Get started

Understand the AI market.

Models, benchmarks, prices, adoption and market changes, each figure with its source and the date it was read. 359 models, 713 skus on the register, and 203 dated market changes, free to read with no key.

All of it public but the Decision Console: no account, no key and no quota.

Market observations

What changed recently across the record?

Every published price cut, launch, deprecation and policy change since 2023, dated from its primary source and kept in one log.

Standing events
203
Providers watched
34
Latest date
19 Nov 2026
Standing market log · newest 6GET /api/v1/markets
Newest recorded move19 Nov 2026 · Novita AI

Novita AI Kimi K3 cache write price, 1 hour retention falls to $0.30

Cache write price, 1 hour retention: $6.00 -> $0.30 (USD per 1M tokens)

  1. 19 Nov 2026Novita AINovita AI Kimi K3 cache write price, 1 hour retention falls to $0.30price
  2. 19 Nov 2026Novita AINovita AI Kimi K3 cache write price, 5 minute retention rises to $6.00price
  3. 19 Nov 2026Novita AINovita AI Kimi K3 cached input price rises to $3.00price
  4. 19 Nov 2026Novita AINovita AI Qwen3.8-Max cache write price, 5 minute retention falls to $0.25price
  5. 19 Nov 2026Novita AINovita AI Qwen3.8-Max cached input price rises to $2.50price
  6. 19 Nov 2026Novita AINovita AI Qwen3.8-Flash cache write price, 5 minute retention falls to $0.016price
01 / 05Evidence ladderFour claim tiers · never mixed

Every figure says whose claim it is.

From a provider's word to your own cases, each page is sorted by the kind of evidence it carries so marketing claims and workload evaluations never blur together.

01Declared

Provider declarations

What a provider publishes about its own model: the price, the context window, the modalities. Vitna reads it from the provider's page and keeps the page.

Source fidelity

Provider list prices and resold open-weight rates stay separated by seller type and keep their original source document digest.

Source document archivesha256 verified
anthropic.com/pricingfirst_party
openai.com/api/pricingfirst_party
ai.google.dev/gemini-api/docs/pricingfirst_party
Providers indexed34
Registered skus713

Prices, context windows and modalities are kept verbatim from the seller's own domain.

02Measured

Third-party benchmarks

A score somebody else measured, on their own tasks and their own scale. Vitna names the board and its date, and never adds two boards together.

No composite blends

Vitna prints each benchmark on the author's own scale and date, and never invents a single combined score across unrelated boards.

Separate third-party scalesnever blended
Artificial Analysis58 models

Intelligence, Coding and Agentic indexes kept on their published 0 to 100 scale.

Design Arena132 models

Pairwise Elo and win rates printed on their own arena scale, never merged into an index.

Each board carries its own author, task set and retrieval date.

03Observed

Market observations

What the market did: a price that moved, a model that launched, a share of usage. Each is dated and carries the source it was read from.

Dated market history

Every logged price cut, launch, deprecation and usage share keeps the date it was read and the primary URL or feed behind it.

Observed market feed203 events
  1. Novita AI · 19 Nov 2026Novita AI Kimi K3 cache write price, 1 hour retention falls to $0.30
  2. Novita AI · 19 Nov 2026Novita AI Kimi K3 cache write price, 5 minute retention rises to $6.00
  3. Novita AI · 19 Nov 2026Novita AI Kimi K3 cached input price rises to $3.00

Price shifts, launches, deprecations and OpenRouter weekly usage shares, dated back to 2023.

04Decisions

Workload-specific evaluations

Your own cases, scored by your own harness, against a threshold you wrote down. The only kind of evidence that can decide a production change.

Human-ruled threshold

Public market signals trigger a review brief; a production switch only happens when your own representative cases pass and a person signs off.

Production ruling gateshuman approved
  1. 01 · Quality flooryour evals
  2. 02 · Context and modalityhard check
  3. 03 · Unit economicslive tape
  4. 04 · Production changeperson rules

Public boards shortlist candidates; only your own representative cases can approve a switch.

02 / 05Read at the source34 providers · archived by sha256

34 providers. One record.

