Solunar AI
Solunar AI / Blog / One endpoint, three families

One endpoint, three model families: how we run Claude, GPT and GLM behind a single gateway

Insights · July 2, 2026 · 6 min read

We just shipped our third model family. The same gateway that routes Claude and GPT now also serves GLM-5.2 — one base URL, one balance, one API key. This post is about how we did it without breaking the two promises this gateway is built on: no silent downgrades, and honest data boundaries.

First, the term

A value tier is a lower-priced model tier behind the same gateway, clearly listed as its own thing: built for high-volume, cost-sensitive work; it shares your key and balance with the frontier tiers, but its data boundary is declared separately and honestly — never blended into the frontier tiers' promises.

One endpoint, per-model routing

If you already use an OpenAI-compatible or Anthropic-compatible SDK, integration is a two-line change: point base_url at the gateway, swap the key. Routing happens per request — claude-* models take the Anthropic-native path (with automatic prompt-cache injection, more below), while gpt-* and glm-* take the OpenAI-compatible path. Your code never needs to know how many families sit behind it.

What we deliberately did not build: cross-vendor fallback

Ask for Claude when Claude is down, and you get an error — not a quiet substitution by a different model. A gateway that silently swaps vendors under load is a gateway you can't trust for evals, agent pipelines, or anything where model identity matters. Fallbacks stay within a family, and only where the substitution is strictly upward.

The part that actually saves money

For Claude models, the gateway injects prompt-cache markers into your system prompts automatically — zero client changes. If your workload re-sends stable context (system prompts, tool definitions, long documents — i.e., basically every agent and every coding assistant), repeated input bills at the cached rate instead of full price. You don't have to take our word for it: the usage block of every response carries cache_read_input_tokens, so you can verify it yourself. We've written before about verifying gateway claims; the short version is: don't trust promises — read your own usage fields.

The value tier: GLM-5.2

GLM-5.2 is the new addition: an open-weights 744B MoE model with strong long-horizon coding results and a large context window. Here it's an explicitly-listed value tier — per-token prices are on the pricing page after sign-in, and the numbers speak for themselves.

Where it fits: high-volume, cost-sensitive work — batch processing, drafts, classification, first-pass coding — where you'd rather spend a fraction of frontier price and keep the frontier models for the calls that need them. Because one balance covers both tiers, a cheap-first-pass, frontier-second-pass pipeline is a routing decision, not a procurement project.

Different tiers, different data boundaries — in writing

Here's the part most gateways gloss over: not every model family comes with the same data guarantees, and pretending otherwise is how you lose trust.

For Claude and GPT requests, we use official channels, and under those providers' commercial terms your content is not used for training. For the GLM tier, the supply chain is different, and the guarantees we can make are narrower — so our privacy policy says exactly that, including which jurisdictions may apply and what we cannot contractually promise. GLM is used only when you explicitly select a GLM model, and we tell you plainly: don't send it content you wouldn't want processed under those conditions.

We'd rather publish a narrower promise we can keep than a broad one we can't. That's the whole philosophy of this gateway in one sentence. Per-tier details live in the Privacy Policy and the data-boundary section of /docs.

Solunar Gateway

Sign up at console.solunarai.com, change two lines, and you have Claude, GPT and GLM behind one key. Docs at solunarai.com/docs — including the data-boundary section, which is worth reading before you pick a tier.

Request access → Read the docs

FAQ

Does choosing GLM change where my data goes?
Yes — and we put that in writing. The GLM tier runs on a different supply chain than our official Claude/GPT channels, and the guarantees we can make there are narrower; the privacy policy states per tier exactly what we can and cannot promise. GLM is used only when you explicitly select a GLM model — if you never pick it, your requests never traverse that chain.
Will the gateway ever swap the model I asked for?
No — that's a design position. Cross-vendor fallback is off: ask for Claude and you get Claude, or an error — never a cheaper model quietly standing in. For evals and agent pipelines, model identity has to be a fact, not a suggestion.
When should I use the value tier vs. a frontier model?
Send high-volume, cost-sensitive work — batch jobs, drafts, classification, first-pass coding — to the value tier, and keep frontier models for calls that need top reasoning quality. With one balance and one key, cheap-first-pass, frontier-second-pass is just a routing decision.
Do I need a separate account or balance for GLM?
No. The same key and balance cover all three families; just set the model name to glm-5.2 in your request. Per-token prices are on the pricing page after sign-in — the numbers speak for themselves.

Keep exploring

Where your data goes Verifying the real model Token cost governance Back to Blog

Drafted with AI assistance; facts and positions reviewed by the founder.