{
	"about": [
		"The reviewed input to scripts/generate-model-prices.mjs. Edit THIS file, never",
		"catalogue.generated.ts — the generator overwrites that file and a CI gate",
		"re-derives it from this one and fails on any difference.",
		"",
		"A rate here is a commercial fact about a vendor's published list price. It is",
		"entered by a human who read that vendor's own pricing page, and `verified` is",
		"the date they read it. It is NOT fetched at build time: an auto-refreshing",
		"generator would change every consumer's reported cost AND every consumer's",
		"costLimitUsd behaviour on an unrelated commit, with no review and no diff.",
		"",
		"Absent is a real answer. A model with no row here reports UNKNOWN cost, not",
		"zero, and a turn that sets turnConfig.costLimitUsd against it is refused rather",
		"than run without a budget. So the correct move when a rate cannot be",
		"established is to leave the row out, never to approximate it.",
		"",
		"unmetered: true means this driver bills nothing by construction — local",
		"inference. Its turns are priced at zero, which is KNOWN-free and therefore",
		"distinct from unknown. Such a vendor needs no model rows.",
		"",
		"Matching is EXACT on a normalised id (lowercased, one trailing -YYYYMMDD or",
		"@YYYYMMDD snapshot suffix removed). Deliberately not prefix or substring:",
		"within one vendor here, ids one character apart differ in price by 4x, so a",
		"near-miss would misprice silently. A miss reports unknown, which is safe."
	],
	"cacheTokensAre": [
		"Two of the four rates only mean anything alongside a fact about the DRIVER,",
		"not about the model: whether the prompt-token count it reports already",
		"contains the tokens it read from cache.",
		"",
		"The drivers in this repository disagree, and it is measured, not assumed:",
		"  packages/providers/anthropic/src/client.ts parseUsage — promptTokens is",
		"    input_tokens, which EXCLUDES cache reads and cache writes.",
		"  packages/providers/bedrock/src/client.ts parseBedrockUsage — same.",
		"  packages/providers/openai/src/client.ts parseUsage — promptTokens is",
		"    prompt_tokens, which INCLUDES cached_tokens.",
		"",
		"So the same TokenUsage shape carries two incompatible meanings, and a rate",
		"applied without knowing which one is off by the whole cache-read volume. The",
		"fact is a property of the driver, and the driver's identity is in hand at the",
		"moment cost is accumulated — so it is declared here, per vendor, once.",
		"",
		"promptIncludesCacheReads   true  => billable input is promptTokens minus",
		"                                    cachedTokens.",
		"                           false => billable input is promptTokens, and",
		"                                    cachedTokens are charged on top.",
		"",
		"reportsCacheWrites declares whether the driver ever produces a non-zero",
		"cacheWriteTokens. When false, model rows carry no write rate — a rate for a",
		"quantity nothing produces is a declaration nothing drives. If such a driver",
		"later does report writes, those tokens are counted as unpriced rather than",
		"quietly charged at some other rate, so the contradiction surfaces."
