How Synthetic works
Synthetic is a subscription service that runs open-source coding LLMs on dedicated infrastructure and serves them over an OpenAI-compatible (and Anthropic-compatible) API. Instead of paying per-token, you pay a flat monthly subscription and get a generous rate-limit budget covering every model on the platform.
It’s positioned as a direct alternative to Claude’s $20/month and $200/month plans for developers who do a lot of coding: the same price point, several-times-higher rate limits, and your choice of open-source models.
Mesrai reaches Synthetic through the OpenAI-compatible endpoint, so there are no adapter changes — just BYOK credentials.
Plans at a glance
Pricing and rate limits change. Always confirm at synthetic.new/pricing before choosing a tier.
Subscription (recommended for code review)
| Plan | Price | Rate limit |
|---|---|---|
| Monthly | ~$30/mo ($1/day) | ~500 messages per 5-hour window |
| Enterprise / usage-based | contact Synthetic | pay-per-token option |
- All standard models are covered by the subscription — pick any of them by changing the
modelfield. - Rate limits are enforced on a 5-hour rolling window, not per month — plan concurrency to fit inside that window.
- Synthetic does not train on your prompts/completions and deletes API data within 14 days.
Recommended models
Synthetic exposes models using a HuggingFace-style prefix (hf:org/model). A few strong picks for code review:
| Model id | Notes |
|---|---|
hf:zai-org/GLM-4.6 | GLM family; balanced for general-purpose review. |
hf:moonshotai/Kimi-K2-Instruct | Long-context Kimi K2 — great on large PRs. |
hf:Qwen/Qwen3-Coder-480B-A35B-Instruct | Strong specialized coder. |
hf:deepseek-ai/DeepSeek-V3.1 | DeepSeek V3.1; strong agentic/coding behavior. |
See the full catalog at dev.synthetic.new/docs/openai/models.
Creating an API Key
- Go to synthetic.new and sign up or log in.
- Subscribe to a plan at synthetic.new/pricing.
- Open the developer console and create an API key. Copy it immediately — you will not see the full key again.
Configure Synthetic in Mesrai
The primary flow is BYOK on Mesrai Cloud. Self-hosted installs can use environment variables as a secondary option.
Option 1 — BYOK on Mesrai Cloud (recommended)
-
In the Mesrai web UI, open Settings → BYOK (app.mesrai.com/organization/byok).
-
Synthetic isn’t in the curated catalog — click Configure manually at the bottom of the model list. Use
?slot=fallbackin the URL if configuring a fallback instead of the main model. -
Fill the wizard:
Field Value Provider OpenAI CompatibleBase URL https://api.synthetic.new/openai/v1Model e.g. hf:zai-org/GLM-4.6(use thehf:prefix and the full HuggingFace path)API Key your Synthetic API key Max Concurrent Requests start at 3–5to fit the 5-hour budget on bigger PRs (under Advanced settings) -
Click Test & save. Mesrai probes the endpoint and persists the config on success. 401 means the key is wrong; 404 usually means the base URL or model path is wrong.
The 500-messages-per-5-hours budget is per subscription, not per PR. For teams doing heavy continuous review, keep Max Concurrent Requests low enough that one big multi-file PR doesn’t eat the window, or split the work across multiple subscriptions.
Configure Synthetic as Main and keep an OpenAI or Anthropic key as Fallback — when the 5-hour window fills up, Mesrai automatically fails over and reviews keep flowing.
Enterprise deployments
For Enterprise customers running Mesrai in their own VPC, the Synthetic block is configured via environment variables in the API + worker containers. Your onboarding contact will share the exact .env template and deployment runbook for your tenancy. Contact sales for details.
When to pick Synthetic
- Your primary model is open-source (Kimi, DeepSeek, Qwen Coder, GLM) and you want a single flat-rate bill instead of juggling per-provider token spend.
- You’re comparing to Claude Max ($200/mo) for review throughput — Synthetic at $30/mo delivers several times the rate-limit budget for similarly-capable coding models.
- You want strong data posture — Synthetic is explicit about no-training and 14-day prompt/completion retention.
Pick pay-per-token (Moonshot, Fireworks, Together, Groq) instead if traffic is bursty, or a different subscription (Z.AI GLM Coding Plan) if you specifically want the GLM family with tiered commits.
Troubleshooting
429 Too Many Requests inside the 5-hour window
- The subscription budget is per 5-hour rolling window, not per hour or per month. Check how many requests you’ve sent recently in the Synthetic dashboard.
- Either wait for the window to roll forward, lower
Max Concurrent Requestsin BYOK, or upgrade/stack plans. - Configure a
FallbackBYOK provider so reviews keep running while you’re rate-limited.
401 / authentication errors
- Confirm the key is active and the subscription is current in the Synthetic dashboard.
- Make sure there are no trailing spaces or quotes in the
.envvalue.
Model not found
- Synthetic uses the
hf:org/modelformat. Leaving off thehf:prefix or using a different capitalization will 404. - Cross-check against dev.synthetic.new/docs/openai/models.
Long cold-start on first call
- Some larger models may cold-start the first time they’re used after a period of inactivity.
- If first-call latency is blocking Mesrai health checks, warm the model by issuing a tiny test request before opening PRs.
Connection errors
- Confirm your server can reach
api.synthetic.new. - Review API and worker logs for the exact upstream error.