Every provider claude-multi supports.

How to choose

Start with how you pay. GLM is subscription-only: the Anthropic endpoint requires a Coding Plan, and there is no metered equivalent. DeepSeek and Kimi sell API tokens only. MiniMax sells both from the same URL, so the key type decides which quota is billed. MiMo and Qwen each have a second template for their subscription plans because those plans use a different endpoint, which is why six providers produce eight templates.

Context is no longer the main filter. Every template now maps a 1M-context model: GLM-5.3 and GLM-5.3-Flash, MiniMax-M3.1-Flash, DeepSeek Flash, MiMo V2.6, Kimi K3.1 Pro, and the Qwen 3.8 and 3.7 models. Claude Code caps output at 128,000 tokens for these unrecognized model IDs, so larger model maximums do not change the practical ceiling.

Price and plan coverage are the real differences. MiniMax and DeepSeek put a cheap Flash model in every slot. MiMo's Flash is cheaper still. Kimi charges flagship rates and offers Highspeed at double the price, so standard speed stays the default. Qwen's pay-as-you-go route reaches Qwen3.8-Max and Qwen3.8-Flash, while the Coding Plan covers only qwen3.7-plus among current Qwen models. GLM is a fixed credit plan. Pick the route that matches the models you need, not just the lowest headline rate.

Then check geography. The MiMo Token Plan has CN, SG, and EU endpoints, and Qwen has separate international and China hosts. Pick the endpoint that matches your account and latency, not the one in the example.

Template reference

Each template pre-fills ANTHROPIC_BASE_URL, model mappings, and related env vars into your instance's settings.json. You only need to supply your API key.

The table reflects the current templates. Every template also maps the Fable alias to its primary model, and the effort-capable templates setCLAUDE_CODE_ALWAYS_ENABLE_EFFORT=1. Each provider page lists the live env vars and any manual options, such as DeepSeek V4-Pro.

Pay per token vs. subscription

Some providers offer two access models, a pay-per-token API and a subscription/coding plan, each with a different base URL. Use the right template for your account type:

• GLM: the Anthropic-compatible endpoint (api.z.ai/api/anthropic) is exclusive to the Z.ai Coding Plan subscription. The metered API only exposes an OpenAI-compatible URL. GLM-5.3 fills primary, opus, sonnet, and fable; GLM-5.3-Flash fills haiku. Both use the 1M hint and a 1M compaction window.
• DeepSeek: the template uses deepseek-flash[1m] in primary, opus, sonnet, and fable, and deepseek-flash in haiku and subagents. V4-Pro is the previous generation and remains a manual option. There is no released V4.1 Pro.
• MiMo Token Plan: defaults to the CN regional endpoint. If your subscription is SG or EU, update ANTHROPIC_BASE_URL with the endpoint shown in your subscription console. Both MiMo templates use the V2.6 model IDs.
• MiniMax: both plans use the same api.minimax.io endpoint; the API key type determines which quota is consumed. M3.1 Flash fills every slot, has a 1M token context window and a 512K output maximum, and always thinks. M3 remains available as the previous generation. M3.1 Flash costs $0.15/$0.60 per 1M tokens.
• Kimi: K2.5 was retired on 2026-08-31 and returns an error. K2.6 and K2.7 closed to new users on 2026-10-31 and retire by the end of December. The template uses K3.1 Pro in every slot at standard speed. Highspeed is optional and costs double. K3.1 Pro costs $2/$10 per 1M tokens.
• Qwen: pay-as-you-go uses Qwen3.8-Max, Qwen3.7-Plus, and Qwen3.8-Flash with a 983616 thinking-mode context setting and a 128000 output ceiling. The Coding Plan uses qwen3.7-plus in every slot and does not cover Qwen3.8-Max or Qwen3.8-Flash among current Qwen models. Qwen3-Coder-Next and Qwen3-Coder-Plus are not on the Anthropic-compatible matrix.