Comparison

Claude Code with GLM, Kimi or DeepSeek vs real Claude

Published November 3, 2026 · Last checked November 3, 2026

Claude Code runs on any endpoint that speaks Anthropic's Messages format, and Z.ai (GLM), Moonshot (Kimi) and DeepSeek each publish one. Pointing Claude Code at them can be cheaper, but the answers then come from GLM, Kimi or DeepSeek, not Claude, and some Claude Code features behave differently. This page lists their prices and model mappings from each vendor's own docs, next to Anthropic's prices and CoreCode's, so you can decide which trade you want.

Side by side

Prices are per million tokens (MTok) unless the row says otherwise, as listed on each vendor's page on November 3, 2026. CoreCode's yuan prices are converted at the ¥6.8 per dollar its pricing page uses; see /docs/en/pricing/ for the current table.

The CoreCode prices are what you pay in yuan; the dollar figures are conversions for comparison. Token counts are not comparable across models: each model has its own tokenizer, and one task can take a different number of turns on a different model, so the per-token price is only part of the cost.

Z.ai GLM Coding Plan

The GLM Coding Plan is a monthly subscription. Its overview measures usage in credits with two limits: Lite gets 2,000 credits per 5 hours and 10,000 a week, Pro 12,000 and 60,000, Max 28,000 and 140,000. Calls from Monday to Friday, 14:00 to 18:00 Singapore time (UTC+8), cost the full credit rate; outside those hours they cost half. All plans include GLM-5.3 and GLM-5.3-Flash, and requests for older GLM models are routed to these two.

The Claude Code guide sets ANTHROPIC_BASE_URL to https://api.z.ai/api/anthropic and maps Claude Code's model settings to GLM. Its setup script maps Opus, Sonnet and Haiku all to GLM-5.3-Flash; its manual example uses glm-5.3[1m] for Opus and Sonnet. Check which mapping your settings.json ends up with, because it decides which GLM model answers.

DeepSeek

DeepSeek's Anthropic API guide uses https://api.deepseek.com/anthropic. Model names starting with claude-opus go to deepseek-v4-pro, and names starting with claude-sonnet or claude-haiku go to deepseek-flash, so Claude Code's default model names work without changes. The same page lists what it does not honor:

• cache_control is ignored, on messages and on tools. DeepSeek's price list has its own cache-hit price instead.
• In thinking, budget_tokens is ignored.
• deepseek-v4-pro replaces images with "[Unsupported Image]".
• document, search_result and redacted_thinking content blocks are not supported.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays; other times cost half the peak price.

Kimi

Moonshot's Claude Code with Kimi guide sets ANTHROPIC_BASE_URL to https://api.moonshot.ai/anthropic and maps Opus and Sonnet to kimi-k3[1m], Haiku to kimi-k2.7-code. It warns that kimi-k2.7-code requires thinking to be on and returns a 400 error otherwise, and that setting only some of the variables makes the unset cases fail silently.

On Kimi's price list, kimi-k3 is $3 input and $15 output per million tokens, above Claude Sonnet 5.5. The same page says K2 models are closed to new users from October 31, 2026 and will be retired by the end of December 2026, which includes the Haiku mapping above.

What changes when the model is not Claude

• Support. Anthropic's LLM gateway page says it "doesn't support routing Claude Code to non-Claude models through any gateway". Claude Code still starts, but a problem that comes from the model is not one Anthropic will fix.
• Features that depend on the API. Claude Code sends prompt-cache markers, thinking settings and images. An endpoint that ignores or rejects some of them, as DeepSeek's and Kimi's docs describe, changes cost and behavior.
• Model names. /status and /model show the names Claude Code sent, such as Opus or Sonnet. They do not tell you which model answered.
• Quality. This page makes no claim about which model codes better. Try your own tasks.

None of this is hidden by these vendors: each documents its mapping. The problem is a relay that sells "Claude" and quietly does the same thing. How to check whether a Claude relay is genuine has a script that tells the difference, because only Anthropic rejects a tampered thinking signature.

Which one to choose

• You want the lowest fixed monthly cost and accept another model: a GLM Coding Plan.
• You want pay-as-you-go at the lowest token price and accept another model: DeepSeek, keeping in mind that cache_control is ignored.
• You want Claude and can pay Anthropic: Claude Pro or Max, or the API; see Claude Code subscription vs API.
• You want Claude below Anthropic's price and can pay in yuan: CoreCode. Every request is answered by an official Anthropic Claude account, one Claude Code session stays on one account so the cache keeps hitting, and each request's tokens and cost are in the usage log. Payment is Alipay or WeChat Pay in yuan only, at ¥1 = $1 of credit.

Switch Claude Code back to Claude through CoreCode

1. Remove the old mapping. Delete the ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL and ANTHROPIC_MODEL entries that point at GLM, Kimi or DeepSeek, from ~/.claude/settings.json and from your shell profile.
2. Set two variables. Create a key in the Claude 满血（仅限CC） group on API Keys (the new user guide covers top-up), then:

   {
     "env": {
       "ANTHROPIC_BASE_URL": "https://corecode.live",
       "ANTHROPIC_AUTH_TOKEN": "sk-your-key"
     }
   }

3. Check. Open a new terminal, run claude -p "reply ok", then /status: the base URL should read https://corecode.live. Details are in Claude Code.

FAQ

Sources

• Z.ai: GLM Coding Plan (prices; checked November 3, 2026)
• Z.ai: Coding Plan overview (credits, peak hours, models)
• Z.ai: Claude Code and Z.ai: Pricing
• DeepSeek: Anthropic API and DeepSeek: Pricing
• Kimi: Claude Code with Kimi and Kimi: Pricing
• Claude Code: Other LLM gateways
• Anthropic: Pricing