LLM API Cost Comparison: OpenAI vs Anthropic vs Google (March 2026)
Choosing an LLM API? Cost can make or break your budget. A naïve implementation can burn $10k/month where a smart one costs $500.
This guide breaks down real pricing (March 2026), shows cost-per-task examples, and reveals hidden tricks to slash your bill.
Quick Comparison Table
Legend:
- ⚡ = Slow (10–30 tokens/sec)
- ⚡⚡⚡ = Fast (40–70 tokens/sec)
- ⚡⚡⚡⚡⚡ = Very fast (100+ tokens/sec)
Real-World Cost Examples
Example 1: Customer Support Chatbot
Usage: 100k messages/month, 500 tokens input + 200 tokens output each
Why Gemini wins: 10x cheaper than competitors, 1M context handles long conversations.
Example 2: Code Generation Tool
Usage: 50k requests/month, 2k tokens input + 1k tokens output each
Why Copilot wins: Subsidized pricing (GitHub eats the cost). Only available to Copilot subscribers ($10–20/month).
Example 3: Document Analysis (Long Context)
Usage: 10k docs/month, 50k tokens input + 2k tokens output each
Why Gemini wins: 1M context window = fewer API calls, lower input costs.
Example 4: Summarization Pipeline
Usage: 1M short texts/month, 200 tokens input + 50 tokens output each
Why Gemini wins: Unbeatable pricing for simple tasks.
Hidden Costs to Watch
1. Prompt Caching (Anthropic Only)
What it is: Reuse repeated prompt prefixes, pay 10% of normal input cost.
Example:
- Normal: 100k tokens input = $1.50 (Claude Opus)
- With caching: 10k unique + 90k cached = $0.15 + $0.135 = $0.285 (81% savings)
When it helps: Long system prompts, RAG contexts, repeated instructions.
How to use:
# Anthropic API
response = anthropic.messages.create(
model="claude-opus-4.6",
messages=[...],
system=[
{"type": "text", "text": "Long system prompt...", "cache_control": {"type": "ephemeral"}}
]
)Savings: Up to 90% on input costs.
2. Batch API (OpenAI)
What it is: Submit jobs in bulk, get 50% discount, results in 24h.
When it helps: Non-time-sensitive tasks (data labeling, summarization).
Example:
- Standard API: $15/1M input = $1,500 for 100M tokens
- Batch API: $7.50/1M input = $750 (50% savings)
How to use:
# OpenAI Batch API
client.batches.create(
input_file=batch_file,
endpoint="/v1/chat/completions",
completion_window="24h"
)3. Output Token Costs (Often Overlooked)
Reality check: Output tokens cost 2–5x more than input tokens.
Bad example:
- Generate 10k token response = $0.60 (GPT-5 output)
- Could have used GPT-5 Mini = $0.006 (100x cheaper)
Optimization: Use smaller models for long outputs (summaries, reports).
Cost Optimization Strategies
Strategy 1: Tiered Model Routing
Route requests based on complexity:
Simple tasks → Gemini 2.5 Flash ($0.075 input)
Medium tasks → Claude Haiku / GPT-5 Mini ($0.25 input)
Hard tasks → GPT-5 / Claude Opus ($15 input)Tools: LiteLLM, OpenRouter, custom routing logic.
Savings: 60–80% on total API costs.
Strategy 2: Prompt Compression
Compress prompts without losing context:
Tools:
- PromptCompressor — 50–80% token reduction
- Semantic caching (vector DB + similarity search)
Example:
- Original: 5k tokens = $0.075 (GPT-5.4)
- Compressed: 1.5k tokens = $0.0225 (70% savings)
Strategy 3: Local + Cloud Hybrid
Run cheap tasks locally (Ollama), expensive tasks in cloud:
Draft generation → Ollama Mistral 7B (free)
Final polish → Claude Sonnet 4.5 ($3 input)Savings: 80–90% vs pure cloud.
Strategy 4: GitHub Copilot Arbitrage
If you have Copilot subscription ($10–20/month):
Use Copilot API for everything:
- Claude Sonnet 4.5: $0.50 input (vs $3 direct)
- Claude Opus 4.6: $0.50 input (vs $15 direct)
Catch: 10 req/min rate limit. Fine for low-volume personal projects.
Hidden Pricing Traps
❌ Free Tiers Are Marketing
- OpenAI: $5 free credits expire in 3 months
- Anthropic: No free tier
- Google: $300 credits (90 days) then charges
Trap: Free credits lure you in, then bills hit. Budget from day 1.
❌ Rate Limits Can Cost You
Hitting rate limits = retries = wasted tokens + latency.
Tiers (OpenAI example):
- Tier 1 (new account): 500 req/min
- Tier 5 ($1k+ spent): 10k req/min
Solution: Use multiple API keys, rotate providers, or pay for higher tier.
❌ Context Window Waste
Bad example: Send 50k token context, only need 5k.
Cost:
- Wasted: 45k tokens × $15/1M = $0.675 per request
- Over 100k requests = $67,500 wasted
Solution: Trim context, use RAG (only send relevant chunks).
Which Provider Should You Choose?
Choose OpenAI if:
- You need GPT-5 class performance
- Speed matters (fastest inference)
- Ecosystem matters (most integrations)
Choose Anthropic if:
- Long context (200k+ tokens)
- Safety/refusal behavior matters (most aligned)
- Prompt caching saves you money
Choose Google if:
- Cost is priority #1 (cheapest flagship + flash models)
- 1M context window (process books, codebases)
- Multimodal native (video, audio)
Choose GitHub Copilot if:
- You're already a Copilot subscriber
- Low-volume personal/side projects
- Want flagship models at 90% discount
Cost Calculator
Try this formula:
Monthly cost = (input_tokens × input_price) + (output_tokens × output_price)Example:
- 100M input, 20M output
- GPT-5.4: (100 × $15) + (20 × $60) = $2,700
- Gemini 3.1 Pro: (100 × $7) + (20 × $21) = $1,120
Savings: $1,580/month (58%)
Final Recommendations
For most apps:
- Start with Gemini 2.5 Flash (cheapest, fast)
- Upgrade to Gemini 3.1 Pro if quality suffers
- Add Claude Sonnet 4.5 for edge cases
For high-stakes apps:
- Use Claude Opus 4.6 or GPT-5.4
- Implement prompt caching (Anthropic)
- Route easy tasks to cheaper models
For personal projects:
- Get GitHub Copilot ($10–20/month)
- Use Copilot API for everything
- Fallback to Ollama for free local inference
Resources
What's your monthly API bill? Drop it in the comments — let's compare strategies.
(Affiliate disclosure: Some links may include referral codes. I only recommend tools I actually use.)
