language-models 2.1.3 → 2.3.0

package/src/pricing/table.ts

@@ -0,0 +1,235 @@
+/**
+ * language-models / pricing — canonical model pricing table.
+ *
+ * All rates are USD per 1,000,000 tokens (per1M form, NOT per1k). Sourced
+ * from public list prices on 2026-05-07. Where prior repo tables disagree,
+ * the most-recently-updated source wins:
+ *
+ * - **Vertex Gemini 3.x**: startup-builder/packages/llm-vertex-batch
+ *   (sb-srnl 2026-05-07 verified) is canonical for the flex/batch tier.
+ *   startup-builder/packages/llm-vertex (same date) is canonical for
+ *   standard interactive pricing.
+ * - **Vertex Gemini 2.5**: startup-builder/llm-vertex-batch is canonical
+ *   (icps's `gemini-2.5-flash` rates were ~4× off — likely transposed
+ *   from a different SKU).
+ * - **Bedrock Anthropic**: startup-builder/packages/llm-bedrock is the
+ *   most-recently-curated table (matches AWS Bedrock public pricing on
+ *   2026-05-07).
+ * - **Embeddings**: icps/packages/llm/pricing.ts is the source for
+ *   `gemini-embedding-2`. Bedrock has no first-party Gemini embedding
+ *   SKU; Anthropic has no embedding SKU. Embedding lives under
+ *   `aistudio/` since the cheapest path is the AI Studio API key (icps's
+ *   2026-05-07 fallback chain — the path startup-builder also uses).
+ *
+ * 200K-context tier breakpoint:
+ *
+ * Gemini 3.1 Pro Preview list pricing has a 2× rate above 200K input
+ * tokens. We model this with `contextTierBreakpoint: 200_000` +
+ * `contextTierAbove`. Anthropic Claude has flat pricing (no breakpoint).
+ * Gemini 2.5 + 3.x flash-lite + embedding SKUs are flat too.
+ *
+ * Adding a new SKU:
+ *
+ * 1. Append a new row keyed on `<provider>/<short-slug>` + `tier`. Both
+ *    fields together form the lookup key; duplicates are caught by the
+ *    table-integrity test.
+ * 2. If the SKU has a context-tier breakpoint, add
+ *    `contextTierBreakpoint` + `contextTierAbove`. Otherwise omit them.
+ * 3. If both standard + batch tiers exist, add BOTH rows; downstream
+ *    consumers select via `priceFor({ tier: ... })`. Vertex's batch tier
+ *    is ~50% of standard; that's not enforced — declare the actual
+ *    published rates.
+ * 4. Bump the package version (semver minor for additions).
+ */
+
+import type { ModelPricing } from './types.js'
+
+export const PRICING_TABLE: readonly ModelPricing[] = [
+  // ---------------------------------------------------------------------
+  // Vertex Gemini 3.x (200K-context breakpoint applies to 3.1 Pro)
+  // ---------------------------------------------------------------------
+
+  // gemini-3.1-pro-preview standard interactive
+  // ≤200K: $2/M in, $12/M out  ;  >200K: $4/M in, $18/M out
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3.1-pro',
+    tier: 'standard',
+    inputPer1M: 2.0,
+    outputPer1M: 12.0,
+    contextTierBreakpoint: 200_000,
+    contextTierAbove: { inputPer1M: 4.0, outputPer1M: 18.0 },
+  },
+
+  // gemini-3.1-pro-preview flex/batch (sb-srnl 2026-05-07 verified)
+  // ≤200K: $1/M in, $6/M out  ;  >200K: $2/M in, $9/M out
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3.1-pro',
+    tier: 'batch',
+    inputPer1M: 1.0,
+    outputPer1M: 6.0,
+    contextTierBreakpoint: 200_000,
+    contextTierAbove: { inputPer1M: 2.0, outputPer1M: 9.0 },
+  },
+
+  // gemini-3.1-flash-lite-preview standard
+  // Flat: $0.25/M in, $1.50/M out (no breakpoint per public table)
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3.1-flash-lite',
+    tier: 'standard',
+    inputPer1M: 0.25,
+    outputPer1M: 1.5,
+  },
+
+  // gemini-3.1-flash-lite-preview flex/batch
+  // Flat: $0.13/M in, $0.75/M out
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3.1-flash-lite',
