modelmix 4.7.4 → 5.0.0

package/test/tokens.test.js

@@ -1,5 +1,5 @@
 import { expect } from 'chai';
-import { ModelMix, MixAnthropic, MixCustom, MixGoogle, MixMiMo, MixOpenAIResponses, MixOpenRouter } from '../index.js';
+import { ModelMix, MixAnthropic, MixCustom, MixGoogle, MixMiMo, MixOpenAI, MixOpenAIResponses, MixOpenRouter } from '../index.js';
 import { createRequire } from 'module';
 
 const require = createRequire(import.meta.url);
@@ -18,14 +18,15 @@ describe('Token Usage Tracking', () => {
         nock.activate();
     });
 
-    it('should extract cached tokens from supported provider usage formats', function () {
+    it('should normalize prompt cache usage from supported provider formats', function () {
         const openAIChatTokens = MixCustom.extractTokens({
             usage: {
                 prompt_tokens: 120,
                 completion_tokens: 30,
                 total_tokens: 150,
                 prompt_tokens_details: {
-                    cached_tokens: 80
+                    cached_tokens: 80,
+                    cache_write_tokens: 10
                 }
             }
         });
@@ -35,7 +36,8 @@ describe('Token Usage Tracking', () => {
                 output_tokens: 20,
                 total_tokens: 110,
                 input_tokens_details: {
-                    cached_tokens: 45
+                    cached_tokens: 45,
+                    cache_write_tokens: 15
                 }
             }
         });
@@ -43,7 +45,8 @@ describe('Token Usage Tracking', () => {
             usage: {
                 input_tokens: 60,
                 output_tokens: 15,
-                cache_read_input_tokens: 25
+                cache_read_input_tokens: 25,
+                cache_creation_input_tokens: 10
             }
         });
         const googleTokens = MixGoogle.extractTokens({
@@ -55,10 +58,69 @@ describe('Token Usage Tracking', () => {
             }
         });
 
-        expect(openAIChatTokens.cached).to.equal(80);
-        expect(openAIResponsesTokens.cached).to.equal(45);
-        expect(anthropicTokens.cached).to.equal(25);
-        expect(googleTokens.cached).to.equal(35);
+        expect(openAIChatTokens).to.include({
+            input: 120,
+            output: 30,
+            total: 150,
+            cached: 80,
+            cacheWrite: 10,
+            uncachedInput: 30,
+            cacheHitRate: 0.6667
+        });
+        expect(openAIResponsesTokens).to.include({
+            input: 90,
+            output: 20,
+            total: 110,
+            cached: 45,
+            cacheWrite: 15,
+            uncachedInput: 30,
+            cacheHitRate: 0.5
+        });
+        expect(anthropicTokens).to.include({
+            input: 95,
+            output: 15,
+            total: 110,
+            cached: 25,
+            cacheWrite: 10,
+            uncachedInput: 60,
+            cacheHitRate: 0.2632
+        });
+        expect(googleTokens).to.include({
+            input: 70,
+            output: 10,
+            total: 80,
+            cached: 35,
+            cacheWrite: 0,
+            uncachedInput: 35,
+            cacheHitRate: 0.5
+        });
+    });
+
+    it('should return zero for prompt cache categories omitted by the provider', function () {
+        const tokens = MixCustom.extractTokens({
+            usage: {
+                prompt_tokens: 20,
+                completion_tokens: 5,
+                total_tokens: 25
+            }
+        });
+
+        expect(tokens).to.include({
+            cached: 0,
+            cacheWrite: 0,
+            uncachedInput: 20,
+            cacheHitRate: 0,
+            cost: 0
+        });
+        expect(tokens.costBreakdown).to.deep.equal({
+            uncachedInput: 0,
+            cachedInput: 0,
+            cacheWrite: 0,
+            cacheWrite5m: 0,
+            cacheWrite1h: 0,
+            output: 0,
+            total: 0
+        });
     });
 
     it('should pass OpenAI Responses prompt cache options through the request body', function () {
@@ -76,6 +138,186 @@ describe('Token Usage Tracking', () => {
         expect(request.prompt_cache_retention).to.equal('24h');
     });
 
