0.242—$0.242—$0.242per 1M tokens$0.2403$0.2$8.40264API · 60m1,354 ms60m100.00%60m · 1d 99.94%
06★Sail ResearchOpenRouterZDRUSfp4$0.2542—$0.2542—$0.2542per 1M tokens$1.043$0.187$9.3563Model page · 30mn/a30m100.00%30m · 1d 100.00%
07Sail ResearchOpenRouterZDRfp4$0.2632—$0.2632—$0.2632per 1M tokens$1.402$0.187$9.3568Model page · 30mn/a30m100.00%30m · 1d 100.00%
08WaferOpenRouterZDRmxfp4$0.2646—$0.2646—$0.2646per 1M tokens$1.8$0.18$988Model page · 30mn/a30m100.00%30m · 1d 99.9%
09WaferVercel AI GatewayZDRUS$0.2646—$0.2646—$0.2646per 1M tokens$1.8$0.18$9126API · 60m727 ms60m100.00%60m · 1d 99.98%
10★AkashMLOpenRouterZDRfp4$0.2684—$0.2684—$0.2684per 1M tokens$1.75$0.18$1042Model page · 30mn/a30m100.00%30m · 1d 100.00%
11★DeepInfraOpenRouterZDRbf16$0.294-5%$0.2793—$0.2793per 1M tokens$1.9$0.19$9.550Model page · 30mn/a30m95.2%30m · 1d 98.4%
12★DigitalOceanOpenRouterZDR$0.2793—$0.2793—$0.2793per 1M tokens$1.9$0.19$9.583Model page · 30mn/a30m100.00%30m · 1d 100.00%
13★InferenceNetOpenRouterZDRUSmxfp4$0.2821—$0.2821—$0.2821per 1M tokens$1.8$0.198$993Model page · 30mn/a30m100.00%30m · 1d 100.00%
14★MakoraOpenRouterZDR$0.2915—$0.2915—$0.2915per 1M tokens$2$0.2$9.4948Model page · 30mn/a30m100.00%30m · 1d 99.6%
15★Alibaba CloudVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1039API · 60m1,636 ms60m100.00%60m · 1d 100.00%
16★BasetenVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10183API · 60m916 ms60m100.00%60m · 1d 99.97%
17DeepInfraRequestyZDRbf16$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
18DigitalOceanVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10130API · 60m337 ms60m100.00%60m · 1d 100.00%
19★Fireworks AIOpenRouterZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1041Model page · 30mn/a30m100.00%30m · 1d 98.1%
20Fireworks AIRequestyZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
21Fireworks AIVercel AI GatewayZDRfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$1072API · 60m788 ms60m100.00%60m · 1d 100.00%
22★LyceumOpperZDREU$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
23LyceumRequestyZDREU$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
24★ModalOpenRouterZDRmxfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$1094Model page · 30mn/a30m100.00%30m · 1d 100.00%
25ModalVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1091API · 60m357 ms60m100.00%60m · 1d 100.00%
26★Moonshot AIRequestyZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
27Moonshot AIVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1030API · 60m1,011 ms60m100.00%60m · 1d 99.99%
28★NebiusOpenRouterZDRfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$10104Model page · 30mn/a30m100.00%30m · 1d 100.00%
29NebiusRequestyZDRUSfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
30NebiusVercel AI GatewayZDRfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$10126API · 60m522 ms60m100.00%60m · 1d 100.00%
31★Novita AIRequestyZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
32★ParasailOpenRouterZDRfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$1071Model page · 30mn/a30m100.00%30m · 1d 100.00%
33ParasailVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10209API · 60m310 ms60m100.00%60m · 1d 100.00%
34★Prima LabsOpenRouterZDRmxfp4$0.294—$0.294—$0.294per 1M tokens$2$0.2$10105Model page · 30mn/a30m100.00%30m · 1d 99.9%
35★RiverOpenRouterZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1083Model page · 30mn/a30m100.00%30m · 1d 100.00%
36★RunwareRequestyZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
37RunwareVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10124API · 60m691 ms60m100.00%60m · 1d 97.7%
38★Together AIOpenRouterZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1047Model page · 30mn/a30m100.00%30m · 1d 99.93%
39Together AIRequestyZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$10n/an/an/an/an/an/a · 1d n/a
40Together AIVercel AI GatewayZDR$0.294—$0.294—$0.294per 1M tokens$2$0.2$1045API · 60m1,592 ms60m100.00%60m · 1d 99.96%
41Sail ResearchTrustedRouterZDR$0.297—$0.297—$0.297per 1M tokens$1.583$0.211$10.55n/an/an/an/an/an/a · 1d n/a
42BasetenTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
43BasetenTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
44★TelnyxTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
45TelnyxTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
46WaferTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
47WaferTrustedRouterZDR$0.3102—$0.3102—$0.3102per 1M tokens$2.11$0.211$10.55n/an/an/an/an/an/a · 1d n/a
48MorphVercel AI GatewayZDR$0.358—$0.358—$0.358per 1M tokens$1.676$0.264$12n/aAPI · 60m500 ms60m100.00%60m · 1d 100.00%
49★Amazon BedrockOpenRouterZDR$0.394—$0.394—$0.394per 1M tokens$4$0.2$2087Model page · 30mn/a30m100.00%30m · 1d 97.9%
50★TinfoilTrustedRouterZDR$0.6203—$0.6203—$0.6203per 1M tokens$4.22$0.422$21.1n/an/an/an/an/an/a · 1d n/a
51TinfoilTrustedRouterZDR$0.6203—$0.6203—$0.6203per 1M tokens$4.22$0.422$21.1n/an/an/an/an/an/a · 1d n/a
52★TensorXOpperZDREUmxfp4$0.8775—$0.8775—$0.8775per 1M tokens$3$0.75$15n/an/an/an/an/an/a · 1d n/a
53TensorXRequestyZDREU$0.8775—$0.8775—$0.8775per 1M tokens$3$0.75$15n/an/an/an/an/an/a · 1d n/a
54★BergetOpperZDREUmxfp4$1.097—$1.097—$1.097per 1M tokens$3.297$0.9891$10.99n/an/an/an/an/an/a · 1d n/a
55★Featherless AITrustedRouterZDRcache ≈ input$2.152—$2.152—$2.152per 1M tokens$2.11$2.11$10.55n/an/an/an/an/an/a · 1d n/a
56NebiusTrustedRouterZDRcache ≈ input$2.152—$2.152—$2.152per 1M tokens$2.11$2.11$10.55n/an/an/an/an/an/a · 1d n/a
57NebiusTrustedRouterZDRcache ≈ input$2.152—$2.152—$2.152per 1M tokens$2.11$2.11$10.55n/an/an/an/an/an/a · 1d n/a
List, quoted, and effective prices are all shown in USD per 1M. Original values are stored as USD per token.TPS n/a is shown as n/a. Latency windows: OpenRouter 30m, Vercel 1h, gateway-only sources n/a.
Data quality notes (4)
Method24 hours · 5m

