TokenMeter 0.13.1

TokenMeter

LLM model metadata catalog and cost calculator for .NET.

Provides context windows, pricing, capability flags (vision, audio, reasoning, tool calling, prompt caching), and thinking/reasoning format metadata for 12+ providers including OpenAI, Anthropic, Google, xAI, Mistral, DeepSeek, and more.

Packages

Installation

dotnet add package TokenMeter

Features

Everything is available as soon as the package is referenced — the catalog is embedded and loads on first use; there is no DI extension or setup call. Register ICostCalculator yourself if you want it injected.

Quick Start

Model Lookup

// Find a model by ID or alias
var model = ModelCatalog.FindModel("claude-sonnet-4-6");

Console.WriteLine(model?.ContextWindow);          // 1000000
Console.WriteLine(model?.ReasoningMode);           // Optional
Console.WriteLine(model?.ThinkingFormat);          // Block
Console.WriteLine(model?.PromptCachingMode);       // Explicit
Console.WriteLine(model?.ToolCallingFormat);       // Anthropic
Console.WriteLine(model?.SupportsMcpToolUse);      // True

Note — fuzzy matching: FindModel resolves aliases in 4 passes (exact → alias exact → prefix → contains) to absorb cloud-specific ID variants (Bedrock/Vertex prefixes, date suffixes). Local/self-hosted deployment names that embed a public model name (e.g. deepseek-r1-distill-qwen-7b) can therefore match a catalog entry whose context window and pricing do not describe your deployment. For self-hosted models, take the effective context length from your deployment configuration (e.g. llama.cpp n_ctx), not from the catalog.

Fuzzy passes never cross a version token: an alias ending in a version number (claude-sonnet-5) does not match an id that continues it (claude-sonnet-5-5) — an unknown newer model returns null rather than the older model's row. Snapshot and deployment suffixes (-20250929, -latest, -v1) still match. Nor do they cross into a variant: an id whose remainder names audio, realtime, tts, transcribe, search, image, embedding, live, translate or moderation (gpt-4o-mini-tts) is a different product and resolves only by a row of its own.

The catalog loads with source-generated JSON metadata, so it works under Native AOT and trimming.

// When correctness matters more than recall, bound the fuzziness:
ModelCatalog.FindModel("deepseek-r1-distill-qwen-7b", AliasMatchType.Exact);   // null — no false positive
ModelCatalog.FindModel("us.anthropic.claude-sonnet-4.6-v1");                   // matched via contains alias

// Or inspect which pass produced the match and apply confidence-based fallback:
var match = ModelCatalog.FindModelMatch("deepseek-r1-distill-qwen-7b");
Console.WriteLine(match?.MatchKind);   // Prefix — a fuzzy inference, not an exact hit
Console.WriteLine(match?.Model.ModelId);

Cost Calculation

One request, every part priced the way the vendor prices it — the rates of the call's date, the long-context or time-of-day tier the request falls in (decided by its whole prompt, cache reads and writes included, and applied to the whole request), cache reads, and each cache write at the price for its lifetime:

var usage = new TokenCounts
{
    InputTokens = 20_000,                                   // neither read from nor written to the cache
    CacheReadTokens = 150_000,
    CacheWrites = [new CacheWriteTokenCount(TimeSpan.FromHours(1), 30_000)],   // null lifetime = the vendor default
    OutputTokens = 2_000,                                   // reasoning included
};
var requestCost = model?.CalculateCost(usage, CostContext.At(callTimeUtc));
var viaCalculator = CostCalculator.Default().CalculateCost("claude-haiku-5-5", usage, CostContext.At(callTimeUtc));

// Batch / Flex / Fast and a data-residency region multiply every token price; server-side tool calls add their fee
var batchInUs = model?.CalculateCost(
    usage with { ToolCalls = new Dictionary<string, int> { ["web_search"] = 2 } },
    CostContext.At(callTimeUtc) with { ServiceTier = ServiceTier.Batch, Region = "us" });

The order is: rates of the date (AsOf) → long-context / time-of-day tier → service-tier multiplier → regional multiplier → tool fees. A service tier the model is not offered at, or a tool with no fee in the catalog, makes the result null (unknown) rather than the standard price. Tool names are the vendors' own (web_search, file_search, google_search), and ToolCalls holds the number of calls the response reports (searches, queries); the fee's ToolCallPrice.Unit decides whether each is billed or only the request. Tokens a tool adds to the context are already in the token counts, and a vendor's monthly free allowance is not applied.

