NewHow we catch a seller listing the wrong model, and refund you if one does.
UberLLM

Can Claude Code run on DeepSeek, GLM or Kimi models?

Yes. Through UberLLM’s Anthropic-compatible endpoint, any chat model on the market can be mapped to Claude Code’s model slots, so Claude Code can run on DeepSeek, GLM, Kimi or Qwen models. On November 19, 2026, 70 models from those four makers were taking requests. MoonshotAI: Kimi K3.1 Pro cost $0.048 per million input tokens and $1.6796 per million output tokens. The installer at uberllm.dev/claude-code writes the mapping for you. A request that fails is not billed.

· ·

Prices as of November 19, 2026, 08:00 UTC. This page is rebuilt every hour.

Set it up in one terminal

Claude Code reads two settings to use another endpoint: ANTHROPIC_BASE_URL and ANTHROPIC_AUTH_TOKEN. Set the first to the Anthropic-compatible address of UberLLM, https://api.uberllm.dev (with no /v1 at the end), and the second to an UberLLM key. Then start Claude Code.

bash
export ANTHROPIC_BASE_URL=https://api.uberllm.dev
export ANTHROPIC_AUTH_TOKEN=ull_your_key
claude

That lasts for the current terminal. ull_your_key stands for a key from your keys page. With only these two set, Claude Code asks for its own model names and UberLLM serves the same Claude models. To run on DeepSeek, GLM, Kimi or Qwen, add the model variables from the block for that family below.

One block per model family

Claude Code has four model slots, each with its own variable: ANTHROPIC_DEFAULT_OPUS_MODEL, ANTHROPIC_DEFAULT_SONNET_MODEL, ANTHROPIC_DEFAULT_HAIKU_MODEL (Claude Code also uses it for background work such as titles and summaries) and CLAUDE_CODE_SUBAGENT_MODEL. ANTHROPIC_MODEL is the model the session starts on. Map every slot: a slot left unset keeps asking for a Claude model, and you pay Claude prices for it.

Claude Code knows the context size of Claude models only, and assumes 200,000 tokens for any other model. So a block for a model with a 1M-token context ends the model id with [1m], and a block for a model of another size sets CLAUDE_CODE_MAX_CONTEXT_TOKENS to that size. A context set larger than the model has makes requests fail once the conversation outgrows the model.

Each block below sets every slot to one model, with the model id as it is in the UberLLM catalog today. Any other chat model id from the table on this page that takes tool calls works the same way.

DeepSeek

deepseek/deepseek-v4.1-flash (DeepSeek V4.1 Flash) was the most used DeepSeek model with tool calls on UberLLM on November 19, 2026. It reads up to 1,048,576 tokens of context.

bash
export ANTHROPIC_BASE_URL=https://api.uberllm.dev
export ANTHROPIC_AUTH_TOKEN=ull_your_key
export ANTHROPIC_MODEL='deepseek/deepseek-v4.1-flash[1m]'
export ANTHROPIC_DEFAULT_OPUS_MODEL='deepseek/deepseek-v4.1-flash'
export ANTHROPIC_DEFAULT_SONNET_MODEL='deepseek/deepseek-v4.1-flash'
export ANTHROPIC_DEFAULT_HAIKU_MODEL='deepseek/deepseek-v4.1-flash'
export CLAUDE_CODE_SUBAGENT_MODEL='deepseek/deepseek-v4.1-flash'
claude

How do I use DeepSeek V4.1 Flash in Claude Code?

GLM

z-ai/glm-5.3 (GLM 5.3) was the most used GLM model with tool calls on UberLLM on November 19, 2026. It reads up to 1,000,000 tokens of context.

bash
export ANTHROPIC_BASE_URL=https://api.uberllm.dev
export ANTHROPIC_AUTH_TOKEN=ull_your_key
export ANTHROPIC_MODEL='z-ai/glm-5.3[1m]'
export ANTHROPIC_DEFAULT_OPUS_MODEL='z-ai/glm-5.3'
export ANTHROPIC_DEFAULT_SONNET_MODEL='z-ai/glm-5.3'
export ANTHROPIC_DEFAULT_HAIKU_MODEL='z-ai/glm-5.3'
export CLAUDE_CODE_SUBAGENT_MODEL='z-ai/glm-5.3'
claude

How do I use GLM 5.3 in Claude Code?

