Repository navigation
Go plan: $60/month quota exhausted in 6 days while the client recorded only $14.80 of usage — cache-read billing is invisible, undocumented, and the local cost meter is misleading users #41976
Description
Description
I am a paying Go plan subscriber ($10/mo). Within 6 days of my billing cycle (Aug 6 - Aug 12, 2026), both the dashboard and the opencode UI show my monthly quota ($60) as essentially exhausted - yet the usage recorded locally by the opencode client itself totals only $14.80. That is a 4x discrepancy between what the product tells me I am consuming and what is actually metered against my quota. I find this deeply misleading.
Reproduction steps
- Subscribe to the Go plan.
- Use opencode-go models (deepseek-v4-flash, kimi-k3) for 6 days - 1,791 assistant messages, 5.94M input / 778K output / 371M cache-read tokens (all recorded in the local opencode SQLite database).
- Compare the locally recorded cost with the Go quota dashboard percentages.
Expected behavior
The local cost field should match what is metered against the Go quota, or the metered figure should be transparent (per-request billing breakdown, published cache-read pricing).
Actual behavior
- Local recorded cost: $14.80; dashboard: monthly quota essentially exhausted.
- deepseek-v4-flash alone accumulated 359M cache-read tokens, recorded locally at $0.00.
- No warning before exhaustion; no per-request breakdown available to reconcile the 4x difference.
Impact
I paid for a month of usage and received 6 days of paid models; the remaining ~24 days of my paid cycle are unusable unless I fall back to free models.
Additional context
Local data from the opencode SQLite database (every assistant message this cycle):
| Model | Msgs | Input | Output | Cache read | Locally recorded cost |
|---|---|---|---|---|---|
| opencode-go/kimi-k3 | 238 | 1.87M | 121K | 12.6M | $13.24 |
| opencode-go/deepseek-v4-flash | 1550 | 4.04M | 656K | 359M | $1.52 |
| opencode-go/grok-4.5 / glm-5.2 | 3 | 25K | 351 | 16K | $0.05 |
| Total | 1791 | 5.94M | 778K | 371M | $14.80 |
If cache reads are metered anywhere near full input price, true consumption lands around $40-$70 - consistent with an exhausted $60 cap. The official Go docs state only the three rolling windows ($12/5h, $30/week, $60/month) and say nothing about cache token pricing.
Related reports with the same symptom (visible usage far below exhausted quota):
- Discrepancy between different opencode go usage dashboard #38255 - ~$34 visible usage, monthly quota 100% blocked
- OpenCode Go weekly/monthly quota does not match usage history #41206 - ~$11 visible usage, weekly quota 100%
- Go quota accounting does not match Usage History or documented limits #41391 - maintainer-opened question about Go quota accounting model
I paid for $60 of usage across a month. I got 6 days, with a meter that claimed I had spent a quarter of my quota while draining it entirely. Whether this is an accounting bug or intentional opacity, it needs an answer - I expect a written explanation, a per-request audit of my account, and compensation for the unusable remainder of my billing cycle.
Activity
github-actions commentedon Aug 12, 2026
This issue doesn't fully meet our contributing guidelines.
What needs to be fixed:
Please 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:
If you believe this was flagged incorrectly, please let a maintainer know.
github-actions commentedon Aug 12, 2026
Thank you for updating your issue! It now meets our contributing guidelines and has all the required information. We appreciate the detailed data you provided.
elrender commentedon Aug 14, 2026
Disclaimer: I'm not affiliated with Opencode in anyway, just your case made me curious as I wanted to test Opencode myself.
Looking at your table and official docs pricing: https://opencode.ai/docs/go/#usage-limits
it seems that all of those models have different limits. So I suspect that:
When summing those up for me you are very close to the monthly limit.
I do not know exact counting algorithm on the Opencode server's side -> there might be a discrepancy here that has to be checked, so the Opencode usage should show the same as GUI, but based on my calculations I wouldn't expect too much of a refund :)
Zaczero commentedon Aug 19, 2026
The $14.80 and the empty monthly bar are the same spend after Kimi’s Usage factor.
Quota is
visible × (60 / Usage)on one shared $60 monthly window, not a separate cap per model.359M Flash cache reads at the published $0.007/M are ~$2.50. They are not the missing $45, and the local $0.00 on that column is the client meter, not the quota.
Tongzii commentedon Aug 21, 2026
@Zaczero @elrender — thanks for the math, but it actually proves the opposite of what you intended.
You've confirmed that the "$14.80 locally vs $60 exhausted" gap is by design, stemming from a per-model multiplier the client never shows.
Here are my thoughts:
1. The disclosure is reactive patchwork, not transparency.
The
Usagecolumn and the "Why some models have lower usage" section you cited were last updated 2026-08-20 — eight days after I filed this issue (2026-08-12) and one day after @Zaczero reverse-engineered the formula. If this billing model had always existed, it should have been written into the pricing page and the client before charging money, not patched in after public complaints. This makes us question OpenCode's honesty toward its customers.2. OpenCode Go is a reseller, and the "6x generosity" is a half-truth.
OC is essentially a middleman/relay: it wholesale-purchases open-model compute and resells it under a $60/month cap. The "6x" is the retail margin OC keeps after negotiating bulk discounts — not a free gift. Packaging that margin as "we give you 6x" while quietly applying a 4x multiplier to Kimi-K3 and a 2x multiplier to DeepSeek-V4-Flash is exactly the kind of trust-eroding rhetoric. My $60 cap didn't buy me $60 of usage — it bought me ~$56 of "multiplier-adjusted usage," and the most-used model was squeezed down to 2x.
3. The client's own meter is still broken; those multipliers should be folded into the local count, or at least clearly disclosed.
Even accepting the multiplier, the opencode client records $0.00 for DeepSeek-V4-Flash's 359M cache-read tokens, while the docs price that cache at $0.007/M (~$2.50). The client's local cost display isn't just "different from the server" — it's factually wrong, leaving users with no way to reconcile their own spend. A paying user shouldn't have to rely on community members to reverse-engineer the server's accounting.
Zaczero commentedon Aug 21, 2026
I generally think (I'm unrelated to the OpenCode team in any way) that it's a genuine docs quality and transparency issue. I don't think they did it on purpose, but there are so many people getting caught on that. I'm wondering why it takes them so much time to address it, though.
Tongzii commentedon Aug 21, 2026
If you really have nothing to do with OC, then you should let the official personnel of OC explain. If you are confused about others spending time understanding token spending and transparency issues, I cannot explain others' thoughts to you. For me, this is a question of honesty
ZalmoraDev commentedon Aug 30, 2026
Reading https://opencode.ai/docs/go/ I also find it misleading how the time limit quotas are advertised, compared to how they are calculated. How I interpret the docs is having a total monthly access of $60 worth of model inference, with max monthly usage differing per model (e.g. DeepSeek V4 Pro capped at $15, MiMo V2.5 capped at $60).
So how I read it is still being able to use $45 dollars worth of model inference, after having depleted my DeepSeek V4 Pro $15 limit. But the total usage calculation is based on how much each model's usage percentage is summed together, which wouldn't reflect actual dollars spent against the advertised $60 cap. If this is unintended, then only the $60 monthly limit models are true to how I interpret the docs.
Some clarification from the OC team would be nice.
Tarquinen commentedon Oct 20, 2026
We're tracking this in #42985, so I'm closing this as a duplicate. Please follow that issue for updates.
— from 𝕺𝖕𝖊𝖓𝕮𝖔𝖉𝖊