+    it('should pass GPT-5.6 explicit cache controls and preserve block breakpoints', function () {
+        const breakpoint = { mode: 'explicit' };
+        const request = MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [{
+                role: 'developer',
+                content: [
+                    { type: 'text', text: 'Stable instructions', prompt_cache_breakpoint: breakpoint },
+                    {
+                        type: 'image',
+                        source: { type: 'base64', media_type: 'image/png', data: 'AAAA' },
+                        prompt_cache_breakpoint: breakpoint
+                    },
+                    {
+                        type: 'input_file',
+                        file_id: 'file_123',
+                        prompt_cache_breakpoint: breakpoint
+                    }
+                ]
+            }],
+            prompt_cache_key: 'explicit-cache',
+            prompt_cache_options: { mode: 'explicit', ttl: '30m' }
+        });
+
+        expect(request.prompt_cache_options).to.deep.equal({ mode: 'explicit', ttl: '30m' });
+        expect(request.input[0].content).to.deep.equal([
+            { type: 'input_text', text: 'Stable instructions', prompt_cache_breakpoint: breakpoint },
+            {
+                type: 'input_image',
+                image_url: 'data:image/png;base64,AAAA',
+                prompt_cache_breakpoint: breakpoint
+            },
+            {
+                type: 'input_file',
+                file_id: 'file_123',
+                prompt_cache_breakpoint: breakpoint
+            }
+        ]);
+    });
+
+    it('should translate neutral cache breakpoints for GPT-5.6 and filter them for older models', async function () {
+        const breakpoint = { mode: 'explicit' };
+        const model = ModelMix.new()
+            .addText('Stable text', { role: 'developer', cache: { breakpoint: true } })
+            .addImageFromUrl('data:image/png;base64,AAAA', { cache: { breakpoint: true } });
+        const messages = await model.prepareMessages();
+        const gpt56Request = MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages,
+            prompt_cache_options: { mode: 'explicit', ttl: '30m' }
+        });
+        const olderRequest = MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.4',
+            messages
+        });
+
+        expect(messages[0].content[0]).to.deep.include({ cache: { breakpoint: true } });
+        expect(messages[0].content[0]).to.not.have.property('prompt_cache_breakpoint');
+        expect(gpt56Request.input[0].content[0].prompt_cache_breakpoint).to.deep.equal(breakpoint);
+        expect(gpt56Request.input[1].content[0]).to.deep.equal({
+            type: 'input_image',
+            image_url: 'data:image/png;base64,AAAA',
+            prompt_cache_breakpoint: breakpoint
+        });
+        expect(olderRequest.input[0].content[0]).to.not.have.property('cache');
+        expect(olderRequest.input[0].content[0]).to.not.have.property('prompt_cache_breakpoint');
+        expect(olderRequest.input[1].content[0]).to.not.have.property('prompt_cache_breakpoint');
+    });
+
+    it('should map assistant history to Responses output_text content', function () {
+        const request = MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [
+                { role: 'user', content: [{ type: 'text', text: 'My name is Alice' }] },
+                { role: 'assistant', content: [{ type: 'text', text: 'Nice to meet you' }] },
+                { role: 'user', content: [{ type: 'text', text: 'What is my name?' }] }
+            ]
+        });
+
+        expect(request.input[0].content[0]).to.deep.equal({
+            type: 'input_text',
+            text: 'My name is Alice'
+        });
+        expect(request.input[1].content[0]).to.deep.equal({
+            type: 'output_text',
+            text: 'Nice to meet you'
+        });
+        expect(request.input[2].content[0]).to.deep.equal({
+            type: 'input_text',
+            text: 'What is my name?'
+        });
+    });
+
+    it('should not treat the native OpenAI breakpoint as a fluent API alias', async function () {
+        const nativeBreakpoint = { prompt_cache_breakpoint: { mode: 'explicit' } };
+        const model = ModelMix.new()
+            .addText('Stable text', nativeBreakpoint)
+            .addImageFromUrl('data:image/png;base64,AAAA', nativeBreakpoint);
+        const messages = await model.prepareMessages();
+        const request = MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages,
+            prompt_cache_options: { mode: 'explicit', ttl: '30m' }
+        });
+
+        expect(messages[0].content).to.have.length(2);
+        expect(messages[0].content.every(item => item.cache === undefined)).to.equal(true);
+        const content = request.input.flatMap(item => item.content ?? []);
+        expect(content).to.have.length(2);
+        expect(content.every(item => item.prompt_cache_breakpoint === undefined)).to.equal(true);
+    });
+
+    it('should strip neutral and Anthropic cache metadata from OpenAI-compatible chat requests', function () {
+        const messages = [{
+            role: 'user',
+            content: [{
+                type: 'text',
+                text: 'Stable text',
+                cache: { breakpoint: true },
+                cache_control: { type: 'ephemeral' }
+            }]
+        }];
+        const converted = MixOpenAI.convertMessages(messages, { system: 'System' });
+
+        expect(converted[1].content[0]).to.deep.equal({ type: 'text', text: 'Stable text' });
+        expect(messages[0].content[0]).to.deep.include({ cache: { breakpoint: true } });
+    });
+
+    it('should reject prompt cache controls unsupported by the selected OpenAI model', function () {
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [{ role: 'user', content: 'Hi' }],
+            prompt_cache_retention: '24h'
+        })).to.throw('prompt_cache_options.ttl');
+
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.4',
+            messages: [{ role: 'user', content: 'Hi' }],
+            prompt_cache_options: { mode: 'explicit', ttl: '30m' }
+        })).to.throw('only supported by GPT-5.6');
+
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.4',
+            messages: [{
+                role: 'user',
+                content: [{
+                    type: 'text',
+                    text: 'Hi',
+                    prompt_cache_breakpoint: { mode: 'explicit' }
+                }]
+            }]
+        })).to.throw('prompt_cache_breakpoint is only supported by GPT-5.6');
+    });
+
+    it('should validate GPT-5.6 prompt cache option values', function () {
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [{ role: 'user', content: 'Hi' }],
+            prompt_cache_options: { mode: 'automatic', ttl: '30m' }