Kimi

moonshotai/kimi-k3.1-pro (MoonshotAI: Kimi K3.1 Pro) was the most used Kimi model with tool calls on UberLLM on November 19, 2026. It reads up to 1,048,576 tokens of context.

bash
export ANTHROPIC_BASE_URL=https://api.uberllm.dev
export ANTHROPIC_AUTH_TOKEN=ull_your_key
export ANTHROPIC_MODEL='moonshotai/kimi-k3.1-pro[1m]'
export ANTHROPIC_DEFAULT_OPUS_MODEL='moonshotai/kimi-k3.1-pro'
export ANTHROPIC_DEFAULT_SONNET_MODEL='moonshotai/kimi-k3.1-pro'
export ANTHROPIC_DEFAULT_HAIKU_MODEL='moonshotai/kimi-k3.1-pro'
export CLAUDE_CODE_SUBAGENT_MODEL='moonshotai/kimi-k3.1-pro'
claude

How do I use MoonshotAI: Kimi K3.1 Pro in Claude Code?

Qwen

qwen/qwen3.8-omni-flash (Qwen: Qwen3.8 Omni Flash) was the most used Qwen model with tool calls on UberLLM on November 19, 2026. It reads up to 1,000,000 tokens of context.

bash
export ANTHROPIC_BASE_URL=https://api.uberllm.dev
export ANTHROPIC_AUTH_TOKEN=ull_your_key
export ANTHROPIC_MODEL='qwen/qwen3.8-omni-flash[1m]'
export ANTHROPIC_DEFAULT_OPUS_MODEL='qwen/qwen3.8-omni-flash'
export ANTHROPIC_DEFAULT_SONNET_MODEL='qwen/qwen3.8-omni-flash'
export ANTHROPIC_DEFAULT_HAIKU_MODEL='qwen/qwen3.8-omni-flash'
export CLAUDE_CODE_SUBAGENT_MODEL='qwen/qwen3.8-omni-flash'
claude

How do I use Qwen: Qwen3.8 Omni Flash in Claude Code?

Keep it in a settings file

To keep the set-up without a key in your shell profile, save the key to a file only you can read:

bash
mkdir -p ~/.uberllm && chmod 700 ~/.uberllm
printf '%s\n' 'ull_your_key' > ~/.uberllm/key && chmod 600 ~/.uberllm/key

Then put the address, the key helper and the model variables in ~/.claude/settings.json (every project) or in <project>/.claude/settings.local.json (one project; your other projects stay on your Anthropic plan). This is the DeepSeek block above as a settings file:

json
{
  "apiKeyHelper": "cat ~/.uberllm/key",
  "env": {
    "ANTHROPIC_BASE_URL": "https://api.uberllm.dev",
    "ANTHROPIC_MODEL": "deepseek/deepseek-v4.1-flash[1m]",
    "ANTHROPIC_DEFAULT_OPUS_MODEL": "deepseek/deepseek-v4.1-flash",
    "ANTHROPIC_DEFAULT_SONNET_MODEL": "deepseek/deepseek-v4.1-flash",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "deepseek/deepseek-v4.1-flash",
    "CLAUDE_CODE_SUBAGENT_MODEL": "deepseek/deepseek-v4.1-flash"
  }
}

Claude Code runs apiKeyHelper to get the key, so the key itself is never written into a settings file. Once the file is in place, remove every ANTHROPIC_* variable from your shell and your shell profile: a token exported in the shell replaces the saved key, and ANTHROPIC_MODEL exported in the shell beats the model setting.

The installer writes the same file for you and gets the key through your browser: run npx uberllm init, or use the commands on the Claude Code set-up page. Every setting is explained in the Claude Code guide in the docs.

Check that it works

Start Claude Code and run /status: the API address is https://api.uberllm.dev. Then ask it anything. The request shows in Dashboard → Activity within a minute, with its cost.

