Engineering Whitepaper • Updated 2026
Cheapest LLM API Providers in 2026
Comprehensive analysis of lowest cost per million tokens across DeepSeek, Google, Groq, and OpenRouter.
Key Architectural Takeaways
- Prompt Caching Invalidation: Anthropic caches expire in 5 minutes unless refreshed by subsequent hits. For low-traffic apps (<1 req/5min), write surcharges can increase total bills.
- The Ground-Floor Pricing Era: DeepSeek-V3 ($0.27/M) and Gemini 1.5 Flash ($0.075/M) have made commodity classification and extraction nearly free.
- Hybrid Cascading: Never route 100% of traffic to a single flagship model. 60%-80% of agent tasks can be handled by fast tiers with zero loss in end-to-end task success.
Provider Cost Index (Per 1M Tokens)
| Model | Provider | Input / 1M | Output / 1M | Cache Read / 1M |
|---|---|---|---|---|
| Apodex: Apodex 1.1 Mini (free) | Apodex | $0.000 | $0.000 | $0.000 |
| Baidu: ERNIE-5.5-Flash-Preview (free) | Baidu | $0.000 | $0.000 | $0.000 |
| Cohere: North Mini Code (free) | Cohere | $0.000 | $0.000 | $0.000 |
| Dots Studio: Dots3-Note Preview (free) | Dots-studio | $0.000 | $0.000 | $0.000 |
| Google: Gemma 4 26B A4B (free) | $0.000 | $0.000 | $0.000 | |
| Google: Gemma 4 31B (free) | $0.000 | $0.000 | $0.000 |
Want to simulate your exact workload?
Use the TokenCosts AI interactive calculator with custom volume, cache %, and batch rates.