LLM API Pricing Comparison: Cost-Per-Million-Tokens Across All Providers

Contents

• Understanding LLM API Pricing (March 2026)
• FAQ
• Related Resources
• Sources

Understanding LLM API Pricing (March 2026)

Language model API pricing has fractured into tiers reflecting model capabilities and inference costs. The same provider charges vastly different rates for different models. Understanding actual cost-per-token is critical for budget forecasting.

Provider Pricing market

OpenAI GPT Models:

GPT-4o costs $2.50 per 1M input tokens and $10 per 1M output tokens. GPT-4.1 costs $2.00 / $8.00 per 1M tokens. GPT-5 (as of March 2026) costs $1.25 / $10.00. GPT-4o-mini costs $0.15 / $0.60 per 1M tokens.

The pricing structure penalizes output tokens at 3x input cost. This ratio reflects higher compute requirements for generation versus understanding.

OpenAI also introduced cache features allowing 90% discounts on cached input tokens (first request at full price, subsequent requests using cached context at 10% cost). For applications with repeated context patterns, caching substantially reduces costs.

Anthropic Claude Models:

Claude Opus 4.6 (most capable) costs $5 per 1M input tokens and $25 per 1M output tokens. Claude Sonnet 4.6 costs $3 / $15. Claude Haiku 4.5 (fastest/cheapest) costs $1.00 / $5.00.

Anthropic's pricing emphasizes long context windows. Claude Opus 4.6 and Sonnet 4.6 both have 1M token context windows; Haiku 4.5 has 200K. Extended context doesn't increase per-token cost, unlike some competitors who charge for full window availability.

Google Gemini Models:

Gemini 2.5 Flash costs $0.30 per 1M input tokens and $2.50 per 1M output tokens. Gemini 2.5 Pro costs $1.25 / $10.00 (≤200K context). Both models have 1M token context windows.

Google's Flash model is competitive on pricing: roughly 17x cheaper than Claude Opus for input tokens.

Meta Llama (via Groq, Together, etc.):

Llama 3.1 405B via Together AI costs $8.00 per 1M input tokens and $24 per 1M output tokens. Via Groq (optimized inference), Llama 3 70B costs $0.34 per 1M input and $1.02 per 1M output (extremely cheap).

Open-weight models are dramatically cheaper when optimized inference platforms are used.

Mistral and Other Open Models:

Mistral Large via Mistral AI API costs $2.00 per 1M input tokens and $6.00 per 1M output tokens. Same model via Together AI costs lower rates with different inference optimization.

Pricing Comparison Table

Input vs. Output Token Economics

LLM API pricing fundamentally reflects different costs for input (understanding context) and output (generation) tokens. Understanding this asymmetry is critical for cost optimization.

Input Token Cost Analysis:

Input tokens require fewer compute operations than output tokens. Batch processing multiple requests reduces per-token input cost through parallelization. Most costs reflect prompt context rather than inference compute.

For applications like question-answering over documents, input tokens dominate costs. A 10,000-token document + 100-token question uses 10,100 input tokens. Output might be just 200 tokens. Total cost is heavily input-weighted.

Output Token Cost Analysis:

Generating output is computationally intensive. Each output token requires computing a full transformer pass, sampling, and validation. This explains why output token pricing is 3-15x input costs.

For applications generating large responses (long-form content, code generation), output tokens dominate. A request for 2,000-word article uses minimal input tokens but 2,000+ output tokens.

Optimal Request Design:

Minimize input tokens: pre-process and summarize context before sending to API. Use system prompts efficiently:concise instructions are better than verbose examples.

Control output: set max_tokens parameter to expected output length. Avoid unnecessarily long responses.

For text understanding tasks, focus on input efficiency. For generation tasks, focus on output efficiency.

Caching and Optimization Strategies

OpenAI Prompt Caching:

Feature: First request with a given prompt pays full price. Subsequent requests using cached segments pay 10% for input tokens matching the cache.

Economics: Request 1 = $5 / 1M tokens. Request 2-10 (same context) = $0.50 / 1M tokens. After 10 requests, cache pays for itself.

Best for: Applications reusing fixed context (RAG systems with repeated documents, customer support with shared knowledge base, code analysis with repeated codebase segments).

Anthropic Batch Processing:

Anthropic offers batch API with 50% discount for non-time-critical workloads. Ideal for bulk processing.

Economics: Real-time API costs $15/1M input (Opus). Batch API costs $7.50/1M. For 100M tokens monthly, batch saves $750.

Best for: Bulk document processing, classification, summarization where latency tolerance exists.

Model Selection Optimization:

Choosing cheaper models when quality permits yields substantial savings:

• Haiku (Anthropic) or Gemini Flash instead of Opus for simple tasks: 15-75x cheaper
• GPT-3.5-turbo instead of GPT-4o for straightforward tasks: 10x cheaper
• Llama 3 70B via Groq instead of proprietary models: 20-50x cheaper (if quality sufficient)

Quality-adjusted cost analysis: Test task completion with cheaper models first. Use expensive models only if cheaper alternatives fail.

