TokenMeter 0.6.4

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

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.

// 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

// 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);

// Via CostCalculator (DI-friendly)
ICostCalculator calc = CostCalculator.Default();
var price = calc.CalculateCost("gpt-4o", inputTokens: 1_000, outputTokens: 500);

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 (the built-in catalog currently contains Chat models only)
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);

Model Metadata

ModelInfo provides the following metadata:

Identity

Limits

Pricing (USD / 1M tokens)

Note — one rate per model: these fields hold a provider's standard rate for a model. Where a provider charges more above a prompt-length threshold (several now publish a second, higher tier for long-context requests), the catalog carries the base tier only, so cost for a request past that threshold is understated. 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.

Input Modalities

API Capabilities

Reasoning & Thinking

Tool Calling Wire Format

Supported Providers

Data Freshness

Console.WriteLine(ModelCatalog.LastUpdated);        // 2026-08-01
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.

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.

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