From 043a8416b0177a8596d5ab979f43e5ed3d52b6bf Mon Sep 17 00:00:00 2001 From: Mario Zechner Date: Tue, 16 Dec 2025 23:01:46 +0100 Subject: [PATCH] Update image limits test with comprehensive 8k stress test results Tested max 8kx8k images per provider: - Anthropic: 100 (explicit limit, fails at 101) - OpenAI: 100-200 (100 works, 200 times out) - Mistral: 8 (explicit limit, fails at 9) - xAI: 100-150 (100 works, 150 times out) - Groq: 0 (8k exceeds 33M pixel limit) - zAI: 400 (context window limited at 500) - OpenRouter: 40 (context window limited at 50) - Gemini: untested (no API key in test env) Key finding: Anthropic's 'many images' rule does NOT cause API errors. 100 x 8kx8k images work fine. Anthropic likely auto-resizes internally. Related to #120 --- packages/ai/test/image-limits.test.ts | 93 ++++++++++++++------------- 1 file changed, 49 insertions(+), 44 deletions(-) diff --git a/packages/ai/test/image-limits.test.ts b/packages/ai/test/image-limits.test.ts index 9a0396c0..7be5cdd5 100644 --- a/packages/ai/test/image-limits.test.ts +++ b/packages/ai/test/image-limits.test.ts @@ -11,33 +11,32 @@ * DISCOVERED LIMITS (Dec 2025): * ============================================================================ * - * | Provider | Model | Max Images | Max Size | Max Dimension | Max 8k Images | - * |-------------|--------------------|------------|----------|---------------|---------------| - * | Anthropic | claude-3-5-haiku | 100 | 5MB | 8000px | 100 | - * | OpenAI | gpt-4o-mini | 500 | ≥25MB | ≥20000px | ≥50 | - * | Gemini | gemini-2.5-flash | ~2000+* | ≥40MB | 8000px | (untested) | - * | Mistral | pixtral-12b | 8 | ~15MB | 8000px | 8 | - * | OpenRouter | z-ai/glm-4.5v | ~40** | ~10MB | ≥20000px | ≥10 | - * | xAI | grok-2-vision | ≥100 | 25MB | 8000px | ≥50 | - * | Groq | llama-4-scout-17b | 5 | ~5MB | ~5760px*** | 0**** | - * | zAI | glm-4.5v | ≥100 | ≥20MB | 8000px | ≥50 | + * | Provider | Model | Max Images | Max Size | Max Dim | Max 8k Imgs | + * |-------------|--------------------|------------|----------|----------|-------------| + * | Anthropic | claude-3-5-haiku | 100 | 5MB | 8000px | 100 | + * | OpenAI | gpt-4o-mini | 500 | ≥25MB | ≥20000px | 100-200* | + * | Gemini | gemini-2.5-flash | ~2000** | ≥40MB | 8000px | (untested) | + * | Mistral | pixtral-12b | 8 | ~15MB | 8000px | 8 | + * | xAI | grok-2-vision | ≥100 | 25MB | 8000px | 100-150* | + * | Groq | llama-4-scout-17b | 5 | ~5MB | ~5760px | 0*** | + * | zAI | glm-4.5v | ≥100 | ≥20MB | 8000px | 400**** | + * | OpenRouter | z-ai/glm-4.5v | ~40**** | ~10MB | ≥20000px | 40**** | * * Notes: * - Anthropic: Docs mention a "many images" rule (>20 images = 2000px max), * but testing shows 100 x 8k images work fine. Anthropic may auto-resize - * internally. Total request size capped at 32MB. - * - OpenAI: Documented limit is 20MB, but we observed ≥25MB working. - * No dimension limit found up to 20000px. - * - Gemini: * Very permissive on count, hits rate limits before image limits. - * Dimension limit is 8000px (same as Anthropic). - * - Mistral: Very restrictive on image count (only 8 images allowed). - * Dimension limit is 8000px. - * - OpenRouter: ** Limited by context window (65k tokens), not explicit image - * limit. No dimension limit found up to 20000px. - * - xAI: 25MB limit (26214400 bytes exactly). Dimension limit ~8000px. - * - Groq: *** Very restrictive. Max 5 images, ~5MB size, 33177600 pixels max - * (≈5760x5760). **** 8k images exceed pixel limit so 0 supported. - * - zAI: Permissive, similar to Anthropic limits. + * internally. Total request size capped at 32MB. Explicit error at 101+. + * - OpenAI: * 100 x 8k succeeded, 200 x 8k failed with timeout. Actual limit + * likely between 100-200. Documented size limit is 20MB but ≥25MB works. + * - Gemini: ** Very permissive on count, hits rate limits before image limits. + * - Mistral: Very restrictive (8 images max). Explicit error at 9+. + * - xAI: * 100 x 8k succeeded, 150 x 8k timed out. 25MB size limit exact. + * - Groq: *** Most restrictive. 5 images max, 33177600 pixels max (≈5760x5760). + * 8k images (64M pixels) exceed limit, so 0 supported. + * - zAI: **** Context-window limited (65536 tokens). 400 x 8k succeeded, + * 500 x 8k exceeded token limit. + * - OpenRouter: **** Context-window limited (65536 tokens), not explicit + * image limit. 