+    tier: 'batch',
+    inputPer1M: 0.13,
+    outputPer1M: 0.75,
+  },
+
+  // gemini-3-pro-preview — placeholder pricing using 3.1 sibling rates
+  // (Vertex hasn't published separate 3.0 list prices as of 2026-05-07)
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3-pro',
+    tier: 'standard',
+    inputPer1M: 2.0,
+    outputPer1M: 12.0,
+    contextTierBreakpoint: 200_000,
+    contextTierAbove: { inputPer1M: 4.0, outputPer1M: 18.0 },
+  },
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3-pro',
+    tier: 'batch',
+    inputPer1M: 1.0,
+    outputPer1M: 6.0,
+    contextTierBreakpoint: 200_000,
+    contextTierAbove: { inputPer1M: 2.0, outputPer1M: 9.0 },
+  },
+
+  // gemini-3-flash-preview — placeholder using 3.1 flash-lite sibling rates
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3-flash',
+    tier: 'standard',
+    inputPer1M: 0.25,
+    outputPer1M: 1.5,
+  },
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-3-flash',
+    tier: 'batch',
+    inputPer1M: 0.13,
+    outputPer1M: 0.75,
+  },
+
+  // ---------------------------------------------------------------------
+  // Vertex Gemini 2.5 (no context-tier breakpoint per public table)
+  // ---------------------------------------------------------------------
+
+  // gemini-2.5-pro standard: $1.25/M in, $10/M out
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-2.5-pro',
+    tier: 'standard',
+    inputPer1M: 1.25,
+    outputPer1M: 10.0,
+  },
+
+  // gemini-2.5-pro batch: ~50% of standard
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-2.5-pro',
+    tier: 'batch',
+    inputPer1M: 0.625,
+    outputPer1M: 5.0,
+  },
+
+  // gemini-2.5-flash standard: $0.075/M in, $0.30/M out
+  // (startup-builder/llm-vertex-batch source — supersedes icps's stale rate)
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-2.5-flash',
+    tier: 'standard',
+    inputPer1M: 0.075,
+    outputPer1M: 0.3,
+  },
+
+  // gemini-2.5-flash batch: ~50% of standard
+  {
+    provider: 'vertex',
+    slug: 'vertex/gemini-2.5-flash',
+    tier: 'batch',
+    inputPer1M: 0.0375,
+    outputPer1M: 0.15,
+  },
+
+  // ---------------------------------------------------------------------
+  // Bedrock Anthropic Claude (flat pricing — no context-tier breakpoint)
+  // ---------------------------------------------------------------------
+
+  // claude-opus-4-7: $15/M in, $75/M out
+  {
+    provider: 'bedrock',
+    slug: 'bedrock/claude-opus-4-7',
+    tier: 'standard',
+    inputPer1M: 15.0,
+    outputPer1M: 75.0,
+  },
+
+  // claude-opus-4-6: same as 4-7
+  {
+    provider: 'bedrock',
+    slug: 'bedrock/claude-opus-4-6',
+    tier: 'standard',
+    inputPer1M: 15.0,
+    outputPer1M: 75.0,
+  },
+
+  // claude-sonnet-4-7: $3/M in, $15/M out
+  {
+    provider: 'bedrock',
+    slug: 'bedrock/claude-sonnet-4-7',
+    tier: 'standard',
+    inputPer1M: 3.0,
+    outputPer1M: 15.0,
+  },
+
+  // claude-sonnet-4-6: same as 4-7
+  {
+    provider: 'bedrock',
+    slug: 'bedrock/claude-sonnet-4-6',
+    tier: 'standard',
+    inputPer1M: 3.0,
+    outputPer1M: 15.0,
+  },
+
+  // claude-haiku-4-5: $1/M in, $5/M out
+  {
+    provider: 'bedrock',
+    slug: 'bedrock/claude-haiku-4-5',
+    tier: 'standard',
+    inputPer1M: 1.0,
+    outputPer1M: 5.0,
+  },
+
+  // ---------------------------------------------------------------------
+  // Google AI Studio embeddings
+  // ---------------------------------------------------------------------
+
+  // gemini-embedding-2: $0.15/M input (no output side — pure embedding SKU)
+  {
+    provider: 'google-ai-studio',
+    slug: 'aistudio/gemini-embedding-2',
+    tier: 'standard',
+    inputPer1M: 0.15,
+    outputPer1M: 0,
+  },
+]

