Guides
1M Context (Claude)
Enable real 1 million-token context on Claude models via a beta header.
Claude Code — env vars
Prismma serves real 1 million token context for Claude models. Use a bare model id (no [1m] suffix) and set ANTHROPIC_BETAS=context-1m-2025-08-07.
# Claude Code — 1M context via Prismma (bare model id required; no [1m] suffix) unset ANTHROPIC_API_KEY # clear stale direct-Anthropic key — Claude Code prefers it export ANTHROPIC_BASE_URL="https://api.prismma.ai" export ANTHROPIC_AUTH_TOKEN="sk-eu-YOUR-KEY" # Claude (group 2) key export ANTHROPIC_BETAS="context-1m-2025-08-07" # required — without it Claude Code compacts at 200K claude # launch Claude Code — now routed through Prismma
Alternative: settings.json
// ~/.claude/settings.json (alternative to env vars)
{
"env": {
"ANTHROPIC_BASE_URL": "https://api.prismma.ai",
"ANTHROPIC_AUTH_TOKEN": "sk-eu-YOUR-KEY",
"ANTHROPIC_MODEL": "claude-opus-4-8",
"ANTHROPIC_BETAS": "context-1m-2025-08-07",
"ANTHROPIC_API_KEY": ""
}
}Important: Claude Code v2.x prefers ANTHROPIC_API_KEY over ANTHROPIC_AUTH_TOKEN when both are set — clear any stale ANTHROPIC_API_KEY first (unset ANTHROPIC_API_KEY) so it doesn't bypass the relay.
ANTHROPIC_BETAS=context-1m-2025-08-07 is required — without it Claude Code compacts context at 200K instead of using the full 1M window. Use bare model ids like claude-opus-4-8; the gateway now also accepts an [1m] suffix without erroring (verified 2026-07-23), but the header above is what actually enables the 1M window, not the suffix.