How the numbers are computed

Every displayed figure comes from a stored observation and a deterministic rule. The index keeps the quote, discount, peak window, effective price, retention class, and raw payload so that any value can be recalculated rather than merely trusted.

Collection

A systemd timer triggers an independent collector every five minutes. It fetches four models from five gateways: OpenRouter endpoints, Vercel AI Gateway endpoints, the Requesty public model catalog, Opper public model routes, and TrustedRouter endpoint lists. Payloads are validated, normalized, archived as gzip, written to SQLite, and projected into per-model read caches.

A single gateway or model failure does not stop the others. Failed combinations keep their last successful offers and are labeled partial or error. If every combination fails, the previous cache is left untouched and health turns stale; no synthetic point is inserted.

Offer identity

An offer is a model × gateway × provider × tier × region × quantization × route configuration. Tier, region, quantization, and gateway route are never merged into one price. OpenRouter tags such as /us and /fp8 become region and quantization attributes; /flex, /priority, and /spot become tiers. Vercel inference regions become separate offers. Opper service routes and TrustedRouter endpoint IDs become the route dimension when they distinguish otherwise identical offers.

Price formula

Blended price = input × input_weight + cache_read × cache_weight + output × output_weight. Weights are normalized when their total is not 100%. The default mix is 2.5% / 97% / 0.5%.

List = quoted / (1 − discount). Effective price applies the UTC peak rule or time override at the observation time. Prices are stored in USD per token, so currency and token-unit changes are reversible display transforms. The gateway-fee toggle applies a second, clearly labeled PAYG multiplier.

  • Discount changes are real signals and are not smoothed.
  • Peak windows wrap across midnight and honor ISO weekdays.
  • Historical points use the current FX rate; historical FX is not replayed.
  • The chart draws one real offer history per provider, not a per-bucket minimum across gateways.

Retention

Provider retention is classified as strict, safety, no-training, retained, or unknown. Strict means no routine content storage, logging, moderation retention, or training as reported by the source. Safety means the source allows retention for abuse, safety, or legal review. No-training means content may be retained but is not used for training. Retained means content may be stored and used for training. Unknown means the source did not publish enough detail.