The same key answers over the Anthropic Messages API, which is the address Claude Code calls:

bash
curl https://api.uberllm.dev/v1/messages \
  -H "Authorization: Bearer $(cat ~/.uberllm/key)" \
  -H "anthropic-version: 2023-06-01" \
  -H "content-type: application/json" \
  -d '{"model":"deepseek/deepseek-v4.1-flash","max_tokens":16,"messages":[{"role":"user","content":"ping"}]}'

What works on a model that is not Claude, and what changes

  • Tool calls. Claude Code works through tool calls (reading files, running commands), so it needs a model that takes them; the blocks above use only such models. A request with tools sent to a model that cannot call them is refused with 400 tools_not_supported, and the message names models that can. A reply that calls a tool always ends with stop_reason: "tool_use", streamed or not.
  • Thinking. On /v1/messages a model thinks only when the request asks for it ("thinking": {"type": "enabled"} or "adaptive"), as on the API of Anthropic, and its thinking comes back whichever provider served the call. A few models cannot turn thinking off; they think either way.
  • Thinking blocks. Claude Code requires a signature field on thinking it gets back. Some providers sent the reasoning of DeepSeek without it, and a long session could stop. Since October 28, 2026 replies always include the field.
  • DeepSeek, thinking and tools. The thinking mode of DeepSeek wants the reasoning of every earlier turn sent back when tools are used, and answers 400 when it is missing (DeepSeek: thinking mode). When a DeepSeek provider refuses a tool call for that reason, UberLLM answers the call with thinking off instead of failing it.
  • Kimi sampling. Kimi K3 and Kimi K3.1 Pro run at a fixed temperature and top_p and refuse a request that sets its own. Since November 18, 2026 such a request is sent again without that setting and answered at the value of the model.
  • Context. A prompt longer than the model takes is refused with 400 context_length_exceeded; the message gives both sizes. When max_tokens plus the prompt is more than the model can hold, the reply gets the room that is left instead of an error.
  • The cost Claude Code shows is not your bill. Claude Code works it out on your computer from the list prices of Anthropic, and shows nothing useful for other models. What you pay is in Dashboard → Activity, and the status line can show your balance.
  • Your claude.ai login is not used. Features that need it, such as claude.ai connectors, /schedule and Remote Control, are not available in a session that runs on UberLLM.
  • Your code goes to third-party model providers. Everything the session reads is sent to the model, which is run by third-party providers on the UberLLM market. Do not use it in a repository that holds secrets, credentials or customer data. API keys, passwords and tokens found in a prompt are hidden from the provider; when one cannot be hidden, the request is refused with 400 secret_in_request and is not charged.

Troubleshooting

These are the errors people have hit, in the words the API or Claude Code uses. Every refused or failed request is also listed in Dashboard → Activity with the message you were sent, and the X-Request-Id response header is the id of that row.

  • 402 insufficient_credits, “Your balance is too low for this request, so it was not sent.” The prepaid balance cannot cover the request. Add credit in Dashboard → Credits and send it again.
  • 402 key_spend_cap_exceeded, “This API key has reached its spend limit.” Raise the limit of the key on the API keys page, or use another key.
  • 400 max_price_exceeded, “No provider is selling … at its listed price right now. Nothing was charged.” The key is set to never pay more than the listed price, and no seller was that cheap at that moment. Try again in a minute, or turn that setting off for the key on the API keys page. Since November 18, 2026 the listed price moves within minutes to the cheapest price sellers really fill; before, such a key could be refused for a few hours.
  • “Parameter 'temperature'=0.5 is not supported for kimi-k3 model.” Kimi K3 and Kimi K3.1 Pro run at a fixed temperature and top_p. Since November 18, 2026 UberLLM sends such a request again without that setting; a value out of range is still refused.
  • “Missing required field in assistant message: 'signature'.” Fixed on October 28, 2026: replies always include the field. Nothing to change on your side.
  • “Your apiKeyHelper script is failing.” The key file is missing or unreadable. Save the key again with the two commands above, or run the installer again.
  • Claude Code still asks you to log in, or still uses your Anthropic plan. ANTHROPIC_AUTH_TOKEN or ANTHROPIC_API_KEY exported in the shell wins over the saved key, and Claude Code reads the settings of a project only after you trust that folder.
  • Claude models on your UberLLM bill. A model slot was left unset, so Claude Code asked for a Claude model. Set ANTHROPIC_MODEL and all four slot variables.
  • Requests fail once the conversation gets long. The context told to Claude Code is larger than the model has. Use the real context size of the model, which its page on uberllm.dev shows.
  • 404 model_not_found. The model id is not in the catalog. GET https://api.uberllm.dev/v1/models lists the ids.
  • 503 no_available_provider. Nobody can serve that model right now, even after the automatic retries. Retry shortly or pick another model. Nothing was charged.
  • 429 rate_limited. The key went over its request limit per minute. Wait and retry, or raise the rate limit of the key.