	],
	"vendors": [
		{
			"providerId": "anthropic",
			"unmetered": false,
			"verified": "2026-08-10",
			"promptIncludesCacheReads": false,
			"reportsCacheWrites": true,
			"note": "The last two rows exist because the driver's own offline catalogue offers those models and this table did not price them — a lookup-key gap that reads as `cost unknown` and is invisible to the generator's own check, which only proves the module matches this file. `providers/anthropic` asserts the two lists agree. Vendor list prices. Cache read is the vendor's published 0.1x of the input rate; cache write is its published 1.25x for the five-minute entry, which is the default TTL and the one the driver's `cacheControl: { type: 'auto' }` asks for. A time-limited introductory rate applied to one model at the time of entry; this table has no time dimension and records the list rate, so a turn inside a promotional window is over-reported rather than mis-modelled.",
			"models": [
				{
					"id": "claude-fable-5",
					"inputPer1M": 10,
					"outputPer1M": 50,
					"cacheReadPer1M": 1,
					"cacheWritePer1M": 12.5
				},
				{
					"id": "claude-mythos-5",
					"inputPer1M": 10,
					"outputPer1M": 50,
					"cacheReadPer1M": 1,
					"cacheWritePer1M": 12.5
				},
				{
					"id": "claude-opus-5",
					"inputPer1M": 5,
					"outputPer1M": 25,
					"cacheReadPer1M": 0.5,
					"cacheWritePer1M": 6.25
				},
				{
					"id": "claude-opus-4-8",
					"inputPer1M": 5,
					"outputPer1M": 25,
					"cacheReadPer1M": 0.5,
					"cacheWritePer1M": 6.25
				},
				{
					"id": "claude-opus-4-7",
					"inputPer1M": 5,
					"outputPer1M": 25,
					"cacheReadPer1M": 0.5,
					"cacheWritePer1M": 6.25
				},
				{
					"id": "claude-opus-4-6",
					"inputPer1M": 5,
					"outputPer1M": 25,
					"cacheReadPer1M": 0.5,
					"cacheWritePer1M": 6.25
				},
				{
					"id": "claude-sonnet-5",
					"inputPer1M": 3,
					"outputPer1M": 15,
					"cacheReadPer1M": 0.3,
					"cacheWritePer1M": 3.75
				},
				{
					"id": "claude-sonnet-4-6",
					"inputPer1M": 3,
					"outputPer1M": 15,
					"cacheReadPer1M": 0.3,
					"cacheWritePer1M": 3.75
				},
				{
					"id": "claude-haiku-4-5",
					"inputPer1M": 1,
					"outputPer1M": 5,
					"cacheReadPer1M": 0.1,
					"cacheWritePer1M": 1.25
				},
				{
					"id": "claude-sonnet-4-5",
					"inputPer1M": 3,
					"outputPer1M": 15,
					"cacheReadPer1M": 0.3,
					"cacheWritePer1M": 3.75
				},
				{
					"id": "claude-opus-4-1",
					"inputPer1M": 15,
					"outputPer1M": 75,
					"cacheReadPer1M": 1.5,
					"cacheWritePer1M": 18.75
				}
			]
		},
		{
			"providerId": "openai",
			"unmetered": false,
			"verified": "2026-08-10",
			"promptIncludesCacheReads": true,
			"reportsCacheWrites": false,
			"note": "Vendor list prices for text models, cached-input rate included. The driver hard-codes cacheWriteTokens to zero, so no write rate is carried. Embedding models are omitted (no output rate, and a chat turn never reaches one), as are models whose cached-input rate was not published alongside the others — an unestablished rate is left out rather than approximated, which is the rule this whole file follows.",
			"models": [
				{
					"id": "gpt-5.6-sol",
					"inputPer1M": 5,
					"outputPer1M": 30,
					"cacheReadPer1M": 0.5
				},
				{
					"id": "gpt-5.6-terra",
					"inputPer1M": 2,
					"outputPer1M": 12,
					"cacheReadPer1M": 0.2
				},
				{
					"id": "gpt-5.6-luna",
					"inputPer1M": 0.2,
					"outputPer1M": 1.2,
					"cacheReadPer1M": 0.02
				},
				{
					"id": "gpt-5.5",
					"inputPer1M": 5,
					"outputPer1M": 30,
					"cacheReadPer1M": 0.5
				},
				{
					"id": "gpt-5.4",
					"inputPer1M": 2.5,
