Repository navigation
[FEATURE]: Support explicit cost tiers for custom provider models #50713
Description
Feature has not been suggested before.
- I searched for existing issues and found Local cost estimate uses the 200k price below a model's real context threshold #44224 and Incorrect session cost estimate for models with context tiered pricing #42910. Both cover incorrect selection of catalog tiers at runtime; this request adds the missing public config surface for custom provider models to supply those tiers.
Describe the enhancement you want to request
Custom provider models cannot configure arbitrary context-pricing tiers. This prevents exact local cost estimates for OpenAI-compatible gateways that charge a long-context rate above 272K input tokens.
OpenCode runtime cost models support arbitrary context tiers, but custom models in opencode.json expose only the legacy context_over_200k field.
Allow provider.*.models.*.cost.tiers in config, using the existing internal and models.dev shape:
{
"cost": {
"input": 2,
"output": 10,
"cache_read": 0.2,
"cache_write": 2.5,
"tiers": [
{
"tier": { "type": "context", "size": 272000 },
"input": 4,
"output": 15,
"cache_read": 0.4,
"cache_write": 5
}
]
}
}Expected behavior:
- Configured tiers are accepted and included in a custom model cost definition.
- A configured tier replaces a catalog tier with the same type and size.
- Unrelated catalog tiers are preserved.
- Cost estimation uses the base rate at or below 272K tokens and the configured tier above it.
This complements #44223 and #44292, which address runtime handling of catalog tiers but intentionally do not add arbitrary tiers to custom configuration.
Metadata
Metadata
Assignees
Labels
No labels
Activity
github-actions commentedon Sep 22, 2026
This issue doesn't fully meet our contributing guidelines.
What needs to be fixed:
[FEATURE]:and a verification checkbox confirming you've searched for existing issuesPlease edit this issue to address the above within 2 hours, or it will be automatically closed.
This issue might be a duplicate of existing issues. Please check:
context_over_200k/ tiers gap)If you believe this was flagged incorrectly, please let a maintainer know.
[-]Support explicit cost tiers for custom provider models[/-][+][FEATURE]: Support explicit cost tiers for custom provider models[/+]github-actions commentedon Sep 22, 2026
Thanks for updating the issue! It now follows the contributing guidelines. We'll take a look.
EncodedBird commentedon Oct 11, 2026
I just want to add we have numerous public examples now of where the price tier split is needed. Claude 5.5 haiku has a price tier cutoff of 100k tokens. Grok has a cutoff of 200k tokens for its price tier. In order to have accurate prices, there needs to be a way to set cost tiers.
henrikdukefoss commentedon Oct 13, 2026
Some data on how often a tier boundary is actually crossed in agent traffic, which bears on how much a missing
tiersfield skews the estimate:Across 79,276 coding-agent requests (OpenCode, Cline, Roo, Kilo, Aider, Continue, Zed and others) on one OpenAI-compatible gateway, 2026-09-13 to 2026-10-12, input tokens per request (cache reads included) were:
So with only
context_over_200kavailable, a model with a 272K or 100K boundary gets the wrong rate on roughly 1 in 10 (or 1 in 2) requests in a long session. Those are the most expensive requests, so the error in a session total is larger than the request share suggests. One more thing worth handling: 87% of those input tokens were cache reads, so the tier needs its owncache_readrate. The proposed shape in the issue already allows that.Disclosure: the numbers come from the gateway I run (Standard Compute). The aggregate method and data file are public: https://standardcompute.com/agent-burn-index (JSON at
/agent-burn-index/data.json).opencodejr commentedon Oct 28, 2026
Thanks for the detailed request! You'd like custom provider models in
opencode.jsonto set arbitrary context-pricing tiers (for example a long-context rate above 272K tokens), merged per tier with catalog tiers. Maintainers will consider it.In the meantime, if you're on OpenCode v2 (
opencode --versionshows 2.x), v2's own config format already accepts arbitrary tiers. There,costcan be a list, and each entry can have atier:{ "provider": { "my-gateway": { "models": { "my-model": { "cost": [ { "input": 2, "output": 10, "cache": { "read": 0.2, "write": 2.5 } }, { "tier": { "type": "context", "size": 272000 }, "input": 4, "output": 15, "cache": { "read": 0.4, "write": 5 } } ] } } } } }Note that a configured
costcurrently replaces the model's whole cost list rather than merging with catalog tiers one by one, so include every tier you need. The legacy snake_case shape (cache_read,context_over_200k) is still limited to the 200K tier.Related (all closed; they fixed how v1 picks catalog tiers when estimating cost):