openrouter: Anthropic models still get no prompt caching by default in opencode run (#39009 not fully fixed) #51726

Description

Summary

#39009 reported that the v2 OpenRouter route never sets cache_control for Anthropic models. It was closed by #39907. On 2.0.18, opencode run with Anthropic models on OpenRouter still reports zero cache reads and zero cache writes on every step, so every step bills the full prompt as fresh input.

The #39907 PR description says it uses a "stable session prompt_cache_key for default OpenRouter caching without emitting explicit markers". But OpenRouter only caches Anthropic models when explicit cache_control markers are present. So caching looks fixed only when a cache policy is set explicitly. The default case is still broken.

We confirmed three things:

• The same prompt caches on opencode 1.18.32.
• It caches on 2.0.18 when cache_control is set by hand as a model option. This is the workaround from OpenRouter route never sets cache_control for Anthropic models #39009.
• Interactive (TUI/desktop) sessions on 2.0.18 cache normally.

The impact is large for headless, multi-step runs. In our case, a 2-step extraction run over a ~120k-token prompt cost $0.94. We estimate that caching would have cut that roughly in half.

Environment

• opencode version: 2.0.18 (~/.opencode/bin/opencode, install/channel: latest). The comparison is with 1.18.32 (Homebrew).
• OS: Darwin 25.6.0 arm64 (macOS)
• Terminal: iTerm.app, TERM=xterm-256color
• Shell: /bin/zsh
• Provider: openrouter (OpenRouter API key auth). OpenRouter routed every request in these tests to Amazon Bedrock (tool call IDs are toolu_bdrk_…) on both versions and with and without the workaround. So the upstream provider is not the difference.
• Active plugins: local files in ~/.config/opencode/plugins/: temp.ts (v2 Plugin.define), rtk.ts and herdr-agent-state.js. The last two fail to load on v2. None are referenced in opencode.json. The repro uses the default agent and no plugin tools.

Reproduction

In a directory that has a README.md and a pyproject.toml (any two small files will do):

1. opencode 2.0.18, default config:

   opencode run --format json --title cache-repro-v2 \
     --model openrouter/anthropic/claude-sonnet-5 \
     "Read README.md, then read pyproject.toml, then reply with just: ok"

   Then run opencode session export <id> and look at messages[].tokens.cache on the assistant messages.
2. Add this project opencode.json and run the same command again:

   {
     "provider": { "openrouter": { "models": {
       "anthropic/claude-sonnet-5": { "options": { "cache_control": { "type": "ephemeral" } } }
     }}}
   }

3. Optional: run the same prompt on opencode 1.18.32 without the override, and look at part.tokens.cache on the step-finish events.

Expected Behavior

With no extra config, step 1 writes the shared prefix (system prompt, tools, first user message) to the cache, and later steps read it. That is what 1.18.32 does by default:

Actual Behavior

On 2.0.18 with default config, every step reports cache: {read: 0, write: 0} and bills the whole prompt as input:

On 2.0.18 with the cache_control model option (step 2), caching works:

A cache write of 0 on step 1 without the override shows that no breakpoint reaches the provider. The model is not simply missing an existing cache.

Additional Context

Consistent. Every opencode run session with an Anthropic model on 2.0.18 without the override shows zero cache use, across 10 runs and every variant we tried:

Interactive sessions are not affected. A long interactive session on 2.0.18 with openrouter/anthropic/claude-opus-5.5 shows 14.4M cache-read tokens. So the interactive path seems to supply a cache policy or hints that run does not. We have not confirmed this in the code.

Automatic-caching models are not affected. Models that OpenRouter caches automatically (for example moonshotai/kimi-k3) show cache reads under opencode run on 2.0.18. This fits the explanation that the Anthropic cache_control markers are what's missing.

Related:

• OpenRouter route never sets cache_control for Anthropic models #39009: the original report. It has the same symptom and the same workaround, and was closed as fixed by feat(ai): expand OpenRouter native support #39907.
• feat(ai): expand OpenRouter native support #39907: its description says default OpenRouter caching uses prompt_cache_key "without emitting explicit markers". That is not enough for Anthropic models on OpenRouter.
• [FEATURE]: Set prompt_cache_ttl for OpenRouter provider #16848 and fix(opencode): set 1h prompt cache TTL for OpenRouter #16850: an earlier report and PR about OpenRouter and cache_control for Anthropic models.
• v2: session ID in the first instruction block defeats cross-session prompt caching #51007: cross-session caching in v2. It is a different issue, but it is in the same area.

Suggested fix: when an OpenRouter model is Anthropic-family (anthropic/*, ~anthropic/*), emit cache_control markers by default, or the top-level cache_control field. run should get the same behaviour as the interactive path.

Workaround: set "options": { "cache_control": { "type": "ephemeral" } } for each Anthropic model under provider.openrouter.models.

Plugins

Temp plugin

OpenCode version

2.0.18

Steps to reproduce

Run the same prompt in any directory that has a README.md and a pyproject.toml (any two small files will do):

1. opencode 2.0.18:

   ~/.opencode/bin/opencode run --format json --title cache-repro-v2 \
     --model openrouter/anthropic/claude-sonnet-5 \
     "Read README.md, then read pyproject.toml, then reply with just: ok"

   Then opencode session export <id> and look at messages[].tokens.cache on the assistant messages.
2. opencode 1.18.32:

   /opt/homebrew/bin/opencode run --format json \
     --model openrouter/anthropic/claude-sonnet-5 \
     "Read README.md, then read pyproject.toml, then reply with just: ok"

   Look at part.tokens.cache on the step-finish events.

Screenshot and/or share link

No response

Operating System

Mac 25.6.0

Terminal

iTerm2