					"outputPer1M": 15,
					"cacheReadPer1M": 0.25
				},
				{
					"id": "gpt-5.4-mini",
					"inputPer1M": 0.75,
					"outputPer1M": 4.5,
					"cacheReadPer1M": 0.075
				},
				{
					"id": "gpt-5.4-nano",
					"inputPer1M": 0.2,
					"outputPer1M": 1.25,
					"cacheReadPer1M": 0.02
				},
				{
					"id": "gpt-5.2",
					"inputPer1M": 1.75,
					"outputPer1M": 14,
					"cacheReadPer1M": 0.175
				},
				{
					"id": "gpt-5.1",
					"inputPer1M": 1.25,
					"outputPer1M": 10,
					"cacheReadPer1M": 0.125
				},
				{
					"id": "gpt-5",
					"inputPer1M": 1.25,
					"outputPer1M": 10,
					"cacheReadPer1M": 0.125
				},
				{
					"id": "gpt-5-mini",
					"inputPer1M": 0.25,
					"outputPer1M": 2,
					"cacheReadPer1M": 0.025
				},
				{
					"id": "gpt-5-nano",
					"inputPer1M": 0.05,
					"outputPer1M": 0.4,
					"cacheReadPer1M": 0.005
				},
				{
					"id": "gpt-4.1",
					"inputPer1M": 2,
					"outputPer1M": 8,
					"cacheReadPer1M": 0.5
				},
				{
					"id": "gpt-4.1-mini",
					"inputPer1M": 0.4,
					"outputPer1M": 1.6,
					"cacheReadPer1M": 0.1
				},
				{
					"id": "gpt-4.1-nano",
					"inputPer1M": 0.1,
					"outputPer1M": 0.4,
					"cacheReadPer1M": 0.025
				},
				{
					"id": "gpt-4o",
					"inputPer1M": 2.5,
					"outputPer1M": 10,
					"cacheReadPer1M": 1.25
				},
				{
					"id": "gpt-4o-mini",
					"inputPer1M": 0.15,
					"outputPer1M": 0.6,
					"cacheReadPer1M": 0.075
				},
				{
					"id": "o1",
					"inputPer1M": 15,
					"outputPer1M": 60,
					"cacheReadPer1M": 7.5
				},
				{
					"id": "o3",
					"inputPer1M": 2,
					"outputPer1M": 8,
					"cacheReadPer1M": 0.5
				},
				{
					"id": "o3-mini",
					"inputPer1M": 1.1,
					"outputPer1M": 4.4,
					"cacheReadPer1M": 0.55
				},
				{
					"id": "o4-mini",
					"inputPer1M": 1.1,
					"outputPer1M": 4.4,
					"cacheReadPer1M": 0.275
				}
			]
		},
		{
			"providerId": "ollama",
			"unmetered": true,
			"verified": "2026-08-10",
			"note": "Local inference. Nothing is billed for a token, whatever the model, so there is no model list to keep current.",
			"models": []
		},
		{
			"providerId": "lmstudio",
			"unmetered": true,
			"verified": "2026-08-10",
			"note": "Local inference. Same as above.",
			"models": []
		}
	],
	"absent": [
		"Drivers deliberately carrying no rows, so that a turn through them reports",
		"unknown rather than a number nobody checked:",
		"",
		"  bedrock    — partner-operated, on a separate rate card from the vendor whose",
		"               models it serves. Entering the vendor's own rates against it",
		"               would be the exact class of quiet wrongness this file exists to",
		"               stop.",
		"  openrouter — routes to arbitrary upstreams and publishes per-model rates from",
		"               its own catalogue endpoint at run time. That is a live source",
		"               and a static table cannot be one.",
		"  http       — a driver for whatever endpoint the host points it at. There is",
		"               no vendor to have a price.",
		"",
		"  deepseek   — publishes two list prices per model and picks between them by the",
		"               CLOCK: peak (01:00-04:00 and 06:00-10:00 UTC) costs exactly twice",
		"               off-peak, on both input and output. A static table has no hour, so",
		"               whichever of the two numbers were entered here would be wrong for",
		"               half of every day — and wrong in the direction that reads as a",
		"               correct, verified figure. Read 2026-08-17.",
		"Adding one is a normal change: add the vendor here with the rows and the date",
		"they were read, run the generator, commit both files."
	]
}