Table of Contents
🌏 中文版
You want to use Claude Sonnet 5 for coding. Hitting the Anthropic API directly costs $2/$10 per MTok. But the same model might be cheaper on OpenRouter, have volume discounts on AWS Bedrock, or offer caching benefits through third parties. This post maps the actual price differences as of August 2026 and helps you pick the right channel.
Six Channels at a Glance
1. Direct Official API
Buy directly from Anthropic, OpenAI, Google, etc. Transparent pricing, lowest latency, best documentation. The downside: separate API keys, SDKs, and billing for each provider.
Best for: using only one or two providers, needing lowest latency, requiring a direct commercial relationship.
2. OpenRouter — Aggregated Routing
OpenRouter unifies dozens of providers behind a single OpenAI-compatible API. The core feature is real-time cheapest-provider routing — when multiple providers offer the same model, OpenRouter picks the cheapest.
Per the OpenRouter pricing page, Claude Sonnet 5 lists at $2/$10 as of August 2026 (same as direct), but some providers run promotions. OpenRouter's business model adds a thin margin on top of provider prices.
Best for: accessing multiple models, one API key for everything, rapid model evaluation.
3. AWS Bedrock — Enterprise Grade
Bedrock offers Claude, Llama, Mistral, and more. Per modelgrep, Bedrock hosts about 30 models as of August 2026, including Claude Opus 5 and Fable 5.
Bedrock's list pricing matches direct (Claude Sonnet 5 at $2/$10), but offers Provisioned Throughput (guaranteed capacity with monthly commitment discounts) and Cross-Region Inference Profiles (load balancing across regions). The hidden value for enterprises is unified AWS billing, existing compliance frameworks, and no separate vendor accounts.
Best for: already on AWS, need SLAs and compliance, high steady-state volume.
4. Google Vertex AI — The GCP Alternative
Vertex AI is Google's equivalent to Bedrock. Offers native Gemini models plus third-party models (Claude, Llama, etc.). Unified billing through Google Cloud with commitment discounts.
Best for: already on GCP, primarily using Gemini models.
5. Together AI — Open-Source Model Specialists
Together AI focuses on managed inference for open-source models. Per modelgrep, Together hosts 21 models including Kimi K3, Qwen3.8, and DeepSeek V4 Pro.
Open-source models on Together typically cost an order of magnitude less than direct closed-source models. For example, DeepSeek V4 Flash on Together runs about $0.20/$0.60 per MTok — one-tenth of Claude Sonnet 5.
Best for: heavy open-source model usage, managed fine-tuning, cost-sensitive workloads.
6. Fireworks AI — Speed-Optimized
Fireworks AI focuses on inference speed, using techniques like speculative decoding to minimize latency. Per modelgrep, they host 9 models.
Best for: latency-sensitive agentic applications, high-throughput requirements.
Price Comparison Table (August 2026)
Prices sourced from each platform's public pricing pages and CloudZero's compilation (checked 2026-08-20). Unit: $/MTok (million tokens).
| Model | Direct (input/output) | OpenRouter | Bedrock | Together AI |
|---|---|---|---|---|
| Claude Sonnet 5 | $2 / $10 | $2 / $10 | $2 / $10 | — |
| Claude Opus 5 | $5 / $25 | $5 / $25 | $5 / $25 | — |
| GPT-5.6 Sol | $2.50 / $15 | $2.50 / $15 | Listed | — |
| GPT-5.6 Luna | $0.20 / $1.20 | $0.20 / $1.20 | — | — |
| DeepSeek V4 Pro | ~$0.50 / $2.00 direct | Multi-provider bidding | Listed | ✅ Listed |
| Llama 4 Maverick | Free weights | $0.20 / $0.50 | Listed | ✅ Listed |
| Kimi K3 | $3 / $15 | Listed | — | ✅ Listed |
Key observations:
- Closed-source models (Claude, GPT) cost nearly the same across all channels — providers control pricing and aggregators have little room to negotiate
- Open-source models (DeepSeek, Llama, Kimi) are where the real price competition happens — different hosting providers have different inference efficiency, with 2-3× price spreads
- Bedrock and Vertex's hidden value lies in commitment discounts and unified billing — list prices match direct, but volume deals are negotiable
How to Choose: Decision Tree
How many model providers do you need?
├── Just one → Direct official API (simplest, lowest latency)
└── Multiple
├── Are you on AWS/GCP?
│ ├── AWS → Bedrock (unified billing, compliance)
│ └── GCP → Vertex AI
└── Not on cloud / indie developer
├── Mostly closed-source models → OpenRouter (one key for everything)
└── Mostly open-source models
├── Latency-sensitive → Fireworks AI
└── Cost-sensitive → Together AIHidden Costs to Watch
- Prompt caching: Anthropic's prompt caching saves 90% on input costs (OpenAI offers 50%). For applications with repeated system prompts, direct Anthropic API might be cheaper than an aggregator
- Batch API: Both OpenAI and Anthropic offer non-realtime batch APIs at half price. Not all aggregators support this
- Rate limits: Direct APIs generally have higher rate limits. Aggregator limits are shared
- Gateway tools: LiteLLM and Portkey are not providers — they let you build your own routing gateway. You bring your own API keys but get a unified interface, automatic fallbacks, and observability
Bottom Line
The 2026 LLM API market is highly commoditized — there's little to compare on closed-source model pricing across channels. The real price differences are in open-source model hosting. If you're an indie developer or small team, OpenRouter's one-stop experience saves the most hassle. If you're an enterprise, Bedrock/Vertex's unified billing and compliance frameworks are the actual selling point, not the price.
References
- LLM API Pricing Comparison — CloudZero (2026-08-20)
- LLM API Providers Compared — modelgrep (August 2026, live-updated)
- LLM API Pricing Comparison — CostGoat
- OpenRouter Deep Dive — this site (in Chinese)
- AWS Bedrock Deep Dive — this site (in Chinese)
- Google Vertex AI Deep Dive — this site (in Chinese)
- Together AI Deep Dive — this site (in Chinese)
- Fireworks AI Deep Dive — this site (in Chinese)
- LiteLLM Gateway — this site (in Chinese)
- Portkey AI Gateway — this site (in Chinese)
Loading...