package/src/pricing/types.ts

@@ -0,0 +1,90 @@
+/**
+ * language-models / pricing — type definitions for the canonical pricing table.
+ *
+ * Schema design notes (sb-ncer 2026-05-07):
+ *
+ * 1. **Provider** is the upstream API surface, not the model family. AWS
+ *    Bedrock hosts Anthropic Claude, but `provider: 'bedrock'` (not
+ *    `'anthropic'`) — same model, different cost when consumed via
+ *    Anthropic's first-party API key (provider: 'anthropic' would be a
+ *    DIFFERENT row with potentially different rates).
+ *
+ * 2. **Slug** is the caller-facing string: `<provider>/<short-name>`.
+ *    Multiple slugs (the cascade short slug, the SDK-native id, etc.)
+ *    can map to the same logical SKU — but for the canonical primitive
+ *    we only carry the cascade short slug (`vertex/gemini-3.1-pro`,
+ *    `bedrock/claude-opus-4-7`). Adapter packages can layer their own
+ *    rewrite tables on top.
+ *
+ * 3. **Tier** distinguishes pricing modes for the same SKU:
+ *    - `standard`: synchronous interactive pricing (full price)
+ *    - `batch`: async batch-prediction pricing (typically 50% discount)
+ *    - `flex`: flex-tier pricing (Vertex's name for batch, kept as alias)
+ *    - `provisioned`: provisioned-throughput pricing (per-hour, not
+ *      currently modeled — placeholder for future PT entries)
+ *
+ * 4. **contextTierBreakpoint + contextTierAbove**: Gemini 3.x SKUs apply
+ *    a 2× rate above 200K input tokens; Anthropic Claude pricing is flat.
+ *    Optional fields — when absent, the base rate applies regardless of
+ *    input size.
+ */
+
+export type Provider = 'vertex' | 'bedrock' | 'openai' | 'anthropic' | 'google-ai-studio'
+
+export type PricingTier = 'standard' | 'batch' | 'flex' | 'provisioned'
+
+/** Rates expressed in USD per 1,000,000 tokens. */
+export interface RateBlock {
+  /** USD per 1M input tokens. */
+  readonly inputPer1M: number
+  /** USD per 1M output (completion) tokens. */
+  readonly outputPer1M: number
+  /**
+   * Optional USD per 1M cached input tokens (prompt-caching tier). When
+   * absent, callers should fall back to inputPer1M (no cache discount).
+   */
+  readonly cachedInputPer1M?: number
+}
+
+/**
+ * Single canonical pricing row. Identity is `(slug, tier)` — provider is
+ * derived from the slug prefix and stored explicitly only for tooling
+ * convenience.
+ */
+export interface ModelPricing extends RateBlock {
+  readonly provider: Provider
+  readonly slug: string
+  readonly tier: PricingTier
+  /**
+   * Token count at which pricing changes (inclusive — i.e. inputs >=
+   * breakpoint use contextTierAbove). Currently only Gemini 3.x SKUs
+   * have a breakpoint (200_000). Anthropic Claude has flat pricing.
+   */
+  readonly contextTierBreakpoint?: number
+  /** Rates that apply when inputTokens >= contextTierBreakpoint. */
+  readonly contextTierAbove?: RateBlock
+}
+
+export interface PriceForArgs {
+  readonly slug: string
+  readonly tier: PricingTier
+  readonly inputTokens: number
+  readonly outputTokens: number
+  /**
+   * Optional cached input tokens (subset of inputTokens that hit a
+   * prompt-caching tier). Billed at cachedInputPer1M when the row
+   * defines it; otherwise at inputPer1M (i.e. no discount).
+   */
+  readonly cachedInputTokens?: number
+}
+
+export interface PriceForResult {
+  readonly inputUsd: number
+  readonly outputUsd: number
+  readonly totalUsd: number
+}
+
+export interface HasPricingArgs {
+  readonly slug: string
+  readonly tier: PricingTier
+}

package/{src → test}/aliases.test.ts

@@ -6,7 +6,7 @@
  */
 
 import { describe, it, expect } from 'vitest'
-import { ALIASES } from './aliases.js'
+import { ALIASES } from '../src/aliases.js'
 