+        })).to.throw('mode must be "implicit" or "explicit"');
+
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [{ role: 'user', content: 'Hi' }],
+            prompt_cache_options: { mode: 'explicit', ttl: '1h' }
+        })).to.throw('ttl must be "30m"');
+
+        expect(() => MixOpenAIResponses.buildResponsesRequest({
+            model: 'gpt-5.6-luna',
+            messages: [{
+                role: 'user',
+                content: [{
+                    type: 'text',
+                    text: 'Hi',
+                    prompt_cache_breakpoint: { mode: 'implicit' }
+                }]
+            }]
+        })).to.throw('breakpoint mode must be "explicit"');
+    });
+
     it('should register GPT-5.5 shortcuts with OpenAI Responses provider', function () {
         const model = ModelMix.new()
             .gpt55()
@@ -100,9 +342,160 @@ describe('Token Usage Tracking', () => {
             'gpt-5.6-luna'
         ]);
         expect(model.models.every(({ provider }) => provider instanceof MixOpenAIResponses)).to.equal(true);
-        expect(ModelMix.calculateCost('gpt-5.6-sol', { input: 1_000_000, output: 1_000_000 })).to.equal(35);
-        expect(ModelMix.calculateCost('gpt-5.6-terra', { input: 1_000_000, output: 1_000_000 })).to.equal(17.5);
-        expect(ModelMix.calculateCost('gpt-5.6-luna', { input: 1_000_000, output: 1_000_000 })).to.equal(7);
+        expect(ModelMix.calculateCost('gpt-5.6-sol', { input: 1_000_000, output: 1_000_000 })).to.equal(55);
+        expect(ModelMix.calculateCost('gpt-5.6-terra', { input: 1_000_000, output: 1_000_000 })).to.equal(22);
+        expect(ModelMix.calculateCost('gpt-5.6-luna', { input: 1_000_000, output: 1_000_000 })).to.equal(2.2);
+    });
+
+    it('should calculate GPT-5.6 cache reads and writes at their actual rates', function () {
+        const tokens = ModelMix.normalizeTokenUsage({
+            input: 1200,
+            output: 50,
+            cached: 800,
+            cacheWrite: 200
+        });
+
+        expect(ModelMix.calculateCostBreakdown('gpt-5.6-terra', tokens)).to.deep.equal({
+            uncachedInput: 0.0004,
+            cachedInput: 0.00016,
+            cacheWrite: 0.0005,
+            cacheWrite5m: 0,
+            cacheWrite1h: 0,
+            output: 0.0006,
+            total: 0.00166
+        });
+        expect(ModelMix.calculateCost('gpt-5.6-terra', tokens)).to.equal(0.00166);
+    });
+
+    it('should apply GPT-5.6 long-context multipliers to the entire request', function () {
+        const tokens = ModelMix.normalizeTokenUsage({
+            input: 300_000,
+            output: 1_000,
+            cached: 100_000,
+            cacheWrite: 100_000
+        });
+
+        expect(ModelMix.calculateCostBreakdown('gpt-5.6-luna', tokens)).to.deep.equal({
+            uncachedInput: 0.04,
+            cachedInput: 0.004,
+            cacheWrite: 0.05,
+            cacheWrite5m: 0,
+            cacheWrite1h: 0,
+            output: 0.0018,
+            total: 0.0958
+        });
+        expect(ModelMix.calculateCacheMetrics('gpt-5.6-luna', tokens)).to.deep.equal({
+            cacheSavings: 0.036,
+            cacheWritePremium: 0.01,
+            breakEvenHits: 0.2778
+        });
+    });
+
+    it('should not apply GPT-5.6 long-context multipliers at exactly 272K input tokens', function () {
+        const tokens = ModelMix.normalizeTokenUsage({
+            input: 272_000,
+            output: 1_000,
+            cached: 100_000,
+            cacheWrite: 100_000
+        });
+
+        expect(ModelMix.calculateCostBreakdown('gpt-5.6-luna', tokens)).to.deep.equal({
+            uncachedInput: 0.0144,
+            cachedInput: 0.002,
+            cacheWrite: 0.025,
+            cacheWrite5m: 0,
+            cacheWrite1h: 0,
+            output: 0.0012,
+            total: 0.0426
+        });
+    });
+
+    it('should price Anthropic 5-minute and 1-hour cache writes separately', function () {
+        const tokens = MixAnthropic.extractTokens({
+            usage: {
+                input_tokens: 10,
+                output_tokens: 5,
+                cache_read_input_tokens: 400,
+                cache_creation_input_tokens: 300,
+                cache_creation: {
+                    ephemeral_5m_input_tokens: 100,
+                    ephemeral_1h_input_tokens: 200
+                }
+            }
+        });
+
+        expect(tokens).to.include({
+            input: 710,
+            cached: 400,
+            cacheWrite: 300,
+            cacheWrite5m: 100,
+            cacheWrite1h: 200,
+            uncachedInput: 10
+        });
+        expect(ModelMix.calculateCostBreakdown('claude-haiku-4-5-20251001', tokens)).to.deep.equal({
+            uncachedInput: 0.00001,
+            cachedInput: 0.00004,
+            cacheWrite: 0.000525,
+            cacheWrite5m: 0.000125,
+            cacheWrite1h: 0.0004,
+            output: 0.000025,
+            total: 0.0006
+        });
+        expect(ModelMix.calculateCacheMetrics('claude-haiku-4-5-20251001', tokens)).to.deep.equal({
+            cacheSavings: 0.00036,
+            cacheWritePremium: 0.000225,
+            breakEvenHits: 0.8333
+        });
+    });
+
+    it('should expose normalized cache costs through raw() and lastRaw', async function () {
+        const provider = new MixCustom();
+        provider.create = async () => ({
+            message: 'ok',
+            think: null,
+            toolCalls: [],
+            tokens: MixOpenAIResponses.extractResponsesTokens({
+                usage: {
+                    input_tokens: 1200,
+                    output_tokens: 50,
+                    total_tokens: 1250,
+                    input_tokens_details: {
+                        cached_tokens: 1024,
+                        cache_write_tokens: 0
+                    }
+                }
+            }),
+            response: {}
+        });
+
+        const model = ModelMix.new()
+            .attach('gpt-5.6-luna', provider)
+            .addText('test');
+        const result = await model.raw();
+
+        expect(result.tokens).to.include({
+            input: 1200,
+            output: 50,
+            total: 1250,
+            cached: 1024,
+            cacheWrite: 0,
+            uncachedInput: 176,
+            cacheHitRate: 0.8533,
+            cacheSavings: 0.00018432,
+            cacheWritePremium: 0,
+            breakEvenHits: 0,
+            cost: 0.00011568
+        });
+        expect(result.tokens.costBreakdown).to.deep.equal({
+            uncachedInput: 0.0000352,
+            cachedInput: 0.00002048,
+            cacheWrite: 0,
+            cacheWrite5m: 0,
+            cacheWrite1h: 0,
+            output: 0.00006,
+            total: 0.00011568
+        });
+        expect(model.lastRaw.tokens).to.deep.equal(result.tokens);
     });
 
