go: single account-wide usage pool contradicts documented per-model limits #52962

Summary

OpenCode Go appears to enforce a single account-wide usage pool shared across all paid models, rather than the independent per-model allowances the documentation describes. After exhausting the Kimi K3 allowance, every other paid model — including models with zero recorded usage — returns Go usage limit exceeded. Total paid spend at that point was $4.38.

Environment

• opencode version: 2.0.21 (latest)
• OS: Linux 7.2.8-1-cachyos x86_64
• Terminal: TERM_PROGRAM=Orca, TERM=xterm-256color, COLORTERM=truecolor
• Shell: /bin/fish
• Install/channel: latest
• Active plugins: ~/.config/opencode/plugins/orca-opencode2-status.js
• Plan: Go ($10/month)

Reproduction

1. Subscribe to the Go plan ($10/month).
2. Use kimi-k3 until requests begin failing with Go usage limit exceeded.
3. Issue the same request against deepseek-v4.1-flash — which has consumed $0.29 of its documented $60 monthly limit (0.49%).
4. Issue requests against grok-4.7, glm-5.3, qwen3.8-max, and gpt-6-luna — models that have never been used on this account at all.
5. Observe: all five paid models return the identical Go usage limit exceeded error.
6. Issue requests against space-bunny-free or longcat-2.5-preview-free.
7. Observe: both succeed.

Expected Behavior

Per the Go documentation, each model carries its own monthly limit (Kimi K3 $15, DeepSeek V4.1 Flash $60, GLM-5.3-Flash $60, MiniMax M3 $60, etc.), each with sub-caps of "5-hour — 20% of the monthly limit; weekly — 50%; and monthly — 100%":

Each model has the following usage limits: 5-hour — 20% of the monthly limit; weekly — 50%; and monthly — 100%.

Each model's monthly limit below determines how its usage counts toward those allowances.

I read this as independent per-model budgets. On that reading, exhausting Kimi K3 should disable only Kimi K3, while $60-limit models retain a nearly untouched allowance. Models never invoked at all should be fully available.

Actual Behavior

All paid models are gated by one shared pool. A model with zero usage cannot have exhausted its own allowance, which indicates the caps are not independent per model.

Measured usage at the time of failure (token counts from local session DB, valued at published Go pricing):

Exact error returned by each:

Go usage limit exceeded

All activity occurred 2026-10-01 21:05 → 2026-10-03 12:00 UTC.

Additional Context

• What I was trying to do: fall back to a cheaper model once Kimi K3 was rate-limited.
• What happened: the fallback was impossible; every paid model was simultaneously unavailable, leaving only the limited-time free models (space-bunny-free, longcat-2.5-preview-free) as a workaround.
• Frequency: 100% reproducible across 6 consecutive models.
• Workarounds tried: free models (work, but are documented as "limited time" and may be withdrawn).
• The apparent effective ceiling is roughly $4–5 of paid-model spend per rolling 5-hour window, pooled across all models. If that is the intended behavior for the $10 plan, the "monthly limit" column in the docs table does not communicate it — the table implies a total included value of roughly $1,000+ per month across the listed models, whereas observed behavior suggests a pooled ceiling an order of magnitude smaller.

Questions:

1. Are Go's limits intended to be per-model or pooled per account? If pooled, what is the actual aggregate ceiling for the $10 Go plan?
2. If pooled, can the docs be corrected to state the real aggregate allowance, so the per-model table is not read as independent budgets?
3. Should models with genuinely unused per-model allowances remain available as a fallback? If not, is that documented anywhere?