Synthetic is an open-source LLM hosting service operated by Synthetic Lab, Co. in the United States. It runs publicly available models such as GLM, Kimi, Qwen, and GPT-OSS in its own managed datacenters and exposes them through OpenAI-compatible and Anthropic-compatible APIs. Its two defining traits are an explicit stance that user data is never used for training and that API prompts and completions are never stored, and a flat $30/month price instead of per-token billing. Because existing coding agents such as Claude Code, OpenCode, and Cline only need their endpoint swapped, it has become a landing spot for developers who cannot predict their usage-based bills from commercial models.
Key Features
- Open-source LLMs hosted in one place: Kimi K3 (512k context), GLM 5.2, GLM 4.7 Flash, Qwen3.8 27B, GPT-OSS-120B, and NVIDIA Nemotron-3 Super — the major vLLM-compatible open models — are all reachable from a single API. Nomic Embed Text v1.5 is available for embeddings
- Both OpenAI- and Anthropic-compatible endpoints:
https://api.synthetic.new/openai/v1accepts OpenAI-style chat/completions and embeddings, whilehttps://api.synthetic.new/anthropic/v1accepts Anthropic-style messages. Existing code usually migrates by changing only the base URL and the API key - Model aliases: Instead of pinning a specific model name, you can specify role-based aliases such as
syn:large:textandsyn:small:text, so your application keeps working as models are replaced by newer generations - Designed not to train or store: The company states that it never trains on user data and never stores API prompts or completions, and points to a GDPR-compliant privacy policy. That is a concrete input when you handle code that cannot leave your organization
- Flat-rate subscription: Rather than accumulating per-token charges, you buy a monthly pack. Each pack carries a request ceiling, and buying additional packs raises the limit and the number of concurrent requests
- Drops into existing agent tools: It works with anything that lets you configure an OpenAI-compatible endpoint, including Claude Code, OpenCode, GitHub Copilot, Cline, Roo, Octofriend, OpenClaw, and Xcode Intelligence
- Synbad evaluation suite published openly: The company open-sources an evaluation suite built from real bugs encountered by coding agents, making its model-selection evidence available to outsiders
Pricing
| Plan | Price | What’s included |
|---|---|---|
| Subscription pack | $30/month (or $1/day) | 500 requests per 5 hours, 1 concurrent request per model, access to both the UI and the API. Buy additional packs to raise the limits and concurrency |
| Usage-based | Contact required | Pay only for what you use, with no upfront cost. Aimed at larger and enterprise customers |
Embedding requests do not count against the subscription’s rate limits.
Pricing is current as of August 2026. Check the official site for the latest information.
Pros & Cons
✅ Pros
- A fixed monthly price means costs stay predictable even when you run coding agents for hours
- The stated policy of not training on user data and not storing prompts makes it easier to choose when your inputs are sensitive
- Support for both OpenAI- and Anthropic-compatible APIs means almost no changes to existing tools and code
- Model aliases keep your application from being disrupted every time models turn over
- One contract and one key let you compare several open-source models side by side
⚠️ Cons
- Only open-source models are offered — commercial flagships such as GPT, Claude, and Gemini are not available
- The default of one concurrent request per model means workloads that need parallelism require buying extra packs
- The per-5-hour request ceiling makes it a poor fit for large batch processing
- Usage-based unit prices are not published on the pricing page and require an inquiry
- There is no Japanese documentation or Japan-based support
Comparison with Similar Services
| Criteria | Synthetic | OpenRouter | Together AI | Groq |
|---|---|---|---|---|
| Models offered | Mainly open-source LLMs | Both commercial and open-source | Mainly open-source | Mainly open-source |
| Billing | Flat monthly rate first (usage-based also available) | Usage-based | Usage-based | Usage-based |
| API compatibility | OpenAI + Anthropic | OpenAI-compatible | OpenAI-compatible | OpenAI-compatible |
| Strength | Predictable cost, explicit no-storage stance | Breadth of model choice | Model count and fine-tuning | Inference speed |
| Best suited for | Everyday agent use | Comparing and switching models | Both training and inference | Low-latency workloads |
Who Is It For
- Developers who use coding agents such as Claude Code or Cline every day and struggle with unpredictable usage-based bills
- Anyone who wants code or text processed on the premise that it will not be trained on or stored
- People who want to try GLM, Kimi, and Qwen side by side without setting up their own infrastructure
- Teams looking for a low-friction alternative that only requires swapping an OpenAI-compatible endpoint
Summary
Synthetic is an API service that commits fully to delivering open-source LLMs at a flat rate, without training on or storing your data. It is not the right choice when you need a commercial flagship model, but for running coding agents day in and day out, the fixed monthly cost structure is itself the value. Start with the $30/month pack, point your existing agent at it, and check whether the rate ceiling fits the way you actually work.