The overloads below are the same calculation for one part at a time:

// Basic cost (input + output tokens)
var cost = model?.CalculateCost(inputTokens: 500_000, outputTokens: 200_000);

// Cost including prompt cache tokens
var costWithCache = model?.CalculateCost(
    inputTokens: 100_000,
    outputTokens: 50_000,
    cacheReadTokens: 400_000,
    cacheWriteTokens: 50_000);

// Cache writes priced by lifetime (Anthropic: 5-minute vs 1-hour writes — CacheWritePrices)
var costOneHourCache = model?.CalculateCost(100_000, 50_000, 400_000, 50_000, cacheWriteTtl: TimeSpan.FromHours(1));

// The rates in effect on a date — applies vendor-announced price changes (ScheduledPrices)
var costNextYear = model?.AsOf(new DateOnly(2027, 1, 1)).CalculateCost(500_000, 200_000);

// Via CostCalculator (DI-friendly) — the basic and cache overloads, the tiered one (see Tiered Pricing below) and TokenCounts
ICostCalculator calc = CostCalculator.Default();
var price = calc.CalculateCost("gpt-4o", inputTokens: 1_000, outputTokens: 500);

Tiered Pricing

Some providers charge more than the representative rate under a specific condition — a long-context surcharge past a prompt-length threshold (xAI), or a peak-hour surcharge during specific UTC windows (DeepSeek). ModelInfo.PricingTiers carries those bands; passing a PricingTierContext picks the right one automatically. Omitting the context, or a model with no tiers, reproduces the representative-rate calculation exactly — existing callers are unaffected. ContextLengthTokens is the request's prompt length — every input token, cache reads and writes included — which is how the vendors decide the tier; CalculateCost(TokenCounts, …) computes it for you. A tier can carry its own cache read and write prices (Claude Haiku 5.5 over 100K, Gemini 2.5 Pro / 3.1 Pro Preview over 200K).

var model = ModelCatalog.FindModel("grok-4.6");

// Below the tier's threshold — representative rate ($2.00 / $6.00 per 1M)
var normal = model?.CalculateCost(50_000, 10_000, new PricingTierContext { ContextLengthTokens = 50_000 });

// At/above 200K context — the long-context tier applies ($4.00 / $12.00 per 1M)
var longContext = model?.CalculateCost(50_000, 10_000, new PricingTierContext { ContextLengthTokens = 250_000 });

// DeepSeek: peak-hour tier, keyed off wall-clock UTC time instead of context length
var deepseek = ModelCatalog.FindModel("deepseek-v4-pro");
var peakCost = deepseek?.CalculateCost(50_000, 10_000,
    new PricingTierContext { CallTimeUtc = TimeOnly.FromDateTime(DateTime.UtcNow) });

Browsing the Catalog

// By provider — typed convenience property
foreach (var m in ModelCatalog.Anthropic.Values)
    Console.WriteLine($"{m.ModelId}: ctx={m.ContextWindow}, ${m.InputPricePerMillion}/M");

// By provider — string-keyed (when the name is only known at runtime)
var openai = ModelCatalog.GetProvider("OpenAI");   // dict, empty if unknown

// By model type (mostly Chat, plus Gemini ImageGeneration/TextToSpeech and OpenAI TextToSpeech/SpeechToText models)
var chatModels = ModelCatalog.GetByType(ModelType.Chat);

// All providers
var providers = ModelCatalog.GetProviderNames();

