241 viewsEnglishComparison
AI API pricing changes fast. New models launch monthly, prices drop, and keeping track of what each provider charges is a job in itself. This is our May 2026 edition of the comprehensive AI API pricing guide — covering every major provider, every model tier, and practical advice on cutting costs.
These are the most capable models available. They handle complex reasoning, long-form generation, advanced coding, and multimodal tasks.
| Model | Provider | Input (per 1M tokens) | Output (per 1M tokens) | Context Window |
|---|---|---|---|---|
| GPT-5 | OpenAI | $3.00 | $15.00 | 256K |
| Claude Opus 4.6 | Anthropic | $15.00 | $75.00 | 200K |
| Gemini 3 Pro | $2.50 | $10.00 | 2M | |
| Grok 4 | xAI | $3.00 | $15.00 | 256K |
| DeepSeek R2 | DeepSeek | $2.00 | $8.00 | 128K |
Cheapest frontier model: DeepSeek R2 at 2.00/2.00/8.00 — roughly 60% cheaper than GPT-5 on output tokens.
Best value for long context: Gemini 3 Pro with its 2M token context window at just 2.50/2.50/10.00.
Most expensive: Claude Opus 4.6 at 15.00/15.00/75.00 — but many developers consider it worth the premium for complex analytical tasks.
For most production workloads, mid-tier models offer the best balance of quality and cost.
| Model | Provider | Input (per 1M tokens) | Output (per 1M tokens) | Context Window |
|---|---|---|---|---|
| GPT-4.1 | OpenAI | $2.00 | $8.00 | 1M |
| Claude Sonnet 4.5 | Anthropic | $3.00 | $15.00 | 200K |
| Gemini 2.5 Flash | $0.15 | $0.60 | 1M | |
| Grok 4 Mini | xAI | $0.50 | $2.00 | 128K |
| DeepSeek V3.5 | DeepSeek | $0.27 | $1.10 | 64K |
| Moonshot Kimi K2 | Moonshot | $0.60 | $2.40 | 128K |
Standout value: Gemini 2.5 Flash at 0.15/0.15/0.60 is absurdly cheap for its capability level. It handles most coding and writing tasks well.
Best reasoning budget model: Grok 4 Mini and DeepSeek V3.5 both deliver strong results under $2/M output tokens.
When you need high throughput at minimal cost — classification, summarization, simple extraction.
| Model | Provider | Input (per 1M tokens) | Output (per 1M tokens) | Context Window |
|---|---|---|---|---|
| GPT-4.1 Mini | OpenAI | $0.40 | $1.60 | 1M |
| GPT-4.1 Nano | OpenAI | $0.10 | $0.40 | 1M |
| Claude Haiku 3.5 | Anthropic | $0.80 | $4.00 | 200K |
| Gemini 2.0 Flash Lite | $0.075 | $0.30 | 1M | |
| DeepSeek V3.5 Lite | DeepSeek | $0.07 | $0.28 | 32K |
Rock bottom: DeepSeek V3.5 Lite and Gemini 2.0 Flash Lite both come in under 0.10/0.10/0.30 — perfect for high-volume pipelines.
Essential for RAG, search, and similarity applications:
| Model | Provider | Price (per 1M tokens) | Dimensions |
|---|---|---|---|
| text-embedding-3-large | OpenAI | $0.13 | 3072 |
| text-embedding-3-small | OpenAI | $0.02 | 1536 |
| embed-v4 | Cohere | $0.10 | 1024 |
| Gemini Embedding | $0.01 | 768 |
| Model | Provider | Price per Image | Resolution |
|---|---|---|---|
| DALL-E 3 | OpenAI | 0.040−0.040 - 0.120 | Up to 1792x1024 |
| Imagen 3 | 0.020−0.020 - 0.060 | Up to 2048x2048 | |
| Flux 1.1 Pro | BFL | $0.040 | Up to 2048x2048 |
| Ideogram 3.0 | Ideogram | 0.020−0.020 - 0.080 | Up to 2048x2048 |
| Model | Provider | Price | Duration |
|---|---|---|---|
| Sora | OpenAI | $0.20/sec | Up to 20s |
| Veo 3 | $0.15/sec | Up to 8s | |
| Kling 2.1 | Kuaishou | $0.10/sec | Up to 10s |
| Runway Gen 4 Turbo | Runway | $0.25/sec | Up to 10s |
| Seedance 2.0 | ByteDance | $0.08/sec | Up to 10s |
Instead of managing accounts with 6+ providers, use a unified gateway. Crazyrouter offers access to 300+ models through a single OpenAI-compatible API with discounted pricing:
| Model | Official Price | Crazyrouter Price | Savings |
|---|---|---|---|
| GPT-5 | 3.00/3.00 / 15.00 | 2.10/2.10 / 10.50 | 30% |
| Claude Opus 4.6 | 15.00/15.00 / 75.00 | 10.50/10.50 / 52.50 | 30% |
| Grok 4 | 3.00/3.00 / 15.00 | 2.10/2.10 / 10.50 | 30% |
| Gemini 3 Pro | 2.50/2.50 / 10.00 | 1.75/1.75 / 7.00 | 30% |
| DeepSeek R2 | 2.00/2.00 / 8.00 | 1.40/1.40 / 5.60 | 30% |
Don't use GPT-5 for everything. Build a routing layer:
Cache identical or near-identical requests to avoid paying twice:
OpenAI and Anthropic both offer 50% discounts on batch processing:
| Use Case | Recommended Model | Monthly Requests | Est. Monthly Cost |
|---|---|---|---|
| Chatbot (small) | GPT-4.1 Mini | 10,000 | $12 |
| Chatbot (enterprise) | GPT-5 | 100,000 | $2,700 |
| Code assistant | Grok 4 via Crazyrouter | 50,000 | $945 |
| RAG pipeline | Gemini 2.5 Flash + embeddings | 200,000 | $180 |
| Content generation | Claude Sonnet 4.5 | 20,000 | $450 |
| Data extraction | GPT-4.1 Nano | 500,000 | $150 |
For general-purpose tasks, DeepSeek V3.5 Lite (0.07/0.07/0.28 per million tokens) and Google's Gemini 2.0 Flash Lite (0.075/0.075/0.30) are the cheapest options. Through Crazyrouter, you can access both at even lower rates.
Gemini 2.5 Flash offers the best quality-to-price ratio at 0.15/0.15/0.60 per million tokens with a 1M context window. For frontier-level tasks, DeepSeek R2 at 2.00/2.00/8.00 undercuts GPT-5 and Grok 4 significantly.
The most effective strategies are: (1) route tasks to the cheapest capable model, (2) use an aggregator like Crazyrouter for bulk discounts, (3) implement response caching, and (4) use batch APIs for non-real-time workloads.
Using multiple providers through an aggregator like Crazyrouter is typically cheaper. You get volume discounts, can route each task to the cheapest capable model, and avoid vendor lock-in.
Major providers adjust pricing every 1-3 months. New model releases often come with price drops on older models. We update this comparison monthly — bookmark this page for the latest data.
The AI API landscape in May 2026 offers more choice and better pricing than ever. The gap between frontier and budget models continues to narrow, and smart routing between models can cut costs by 60-80% without meaningful quality loss.
For the simplest path to cost optimization, Crazyrouter gives you a single API key for every model listed above, with built-in discounts and the flexibility to switch models with one line of code.
This pricing guide is updated monthly. Last updated: May 2026.
此内容由惯性聚合(RSS阅读器)自动聚合整理，仅供阅读参考。 原文来自 — 版权归原作者所有。