 describe('ALIASES', () => {
   it('is an object', () => {
@@ -217,7 +217,7 @@ describe('ALIASES', () => {
     })
 
     it('has unique lowercase keys', () => {
-      const lowerKeys = Object.keys(ALIASES).map(k => k.toLowerCase())
+      const lowerKeys = Object.keys(ALIASES).map((k) => k.toLowerCase())
       const uniqueLowerKeys = new Set(lowerKeys)
       expect(lowerKeys.length).toBe(uniqueLowerKeys.size)
     })
@@ -249,9 +249,7 @@ describe('ALIASES', () => {
 
   describe('provider coverage', () => {
     it('covers major AI providers', () => {
-      const providers = new Set(
-        Object.values(ALIASES).map(v => v.split('/')[0])
-      )
+      const providers = new Set(Object.values(ALIASES).map((v) => v.split('/')[0]))
 
       expect(providers.has('anthropic')).toBe(true)
       expect(providers.has('openai')).toBe(true)
@@ -278,29 +276,29 @@ describe('ALIASES', () => {
     it('matches all aliases documented in README', () => {
       // These are the aliases listed in the README.md table
       const documentedAliases = {
-        'opus': 'anthropic/claude-opus-4.5',
-        'sonnet': 'anthropic/claude-sonnet-4.5',
-        'haiku': 'anthropic/claude-haiku-4.5',
-        'claude': 'anthropic/claude-sonnet-4.5',
-        'gpt': 'openai/gpt-4o',
+        opus: 'anthropic/claude-opus-4.5',
+        sonnet: 'anthropic/claude-sonnet-4.5',
+        haiku: 'anthropic/claude-haiku-4.5',
+        claude: 'anthropic/claude-sonnet-4.5',
+        gpt: 'openai/gpt-4o',
         'gpt-4o': 'openai/gpt-4o',
         '4o': 'openai/gpt-4o',
-        'o1': 'openai/o1',
-        'o3': 'openai/o3',
+        o1: 'openai/o1',
+        o3: 'openai/o3',
         'o3-mini': 'openai/o3-mini',
-        'gemini': 'google/gemini-2.5-flash',
-        'flash': 'google/gemini-2.5-flash',
+        gemini: 'google/gemini-2.5-flash',
+        flash: 'google/gemini-2.5-flash',
         'gemini-pro': 'google/gemini-2.5-pro',
-        'llama': 'meta-llama/llama-4-maverick',
+        llama: 'meta-llama/llama-4-maverick',
         'llama-4': 'meta-llama/llama-4-maverick',
         'llama-70b': 'meta-llama/llama-3.3-70b-instruct',
-        'mistral': 'mistralai/mistral-large-2411',
-        'codestral': 'mistralai/codestral-2501',
-        'deepseek': 'deepseek/deepseek-chat',
-        'r1': 'deepseek/deepseek-r1',
-        'qwen': 'qwen/qwen3-235b-a22b',
-        'grok': 'x-ai/grok-3',
-        'sonar': 'perplexity/sonar-pro',
+        mistral: 'mistralai/mistral-large-2411',
+        codestral: 'mistralai/codestral-2501',
+        deepseek: 'deepseek/deepseek-chat',
+        r1: 'deepseek/deepseek-r1',
+        qwen: 'qwen/qwen3-235b-a22b',
+        grok: 'x-ai/grok-3',
+        sonar: 'perplexity/sonar-pro',
       }
 
       for (const [alias, expectedId] of Object.entries(documentedAliases)) {

package/{src → test}/index.test.ts

@@ -17,7 +17,7 @@ import {
   type ProviderEndpoint,
   type ResolvedModel,
   type DirectProvider,
-} from './index.js'
+} from '../src/index.js'
 
 describe('package exports', () => {
   it('exports resolve function', () => {
@@ -166,8 +166,8 @@ describe('end-to-end workflows', () => {
       expect(allModels.length).toBeGreaterThanOrEqual(0)
 
       if (allModels.length > 0) {
-        const anthropicModels = allModels.filter(m => m.id.startsWith('anthropic/'))
-        const openaiModels = allModels.filter(m => m.id.startsWith('openai/'))
+        const anthropicModels = allModels.filter((m) => m.id.startsWith('anthropic/'))
+        const openaiModels = allModels.filter((m) => m.id.startsWith('openai/'))
 
