Claude Code

The bInference Router speaks the same Messages API as Anthropic, so Claude Code talks to it directly. There is nothing extra to install and nothing running on your computer besides Claude Code itself. Every call is paid from your bInference credit.

You need Claude Code installed (claude --version prints a version) and a bInference account or agent with credit.

{
  "model": "sonnet",
  "env": {
    "ANTHROPIC_BASE_URL": "https://binference.io/api",
    "ANTHROPIC_AUTH_TOKEN": "binf_...",
    "ANTHROPIC_API_KEY": "",
    "ANTHROPIC_DEFAULT_HAIKU_MODEL": "claude-haiku-4-5",
    "CLAUDE_CODE_MAX_OUTPUT_TOKENS": "32000"
  }
}

export ANTHROPIC_BASE_URL="https://binference.io/api"
export ANTHROPIC_AUTH_TOKEN="binf_..."
export ANTHROPIC_API_KEY=""
export ANTHROPIC_DEFAULT_HAIKU_MODEL="claude-haiku-4-5"
export CLAUDE_CODE_MAX_OUTPUT_TOKENS=32000
claude --model sonnet

Models to start withCopy Anchor Link

Every model was run through Claude Code on the Router before this page was written, each reading files and running commands with Claude Code's tools. Any other model on Models that handles tools well works the same way: use its id with /model.

Claude Code is built for Claude models. GPT and Abliteration models work through the Router, but Anthropic doesn't support them in Claude Code, so a feature may behave differently. On them, Claude Code prints that it doesn't know the model and keeps the session within 200,000 tokens before it compacts. That is expected.

More Claude models in the menu: add "CLAUDE_CODE_ENABLE_GATEWAY_MODEL_DISCOVERY": "1" to env, and /model also lists every Claude model the Router serves, older versions included, by name.

What a turn holdsCopy Anchor Link

Before each call starts, the Router holds enough credit for the longest answer the call allows, plus the conversation so far. You pay only what the model writes, and the rest comes back when the call ends. See How a call is paid.

Claude Code asks for long answers: room for 128,000 tokens on Opus 5.5. At the start of a session, that holds about $2.40 per turn, and Claude Code's helpers running at the same time each hold their own. With the 32,000 cap from step 2, the same Opus turn holds about $0.67, and a Sonnet 5.5 turn about $0.33. Holds grow as the conversation grows, and pricier models like GPT-6 Astra hold several times more.

When your credit can't cover a hold, Claude Code shows API Error: 402 and that the balance is below the request's reserve. Lower CLAUDE_CODE_MAX_OUTPUT_TOKENS, for example to 16000, pick a cheaper model, or add credit.

What you pay is a different story: Claude Code resends the whole conversation each turn, and the Router passes its cache marks through. In our test on Sonnet 5.5, the second call read 26,463 tokens from the cache and cost about a twelfth of the first.

Good to knowCopy Anchor Link

• Long sessions: a request over 4 MB is refused with 413, usually because of pasted screenshots. Run /compact or /clear. To compact sooner, add "CLAUDE_CODE_AUTO_COMPACT_WINDOW": "400000" to env.
• /cost shows Anthropic's prices, not what you paid. Your real charges are on the Calls page.
• /context counts are estimates: the Router doesn't count tokens ahead of a call, so Claude Code estimates them.
• Auto mode may say the session "isn't eligible" for free safety checks. Press Enter: the checks run as ordinary calls on your key. To skip the notice, add "CLAUDE_CODE_AUTO_MODE_SERVER": "0" to env.
• Remote Control and voice dictation need a Claude sign-in, so they're off while the key is set.
• Claude desktop app: its Code tab keeps its own setting for the address and ignores this file. This page covers Claude Code in a terminal and in VS Code.
• VS Code extension: put the same variables in VS Code's own settings, under claudeCode.environmentVariables, as { "name": "ANTHROPIC_BASE_URL", "value": "https://binference.io/api" } entries. The extension checks for a key there before it starts.
• Let Claude Code watch its credit: connect the MCP server with the same key, and Claude Code can check its own balance, usage and limits.

Go back to your Claude planCopy Anchor Link

Delete the lines you added in step 2, or the whole env block, and start Claude Code again. Your Claude sign-in was kept: /status shows it as the login method again.

If something goes wrongCopy Anchor Link

Every error code is on Errors, with what to do.