Custom Models

var calc = CostCalculator.Default();
calc.RegisterModel(new ModelInfo
{
    ModelId = "my-fine-tuned-model",
    Provider = "MyCompany",
    InputPricePerMillion = 2.00m,
    OutputPricePerMillion = 8.00m,
    ContextWindow = 128_000,
    SupportsToolCalling = true,
    ToolCallingFormat = ToolCallingFormat.OpenAI
});

var cost = calc.CalculateCost("my-fine-tuned-model", 10_000, 5_000);

// A self-hosted name that embeds a public one ("qwen3-8b-local") would match the public model fuzzily.
// Bound the matching so an unregistered name is unknown instead of priced as something else:
var strict = CostCalculator.Default(AliasMatchType.Exact);

Model Metadata

ModelInfo provides the following metadata:

Identity

Limits

Pricing (USD / 1M tokens)

Note — these fields hold the representative rate: the price outside of any pricing tier a model carries. Cache-write cost falls back to the input rate when a provider does not price it separately, which matches how automatic prompt caching is normally billed.

Reasoning tokens are billed at the output rate by OpenAI, Anthropic and Google — pass them as output tokens. OpenAI (output_tokens includes reasoning_tokens) and Anthropic (thinking is part of output_tokens) already count them there; Gemini reports them apart (thoughtsTokenCount), so its output is candidatesTokenCount + thoughtsTokenCount. Prices are in the vendor's currency (USD).

Lifecycle

GetLifecycleStatus(asOf) returns Active / Deprecated / Retired for a date, computed from the dates so an older package still answers correctly. Filled for Anthropic, OpenAI and Google from their deprecation pages; partner clouds (Bedrock, Vertex AI, Azure) keep their own schedules and are not covered. Google publishes a shutdown date for a generally available model from its release (the earliest date it may be shut down), so a Google model can carry a RetirementDate without a DeprecationDate; a model whose requests Google routes to its successor carries only ReplacementModelId.

Input Modalities

API Capabilities

Reasoning & Thinking

Tool Calling Wire Format

Supported Providers

Data Freshness

Console.WriteLine(ModelCatalog.LastUpdated);        // e.g. 2026-10-10
Console.WriteLine(ModelCatalog.DataAgeDays);        // days since last update
Console.WriteLine(ModelCatalog.IsDataStale());      // true if > 90 days old

LastUpdated is derived from the lastUpdated field each bundled provider file declares, and reports the most recent of them. Providers are refreshed independently, so an individual provider's data can be considerably older than this value — read that provider's own date with ModelCatalog.GetLastUpdated("Google") (null for an unknown provider).

Note — what the signal does and does not tell you: it reports when this catalog was last refreshed, not whether a provider has changed its prices since. A vendor can cut a rate the day after a refresh, and IsDataStale() will still answer false while the bundled figure is wrong. Treat catalog pricing as a good default for estimation and budgeting, and read authoritative figures from your provider's billing data when they have to be exact.

Price Source

var model = ModelCatalog.FindModel("llama-4-maverick")!;
Console.WriteLine(model.PriceSource); // ThirdParty

ModelInfo.PriceSource reports how a model's price fields were obtained:

• Official (default) — verified directly against the vendor's own published rate card.
• ThirdParty — the vendor's own rate card could not be read directly (a client-side-rendered pricing page, a docs-only landing page, or no direct per-token price published at all), so the figure was cross-referenced from a third-party aggregator or another vendor's official rate card instead. Currently applies to Meta Llama (Maverick, Scout) and Cohere Command A — see docs/pricing-update-guide.md Version History for the source used per provider. Treat these as estimates for cost-sensitive accounting.

Migration from 0.3.x

The following APIs were removed in 0.4.0:

• TokenMeter.Abstractions package (removed entirely)
• ITokenCounter, TokenCounter — use your own tokenizer library
• IUsageTracker, UsageTracker, UsageRecord, UsageStatistics — implement in your application
• ModelPricing → replaced by ModelInfo
• ModelPricingData → replaced by ModelCatalog
• ICostCalculator.GetPricing() → GetModel()
• ICostCalculator.RegisterPricing() → RegisterModel()

Requirements

• .NET 10.0 or later

License

MIT