opencode-1m-context

OpenCode plugin that enables the 1M token context window for supported Anthropic Claude models on any provider, including GitHub Copilot.

Unlike similar plugins, this one works with all providers — Anthropic direct, third-party gateways/proxies, Amazon Bedrock, GitHub Copilot, etc. — not just providerID === "anthropic".

Supported providers

| Provider | How it works | | --- | --- | | Anthropic direct | Injects anthropic-beta: context-1m-2025-08-07 header | | Third-party gateways / proxies / tunnels | Injects anthropic-beta: context-1m-2025-08-07 header | | Amazon Bedrock | Injects anthropicBeta: ["context-1m-2025-08-07"] option | | GitHub Copilot | Clones models into -1m variants with higher limits + sets Copilot-Integration-Id header |

Supported models

Anthropic providers (header injection)

Any model whose ID contains both claude and one of [1m] or -1m (e.g. claude-opus-4-8[1m], claude-sonnet-4-6-1m).

GitHub Copilot (model cloning)

The plugin automatically creates (1M) variants for:

• Claude Opus 4.6 → claude-opus-4.6-1m
• Claude Opus 4.7 → claude-opus-4.7-1m
• Claude Opus 4.8 → claude-opus-4.8-1m
• Claude Sonnet 4.6 → claude-sonnet-4.6-1m
• GPT-5.4 → gpt-5.4-1m
• GPT-5.5 → gpt-5.5-1m

Install

{
  "plugin": ["opencode-1m-context"]
}

Configuration

Two layers are required and work together:

1. Plugin layer (API header) — injects the anthropic-beta flag so the API accepts up to 1M input tokens
2. Config layer (OpenCode limit override) — sets limit.context so OpenCode's auto-compaction triggers near 1M instead of ~168K. Without this, OpenCode truncates conversation history at 200K before the request is even sent.

Anthropic direct / third-party gateways / tunnels

{
  "plugin": ["opencode-1m-context"],
  "provider": {
    "anthropic": {
      "models": {
        "claude-opus-4-8[1m]": { "limit": { "context": 1000000, "output": 128000 } },
        "claude-sonnet-4-6[1m]": { "limit": { "context": 1000000, "output": 64000 } }
      }
    }
  }
}

Adjust the model IDs and provider name to match your setup. The [1m]/-1m marker in the model ID is what triggers the plugin's header injection.

Amazon Bedrock

{
  "plugin": ["opencode-1m-context"],
  "provider": {
    "amazon-bedrock": {
      "models": {
        "anthropic.claude-opus-4-6-v1:0": { "limit": { "context": 1000000, "output": 128000 } },
        "anthropic.claude-sonnet-4-6-v1:0": { "limit": { "context": 1000000, "output": 64000 } }
      }
    }
  }
}

You don't need to include the region prefix (us., eu., global., etc.) — OpenCode adds it automatically based on your AWS_REGION. The model ID must contain claude and a [1m]/-1m marker to trigger the plugin.

GitHub Copilot

No extra configuration needed — the plugin automatically creates the 1M variants. You'll see new model entries like "Claude Opus 4.8 (1M)" and "GPT-5.5 (1M)" alongside the originals after restarting OpenCode.

Optional: set reasoning effort for the 1M variants:

{
  "plugin": ["opencode-1m-context"],
  "provider": {
    "github-copilot": {
      "models": {
        "claude-opus-4.6-1m": { "reasoningEffort": "high" },
        "claude-opus-4.7-1m": { "reasoningEffort": "medium" }
      }
    }
  }
}

How it works

• HTTP providers (Anthropic direct, third-party gateways, etc.): Injects anthropic-beta: context-1m-2025-08-07 HTTP header via chat.headers hook
• Amazon Bedrock: Injects anthropicBeta: ["context-1m-2025-08-07"] option via chat.params hook
• GitHub Copilot: Clones each supported model into a -1m variant with model-specific long-context limits via provider.models hook, and sets the Copilot-Integration-Id: copilot-developer-cli header on all Copilot requests

Requirements

• Anthropic API key with Usage Tier 4 ($400+ cumulative deposit) or equivalent rate limits (for Anthropic direct provider)
• GitHub Copilot subscription (for GitHub Copilot provider)

Related

• opencode-copilot-1m — same idea, GitHub Copilot only
• opencode-anthropic-context-1m — same idea, Anthropic direct and Bedrock only
• opencode-anthropic-context-1m — same idea, Anthropic direct only

License

MIT