fix(provider): apply config model cost overrides at runtime - #17645

Thanks for updating your PR! It now meets our contributing guidelines. 👍

Any chance this will be merged soon? I use API accounts directly with providers that are often missing from the standard models.dev data, and thus specify most of my model configs (and pricing info) directly in my opencode.json config. This means that OpenCode's cost estimates are wildly inaccurate for me, even though the token usage counts are correct.

In this example, it appears to be pricing the cached reads at $0 even though I have the correct config for $0.06/Mtok. The 1.5M input + 68.7K output looks to equal the $0.54 total it reports, but it's missing $0.66 of cached read cost:

┌────────────────────────────────────────────────────────┐
│                      MODEL USAGE                       │
├────────────────────────────────────────────────────────┤
│ novita-ai/novita-ai/minimax/minimax-m2.7               │
│  Messages                                          216 │
│  Input Tokens                                     1.5M │
│  Output Tokens                                   68.7K │
│  Cache Read                                      11.0M │
│  Cache Write                                         0 │
│  Cost                                          $0.5447 │
└────────────────────────────────────────────────────────┘

Somehow this PR only contains a test but not real change. @mollux Can you fix this?

This PR has been open for nearly 4 months and is critical for anyone using custom providers (LiteLLM proxies, Chinese model providers like GLM/DeepSeek/Qwen, OpenAI-compatible endpoints).

The cost config field in opencode.json model definitions has zero effect — $ Spent stays at $0.00 regardless of token usage. This affects a huge number of users outside the Anthropic/OpenAI ecosystem.

Real-world impact: Users in China using GLM-5.2, DeepSeek V4, MiMo, etc. via @ai-sdk/openai-compatible cannot see any cost tracking. Related issues: #24113, #17223.

The fix looks solid with test coverage. Could the maintainers prioritize reviewing and merging this? Many users are resorting to third-party plugins (opencode-throughput, opencode-tokenscope) as workarounds because the built-in cost tracking is completely broken for custom providers.

Everyone in here who is not a hallucinating bot, raise your hands. ✋

+1

real fix here #36903

This adds nothing, closing