Each price comes from a document on the provider's own domain, archived byte for byte under its sha256.

Inspect every provider
skus on the register
713
moves logged
203
listings
451
keys needed
0
  • Novita AI 99
  • OpenAI 92
  • Google 88
  • AWS 85
  • DeepInfra 78
  • Together AI 68
  • Parasail 39
  • Anthropic 23
  • Baseten 22
  • Mistral AI 14
  • Azure 12
  • SpaceXAI 8
  • Deepgram 8
  • AssemblyAI 8
03 / 05The marketFrontier · spreads · non-token units
The market

Where the same work costs less, read from the record.

Who scores as well as the model you run for less, what one model costs from each company that serves it, and what a unit of everything beyond tokens costs. Every price is one a seller published, and every figure says where it was read.

Intelligence per dollar, today

57 models on one index against their blended list price. The line is the frontier: the best score each budget buys.

  • OpenAI8
  • Anthropic7
  • Google7
  • Meta4
  • DeepSeek3
  • Mistral3
  • NVIDIA3
  • Thinking Machines3

2 models score at least as high as DeepSeek V4.1 Flash 1017 on this index, for less.

The most-used model on OpenRouter this week, at 20.8% of tokens.

One index is not your workload. See how the engine decides
Artificial Analysis Intelligence Index, as OpenRouter republishes it · list prices blended 3 input to 1 output · read 17 Nov
Open weights · 26 models sold by three companies or more · read 2 Nov

One model, many prices

The same weights, priced by each company that serves them: input per million tokens as every seller publishes it, on one scale. Green is the cheapest seller, amber the dearest.

Every resold model
  1. DeepSeek V4 Flash 07313 sellersDeepInfra, $0.06Baseten, $0.13Novita AI, $0.447.3x$0.06 to $0.44
  2. gpt-oss-120b6 sellersNovita AI, $0.05Google, $0.09Baseten, $0.10Parasail, $0.10AWS, $0.15Together AI, $0.153x$0.05 to $0.15
  3. Gemma 3 27B3 sellersParasail, $0.08Novita AI, $0.119AWS, $0.232.9x$0.08 to $0.23
  4. DeepSeek R1 05283 sellersDeepInfra, $0.50Novita AI, $0.70Google, $1.352.7x$0.50 to $1.35
  5. Qwen3.6-35B-A3B3 sellersDeepInfra, $0.10Parasail, $0.15Novita AI, $0.2482.5x$0.10 to $0.248
  6. DeepSeek V3.13 sellersDeepInfra, $0.25AWS, $0.58Google, $0.602.4x$0.25 to $0.60
Beyond tokens · 9 more boards

Not only tokens

Images, speech, video, search, pages and GPU hours, each read from the seller's own page in the unit it bills. The green dot is the cheapest on the board.

Every board
04 / 05Check the evidencesha256 · Merkle leaf · Ed25519

A record a stranger can check.

One fetch, one hash, one comparison, and no account.

  1. 01

    Fetch the archived bytes

    Every figure names the sha256 of the document it was read from. Fetch that document from the evidence endpoint, or from the provider, and you hold the same bytes Vitna read.

  2. 02

    Hash them

    One sha256 over the bytes. It has to equal the hash the tape asserts, or the tape is wrong and says so in public.

  3. 03

    Compare against the signed root

    Each day closes into a Merkle root signed with Ed25519. The proof endpoint hands you the leaf, the path and the signature, and the arithmetic either agrees or it does not.

vitna · verify
$ curl -s "<the archived url>" | sha256sum
0de899a93b1d…  -

$ curl https://vitna.ai/api/v1/observations/1/proof
{
  "leaf": {
    "sourceId": "anthropic-pricing-md",
    "skuId": "anthropic/claude-haiku-4-5",
    "dimensionId": "input_per_mtok",
    "amountUsd": "1.000000",
    "observedAt": "2026-08-21T23:22:08.743Z"
  },
  "proof": [ { "side": "left", "hash": "…" } ],
  "attestation": { "root": "…", "signature": "…", "publicKey": "…" }
}