         if (anthropicModels.length > 0) {
           expect(anthropicModels[0].id).toContain('anthropic/')
@@ -182,7 +182,7 @@ describe('end-to-end workflows', () => {
       const allModels = list()
 
       if (allModels.length > 0) {
-        const directModels = allModels.filter(m => {
+        const directModels = allModels.filter((m) => {
           const provider = m.id.split('/')[0]
           return (DIRECT_PROVIDERS as readonly string[]).includes(provider)
         })
@@ -282,8 +282,8 @@ describe('end-to-end workflows', () => {
       expect(Array.isArray(claudeModels)).toBe(true)
       if (claudeModels.length > 0) {
         expect(
-          claudeModels.some(m =>
-            m.id.includes('claude') || m.name.toLowerCase().includes('claude')
+          claudeModels.some(
+            (m) => m.id.includes('claude') || m.name.toLowerCase().includes('claude')
           )
         ).toBe(true)
       }
@@ -329,7 +329,7 @@ describe('end-to-end workflows', () => {
     it('models may have architecture info', () => {
       const models = list()
       if (models.length > 0) {
-        const modelWithArch = models.find(m => m.architecture)
+        const modelWithArch = models.find((m) => m.architecture)
         if (modelWithArch?.architecture) {
           expect(modelWithArch.architecture.modality).toBeDefined()
           expect(Array.isArray(modelWithArch.architecture.input_modalities)).toBe(true)
@@ -344,7 +344,7 @@ describe('end-to-end workflows', () => {
       const directProviders = ['anthropic', 'openai', 'google']
 
       for (const provider of directProviders) {
-        const models = list().filter(m => m.id.startsWith(`${provider}/`))
+        const models = list().filter((m) => m.id.startsWith(`${provider}/`))
         if (models.length > 0) {
           const resolved = resolveWithProvider(models[0].id)
           expect(resolved.supportsDirectRouting).toBe(true)
@@ -355,7 +355,7 @@ describe('end-to-end workflows', () => {
 
     it('identifies non-direct routing providers', () => {
       const models = list()
-      const nonDirectModel = models.find(m => {
+      const nonDirectModel = models.find((m) => {
         const provider = m.id.split('/')[0]
         return !(DIRECT_PROVIDERS as readonly string[]).includes(provider)
       })

package/{src → test}/models.test.ts

@@ -14,8 +14,8 @@ import {
   DIRECT_PROVIDERS,
   type ModelInfo,
   type ResolvedModel,
-} from './models.js'
-import { ALIASES } from './aliases.js'
+} from '../src/models.js'
+import { ALIASES } from '../src/aliases.js'
 
 describe('list', () => {
   it('returns an array of models', () => {
@@ -93,7 +93,7 @@ describe('search', () => {
       const idPart = model.id.split('/')[0] // Provider name
       const results = search(idPart)
       expect(results.length).toBeGreaterThan(0)
-      expect(results.some(m => m.id.includes(idPart))).toBe(true)
+      expect(results.some((m) => m.id.includes(idPart))).toBe(true)
     }
   })
 
@@ -123,11 +123,13 @@ describe('search', () => {
     const models = list()
     if (models.length > 0) {
       // Find a model and search for part of its name
-      const model = models.find(m => m.name.includes(' '))
+      const model = models.find((m) => m.name.includes(' '))
       if (model) {
         const namePart = model.name.split(' ')[0].toLowerCase()
         const results = search(namePart)
-        expect(results.some(m => m.id === model.id || m.name.toLowerCase().includes(namePart))).toBe(true)
+        expect(
+          results.some((m) => m.id === model.id || m.name.toLowerCase().includes(namePart))
+        ).toBe(true)
       }
     }
   })
@@ -299,7 +301,7 @@ describe('resolveWithProvider', () => {
   it('identifies non-direct providers', () => {
     // Use a model from a provider not in DIRECT_PROVIDERS
     const models = list()
-    const nonDirectModel = models.find(m => {
+    const nonDirectModel = models.find((m) => {
       const provider = m.id.split('/')[0]
       return !(DIRECT_PROVIDERS as readonly string[]).includes(provider)
     })