     it('should register Gemini Flash shortcuts with Google provider', function () {
@@ -182,7 +575,7 @@ describe('Token Usage Tracking', () => {
         this.timeout(30000);
 
         const model = ModelMix.new()
-            .gpt5nano()
+            .gpt56luna()
             .addText('Say hi');
 
         const result = await model.raw();
@@ -247,7 +640,7 @@ describe('Token Usage Tracking', () => {
         this.timeout(60000);
 
         const conversation = ModelMix.new({ config: { max_history: 10 } })
-            .gpt5nano();
+            .gpt56luna();
 
         // First turn
         conversation.addText('My name is Alice');
@@ -272,7 +665,7 @@ describe('Token Usage Tracking', () => {
         this.timeout(30000);
 
         const model = ModelMix.new()
-            .gpt5nano()
+            .gpt56luna()
             .addText('Return a simple greeting');
 
         // Using raw() to get token info
@@ -288,7 +681,7 @@ describe('Token Usage Tracking', () => {
         this.timeout(90000);
 
         const providers = [
-            { name: 'OpenAI', create: (m) => m.gpt5nano() },
+            { name: 'OpenAI', create: (m) => m.gpt56luna() },
             { name: 'Anthropic', create: (m) => m.haiku45() },
             { name: 'Google', create: (m) => m.gemini3flash() }
         ];