Gateway retention is a separate static policy. OpenRouter does not store prompts by default and enforces provider ZDR per request. Vercel deletes prompts and outputs after the request. Requesty logs content by default unless disabled. Opper and TrustedRouter report no default prompt storage. The ZDR filter applies to provider retention; the tooltip explains the gateway layer.

TPS, latency, and uptime

TPS is output tokens/sec p50. Vercel reports p50 throughput in its endpoint API. OpenRouter endpoint TPS is often empty, so the collector parses the public model page per model every 30 minutes. Requesty, Opper, and TrustedRouter do not publish TPS for these routes, so they show n/a. Missing TPS is never estimated.

Latency is time-to-first-token; uptime is the source’s trailing window. These fields are evidence only and are kept in the price table for audit.

Storage, coverage, and preferences

SQLite runs in WAL mode with the collector and web process separate. Raw 5-minute snapshots are kept for 30 days, 15-minute aggregates for 180 days, 30-minute aggregates for 2 years, and gzipped source payloads for 90 days. Each bucket uses its last observation. A gap is carried forward only when the last real observation is at most 10 minutes older than the bucket.

The three ranges use fixed granularity: 24 hours at 5 minutes, 3 days at 15 minutes, and 7 days at 30 minutes. Coverage is the share of buckets with at least one observation. Display preferences, filters, and provider selections are stored in the browser; only the last language/model landing target uses a small cookie so the root redirect works before React loads.

Public source audit

The collector accesses only public endpoints and public model pages. No bearer tokens are used. Source sites’ terms and robots policies are respected; requests run at a five-minute cadence.

FAQmethod, gaps, TPS, and pricing

Questions the data should answer

01What price is actually being compared?

The chart and table use the effective blended price: input × input weight + cache-read × cache weight + output × output weight. Each offer stores input, cache-read, and output prices in USD per token. The default usage mix is 2.5% input, 97% cache read, and 0.5% output. List price is used only for the struck-through reference and discount badge.

02What is the difference between a gateway and a provider?

A gateway is the routing and billing layer (OpenRouter, Vercel AI Gateway, Requesty, Opper, TrustedRouter). A provider is the organization that actually runs inference (DeepInfra, Morph, Wafer, Moonshot AI, Z.AI, and others). One provider can appear through several gateways at different prices. The chart draws the provider’s current best offer; the full snapshot lists every gateway offer so the choice is auditable.

03How is Zero Data Retention determined?

Retention is two layers: gateway retention and provider retention. Provider retention is normalized to strict, safety, no-training, retained, or unknown using the source’s published flags: OpenRouter’s ZDR endpoint list, Vercel’s has_zdr and has_no_training, Requesty’s data_retention fields, Opper’s compliance metadata, and TrustedRouter’s endpoint flags. A strict provider does not guarantee that the gateway itself does not log content; check both layers. The ZDR filter defaults to strict only and can include safety retention.

04Are gateway fees included in the ranking?

By default the index compares published token rates, matching the existing OpenRouter and Vercel data. The “Include gateway fees” toggle applies each gateway’s public PAYG estimate: OpenRouter 5.5%, Requesty 5%, Opper 3%, TrustedRouter 5.5%, and Vercel 0%. BYOK, enterprise contracts, and payment methods can change the real all-in cost, so the toggle is an estimate rather than a quote.

05Why does history start at a recent timestamp?

No gateway exposes historical provider quotes for these models. History begins at the first successful collection and is never back-filled or synthesized. The collector runs every five minutes; new models and new gateways start with an empty series and grow over time. Gaps longer than the ten-minute forward-fill allowance are rendered as hatched bands rather than connected lines.

06Why can the same provider have different prices on different gateways?

Each offer is a specific route: model × gateway × provider × tier × region × quantization × route. A provider may have a discounted endpoint on one gateway, a regional endpoint on another, or a different quantization. The index never merges those into one price. The chart picks the provider’s current best offer for the selected tier, retention filter, gateway filter, and usage mix; the full snapshot keeps every route.

07Does changing currency or token unit change the ranking?

No. Prices are stored in USD per token. Currency and token-unit conversion are display-only multipliers using the latest daily ECB reference rates. Changing the usage mix, tier, ZDR mode, gateway filter, or gateway-fee toggle does recompute the ranking.

08Is this an official DeepSeek, Moonshot, Z.ai, OpenRouter, or Vercel product?

No. This is an independent, self-hosted, read-only index built from public endpoints and public model pages. It does not use bearer tokens or private APIs, is not affiliated with any model provider or gateway, and is not a purchasing commitment or a commercial data license.