OpenRouter - Anthropic: Prompt cache is not working

[jamesst20]

Reproduction steps

1. Setup OpenRouter provider with API Key
2. Open a new Zed Agent
3. Select Claude Sonnet 4.6 from OpenRouter
4. Make a few prompts

Current vs. Expected behavior

native_tokens_cached is always 0

Zed version and system specs

Zed: v0.229.0+stable.206.7c07887d9555953bca7fe78417602114092ef8c8 (Zed)
OS: macOS 26.3
Memory: 48 GiB
Architecture: aarch64

Attach Zed log file

I don't see any log about AI/Agent

Relevant Zed settings

settings.json

// Zed settings
//
// For information on how to configure Zed, see the Zed
// documentation: https://zed.dev/docs/configuring-zed
//
// To see all of Zed's default settings without changing your
// custom settings, run `zed: open default settings` from the
// command palette (cmd-shift-p / ctrl-shift-p)
{
  "edit_predictions": {
    "provider": "copilot",
  },
  "show_wrap_guides": true,
  "wrap_guides": [130],
  "tab_size": 2,
  "hard_tabs": false,
  "agent": {
    "default_model": {
      "provider": "anthropic",
      "model": "claude-sonnet-4-6-latest",
      "enable_thinking": false,
    },
    "favorite_models": [
      {
        "provider": "openrouter",
        "model": "openai/gpt-5.4",
        "enable_thinking": true,
      },
      {
        "provider": "openrouter",
        "model": "anthropic/claude-sonnet-4.6",
        "enable_thinking": true,
      },
    ],
    "tool_permissions": {
      "tools": {
        "terminal": {
          "always_allow": [
            {
              "pattern": "^pnpm\\s+exec(\\s|$)",
            },
            {
              "pattern": "^cat\\b",
            },
          ],
        },
        "fetch": {
          "default": "allow",
        },
      },
    },
    "default_profile": "write",
  },
  "session": {
    "trust_all_worktrees": true,
  },
  "base_keymap": "VSCode",
  "autosave": {
    "after_delay": {
      "milliseconds": 1000,
    },
  },
  "theme": {
    "mode": "system",
    "light": "One Light",
    "dark": "One Dark",
  },
  "languages": {
    "Ruby": {
      "language_servers": ["ruby-lsp", "!sorbet", "!solargraph", "!rubocop"],
    },
    "HTML+ERB": {
      "language_servers": ["ruby-lsp"],
    },
    "JS+ERB": {
      "language_servers": ["ruby-lsp"],
    },
    "YAML+ERB": {
      "language_servers": ["ruby-lsp"],
    },
  },
  "lsp": {
    "tailwindcss-language-server": {
      "settings": {
        "experimental": {
          "classRegex": ["\\bclass:\\s*['\"]([^'\"]*)['\"]"],
        },
      },
    },
  },
}

Relevant Keymap

 N/A

(for AI issues) Model provider details

• Provider: OpenRouter
• Model: Claude Sonnet 4.6
• Mode: Agent Panel

If you are using WSL on Windows, what flavor of Linux are you using?

None

[zed-community-bot]

This issue appears to be a duplicate of:

• AI: Support OpenRouter caching for GLM models #37528

If this is indeed a duplicate:
Please close this issue and subscribe to the linked issue for updates (select "Close as not planned" → "Duplicate")

If this is a different issue:
No action needed. A maintainer will review this shortly.

Why were these issues selected?

#37528: Both issues report that OpenRouter's caching feature is not working in Zed. Issue #37528 specifically mentions 'Zed does not use OpenRouter's cache feature for supported models' and the new issue shows 'native_tokens_cached is always 0' for Claude Sonnet 4.6 via OpenRouter. Both indicate the same underlying problem: Zed's OpenRouter integration lacks cache header/parameter support.

Shared root cause: OpenRouter provider not implementing prompt caching functionality for Anthropic models

---

_{This is an automated analysis and might be incorrect.}

[jamesst20]

I took a quick glance at the Rust provider code

1. According to the OpenRouter documentation, we can inject a top-level cache-control field at the top level of the request and it will take care of everything https://openrouter.ai/docs/guides/best-practices/prompt-caching#anthropic-claude.
2. The last segment of a message is marked as cached ephemeral in Anthropic provider https://github.com/zed-industries/zed/blob/main/crates/language_models/src/provider/anthropic.rs#L694 but there is no caching equivalent in the OpenRouter provider https://github.com/zed-industries/zed/blob/main/crates/language_models/src/provider/open_router.rs

[thelok1s]

I am experiencing the same issue with claude-4.6-opus via openrouter. In a session with just 4 prompts, the agent performed 130 tool calls. Because prompt caching wasn't working (same result of 0 tokens cached according to openrouter activity log), each request resent the full context. In result 4 prompts that would take approximately 3$ were priced at least 7 times (!!) more.

[lixiaoyan]

Just to add to this: similarly, Zed currently haven't implemented Claude prompt caching for Vercel AI Gateway or OpenAI-compatible endpoints either. I think we should add support across all these similar scenarios.

[miguelraz]

@jamesst20 thanks for the report!

Where are you clicking / navigating to to see this "native_tokens_cached" information?

I don't see anything in the Zed.log, the Telemetry log, or in dev: open acp logs about this.

[jamesst20]

@jamesst20 thanks for the report!

Where are you clicking / navigating to to see this "native_tokens_cached" information?

I don't see anything in the Zed.log, the Telemetry log, or in dev: open acp logs about this.

No problem!
These logs can be found directly on OpenRouter usage history