40 x 8k succeeded, 50 x 8k exceeded token limit. * * ============================================================================ * PRACTICAL RECOMMENDATIONS FOR CODING AGENTS: @@ -858,14 +857,14 @@ describe("Image Limits E2E Tests", () => { ); }); - // Anthropic - known to have "many images" rule (>20 images = 2000px max dimension) - // Testing to find actual limit with 8k dimension images + // Anthropic - known 100 image limit, testing if 8k dimensions change this it.skipIf(!process.env.ANTHROPIC_API_KEY)( "Anthropic: max 8k images before rejection", - { timeout: 600000 }, + { timeout: 900000 }, async () => { const model = getModel("anthropic", "claude-3-5-haiku-20241022"); - const counts = [5, 10, 15, 20, 21, 25, 30, 40, 50, 60, 80, 100]; + // Known limit is 100 images - test around that boundary + const counts = [10, 20, 50, 80, 100, 110, 120]; let lastSuccess = 0; let lastError: string | undefined; @@ -888,13 +887,14 @@ describe("Image Limits E2E Tests", () => { }, ); - // OpenAI + // OpenAI - known 500 image limit it.skipIf(!process.env.OPENAI_API_KEY)( "OpenAI: max 8k images before rejection", - { timeout: 600000 }, + { timeout: 1800000 }, async () => { const model = getModel("openai", "gpt-4o-mini"); - const counts = [5, 10, 20, 30, 50]; + // Known limit is 500 images - test around that boundary + const counts = [50, 100, 200, 300, 400, 500, 550]; let lastSuccess = 0; let lastError: string | undefined; @@ -917,13 +917,14 @@ describe("Image Limits E2E Tests", () => { }, ); - // Gemini + // Gemini - known to be very permissive (~2000+ small images), but 8k may differ it.skipIf(!process.env.GOOGLE_API_KEY)( "Gemini: max 8k images before rejection", - { timeout: 600000 }, + { timeout: 1800000 }, async () => { const model = getModel("google", "gemini-2.5-flash"); - const counts = [5, 10, 20, 30, 50]; + // Test progressively - 8k images are large so limit may be lower + const counts = [10, 50, 100, 200, 500, 1000]; let lastSuccess = 0; let lastError: string | undefined; @@ -946,13 +947,14 @@ describe("Image Limits E2E Tests", () => { }, ); - // Mistral - already very limited (8 images max) + // Mistral - known 8 image limit it.skipIf(!process.env.MISTRAL_API_KEY)( "Mistral: max 8k images before rejection", { timeout: 600000 }, async () => { const model = getModel("mistral", "pixtral-12b"); - const counts = [1, 2, 4, 6, 8]; + // Known limit is 8 images - test around that boundary + const counts = [1, 2, 4, 6, 8, 9, 10]; let lastSuccess = 0; let lastError: string | undefined; @@ -975,10 +977,11 @@ describe("Image Limits E2E Tests", () => { }, ); - // xAI - it.skipIf(!process.env.XAI_API_KEY)("xAI: max 8k images before rejection", { timeout: 600000 }, async () => { + // xAI - tested up to 100 small images successfully + it.skipIf(!process.env.XAI_API_KEY)("xAI: max 8k images before rejection", { timeout: 1200000 }, async () => { const model = getModel("xai", "grok-2-vision"); - const counts = [5, 10, 20, 30, 50]; + // Test around the expected boundary + const counts = [10, 50, 100, 150, 200]; let lastSuccess = 0; let lastError: string | undefined; @@ -1031,10 +1034,11 @@ describe("Image Limits E2E Tests", () => { }, ); - // zAI - it.skipIf(!process.env.ZAI_API_KEY)("zAI: max 8k images before rejection", { timeout: 600000 }, async () => { + // zAI - tested up to 100 small images successfully, very permissive + it.skipIf(!process.env.ZAI_API_KEY)("zAI: max 8k images before rejection", { timeout: 1800000 }, async () => { const model = getModel("zai", "glm-4.5v"); - const counts = [5, 10, 20, 30, 50]; + // Very permissive - extend to find actual limit + const counts = [50, 100, 200, 300, 400, 500]; let lastSuccess = 0; let lastError: string | undefined; @@ -1056,13 +1060,14 @@ describe("Image Limits E2E Tests", () => { expect(lastSuccess).toBeGreaterThanOrEqual(5); }); - // OpenRouter + // OpenRouter - context-window limited (~40 small images), 8k will be fewer it.skipIf(!process.env.OPENROUTER_API_KEY)( "OpenRouter: max 8k images before rejection", - { timeout: 600000 }, + { timeout: 900000 }, async () => { const model = getModel("openrouter", "z-ai/glm-4.5v"); - const counts = [1, 2, 3, 5, 10]; + // 8k images consume more tokens, so limit will be lower than 40 + const counts = [1, 2, 5, 10, 20, 30, 40, 50]; let lastSuccess = 0; let lastError: string | undefined;