Real-World Cost Scenarios

Scenario 1: Customer Support Chatbot (1,000 daily conversations)

Assumptions:

• 2,000 token average context (customer history + knowledge base)
• 200 token average response
• Using GPT-3.5-turbo

Daily cost:

• Input: 1,000 × 2,000 tokens × $0.50/1M = $1.00
• Output: 1,000 × 200 tokens × $1.50/1M = $0.30
• Total daily: $1.30
• Monthly: $39

Same with Claude Haiku 4.5:

• Input: 1,000 × 2,000 × $1.00/1M = $2.00
• Output: 1,000 × 200 × $5.00/1M = $1.00
• Total daily: $3.00
• Monthly: $90

Same with Gemini 2.5 Flash:

• Input: 1,000 × 2,000 × $0.30/1M = $0.60
• Output: 1,000 × 200 × $2.50/1M = $0.50
• Total daily: $1.10
• Monthly: $33 (15% savings vs. GPT-3.5-turbo at ~$39)

Scenario 2: Batch Document Summarization (100,000 documents)

Assumptions:

• 3,000 token average document
• 500 token summary
• Can use batch API (non-real-time)

Using Claude Opus 4.6 (real-time):

• Input: 100K × 3,000 × $5/1M = $1,500
• Output: 100K × 500 × $25/1M = $1,250
• Total: $2,750

Using Claude Opus 4.6 (batch at 50% input discount):

• Total: ~$2,000

Using Llama 3 70B via Together (batch equivalent):

• Input: 100K × 3,000 × $0.40/1M = $120
• Output: 100K × 500 × $1.20/1M = $60
• Total: $180 (23x cheaper than Opus batch)

Scenario 3: Code Generation for Development (50 prompts daily)

Assumptions:

• 500 token average prompt (code context + instructions)
• 800 token average generated code
• Needs high quality (GPT-4o)

Daily cost:

• Input: 50 × 500 × $2.50/1M = $0.0625
• Output: 50 × 800 × $10/1M = $0.40
• Daily: $0.4625
• Monthly: $13.88

With GPT-4o-mini (acceptable for routine tasks):

• Input: 50 × 500 × $0.15/1M = $0.00375
• Output: 50 × 800 × $0.60/1M = $0.024
• Daily: $0.02775
• Monthly: $0.83 (94% savings vs GPT-4o)

With Mistral Large:

• Input: 50 × 500 × $2.00/1M = $0.05
• Output: 50 × 800 × $6.00/1M = $0.24
• Daily: $0.29
• Monthly: $8.70

Quality vs. Cost Tradeoffs

Not all models are equal. Quality-adjusted pricing requires testing:

Simple Classification Tasks:

GPT-3.5-turbo and Claude Haiku both achieve 95%+ accuracy at 20-60x lower cost than premium models. Quality-adjusted savings are massive.

Complex Reasoning:

Claude Opus and GPT-4o maintain significant quality advantages for multi-step reasoning. Premium models justify costs if accuracy above 85-90% is required.

Code Generation:

GPT-4o leads for complex problems. For routine patterns, GPT-3.5 is competitive. Haiku sometimes struggles with intricate requirements.

Creative Writing:

Claude excels due to long context handling and instruction-following. GPT-4o similar quality. Cheaper models produce acceptable but lower-quality output.

Recommendation: Test each model with the representative workload. If all achieve 90%+ accuracy, use cheapest option.

Multi-Provider Strategy

Many teams balance cost and capability with multi-provider approaches:

• Premium tasks (reasoning, analysis): Claude Opus or GPT-4o
• Standard tasks (classification, summarization): GPT-3.5 or Claude Sonnet
• Cost-sensitive tasks: Gemini Flash or Llama via Groq
• Bulk processing: Batch APIs (50% discount)

Route requests based on complexity. Use cheap models first; escalate to expensive models only when cheaper ones fail.

Relevant Pricing Context

For broader context on AI infrastructure costs, check GPU pricing to understand self-hosted inference costs. For embedding APIs, review embedding model pricing.

For specific provider details, see OpenAI API pricing and Anthropic API pricing.

Compare against self-hosted inference with GPU cloud pricing tracker methodology.

FAQ

What's the actual cheapest way to run LLMs?

Self-hosting open models on GPU infrastructure (Groq, Together, RunPod) costs 50-500x less per token than proprietary APIs. However, it requires engineering overhead. For small-scale use (under $500/month), APIs are simpler and sometimes cheaper. For large-scale (over $5K/month), self-hosting is optimal.

Should I commit to spending with providers for discounts?

Only if your volume is verified and stable. OpenAI/Anthropic discounts require $100K+ commitments typically. Ensure you'll actually use committed volume before signing.

How much does caching actually save?

For applications with repeated context, 50-90% savings on input tokens (caching reduces to 10% cost). If you send same 10K-token context 100 times monthly, caching saves $45. Only beneficial for high-context, repeated-request patterns.

Can I mix providers based on cost?

Yes, it's an advanced optimization. Route simple tasks to cheap providers, complex tasks to expensive ones. Requires infrastructure coordination but yields significant savings at scale.

What about privacy and using proprietary models?

All proprietary APIs (OpenAI, Anthropic, Google) log requests for monitoring. For sensitive data, self-host open models or use dedicated private endpoints (more expensive). Understand privacy requirements before choosing providers.

How do I forecast my LLM API costs?

Estimate: (average requests per month) × (average input tokens per request) × (input rate) + (average output tokens per request) × (output rate). Test this with pilot queries for 1-2 weeks to calibrate estimates.

Related Resources

• OpenAI API Pricing - OpenAI detailed pricing
• Anthropic API Pricing - Anthropic Claude pricing
• Embedding Model Pricing - Semantic search APIs
• GPU Pricing Guide - Self-hosted inference infrastructure
• GPU Cloud Price Tracker - Monitoring infrastructure costs

Sources

• OpenAI, Anthropic, Google official API pricing (as of March 2026)
• Groq, Together AI, Mistral pricing documentation (as of March 2026)
• DeployBase.AI LLM cost analysis (as of March 2026)
• Community benchmarking on model quality across price tiers
• Case studies on cost optimization strategies from 2025-2026