Switch back to Claude

To keep paying per token but run on Claude models again, keep the address and the key and remove only the model variables: Claude Code then asks for its own model names and UberLLM serves the same Claude models.

To go back to your Anthropic login, close the terminal, or clear the variables in it:

bash
unset ANTHROPIC_BASE_URL ANTHROPIC_AUTH_TOKEN ANTHROPIC_MODEL \
  ANTHROPIC_DEFAULT_OPUS_MODEL ANTHROPIC_DEFAULT_SONNET_MODEL \
  ANTHROPIC_DEFAULT_HAIKU_MODEL CLAUDE_CODE_SUBAGENT_MODEL \
  CLAUDE_CODE_MAX_CONTEXT_TOKENS

For a settings file, remove apiKeyHelper and the ANTHROPIC_* entries of its env block. If the installer wrote the file, it can undo its own work: it backs up the settings file first, removes only what it added, and deletes the saved key file. Claude Code goes back to your Anthropic login on the next start.

bash
curl -fsSL https://uberllm.dev/install/claude-code.sh | sh -s -- --uninstall

Removing the settings does not revoke the key. Delete the key on your keys page to make sure it can never spend again.

The direct route: the endpoint of each maker

Claude Code can also be pointed straight at the Anthropic-compatible endpoint of a model maker, with a key from that maker. These are from each maker’s own page, read on November 19, 2026:

  • Kimi (Moonshot AI): ANTHROPIC_BASE_URL is https://api.moonshot.ai/anthropic and the model is kimi-k3[1m]. Source: Use Kimi in Claude Code.
  • GLM (Z.ai): ANTHROPIC_BASE_URL is https://api.z.ai/api/anthropic. Source: Z.ai: Claude Code.
  • DeepSeek: ANTHROPIC_BASE_URL is https://api.deepseek.com/anthropic. The page lists what that endpoint ignores, such as cache_control and the budget_tokens of thinking. Source: DeepSeek: Anthropic API.

The endpoint of a maker serves the models of that maker, billed by that maker. Through UberLLM one key and one address reach all of these families and Claude, from one prepaid balance, and each request is billed what the seller serving it charged.

ModelInput / 1M tokensOutput / 1M tokensBelow list price
DeepSeek V4.1 Flash$0.00248$0.0099199.2%
DeepSeek V4.1 Flash 1017$0.00718$0.028797.6%
GLM 5.3$0.000663$0.013399.7%
GLM 5.3 Flash$0.002$0.0066698.7%
MoonshotAI: Kimi K3.1 Pro$0.048$1.6796
Kimi K3$0.0154$1.2891.5%
Qwen: Qwen3.8 Omni Flash$0.000146$0.00045699.9%
GLM-5.2$0.00415$0.016299.6%
Qwen3.8 Flash$0.015$0.04790%
DeepSeek V4 Flash$0.0144$0.028884%
Kimi K2.7 Code$0.009$0.052598.5%
Qwen 3.8 27B$0.000158$0.00507
DeepSeek V4 Pro 0813$0.00277$0.110697.2%
DeepSeek V4 Pro$0.0412$0.068696.9%
Z.ai: GLM 5.3 Prime$0.0121$0.038199.3%
Qwen3 Coder Next$0.00119$0.0079299%
Qwen 4 Max$0.045$0.135196.8%
Qwen3 Coder 30B A3B Instruct$0.000693$0.0027799.5%
DeepSeek V4 Flash 0731$0.00293$0.0087995.1%
Qwen3.8 2.4T A95B$0.0106$0.031999.5%

UberLLM adds no fee to these prices · every model and price →

More questions

How do I pick the model?

Claude Code reads its models from settings such as ANTHROPIC_DEFAULT_OPUS_MODEL. Set each one to a model id from the table, or choose a ready-made profile in the installer.

Does it work with other coding agents?

Yes. Codex CLI, OpenCode, Cline, Kilo Code, Cursor, Aider and Continue have setup guides at uberllm.dev/guides.