This is a COPY for the operator's reading room. The original (above) is the authority — if they ever disagree, the original wins and this file is regenerated. Generated 2026-10-12 18:27 by mira.

Opencode GO and ZEN Models

The sheet Young asked for on 2026-10-11: track which models Brain and Heart use, with prices and allowance limits, so a model choice is a lookup and not a re-derivation.

How this file is sourced, because a price sheet that is wrong is worse than none:

1. The live catalogue — what actually exists

Probed 2026-10-11 against both endpoints. The headline: GO is the small pool.

GO's 45 models are mostly things that will block. A $10 package against per-model ceilings is not headroom — it is a menu of things that will run out. ZEN's 12 free models are the real capacity, and they cost nothing.

GO free (2) — probed

ZEN free (12) — probed

Seven of ZEN's twelve also exist on GO. Where a model is on both, use the GO copy — usage inside the $10 package costs nothing extra; ZEN drains the $40 float.

2. GO — paid models with allowance limit

Prices and limits are Young's own figures, quoted 2026-10-10 (opencode.ai/docs/go/#usage-limits). Each model's limit is independent, and the 5-hour window is the binding constraint — 20% of monthly, not 100%.

The two things this table decides

1. Qwen3.8 Flash took $11.17 last week — and it was for nothing. Our own opencode stats --days 7: free models did 5.07b tokens for $0.00 while paid did 573m tokens for $18.14. Free was 8.8× the tokens at zero cost. Head to head on the single largest line: space-bunny-free did 2.0b tokens for $0.00 while qwen3.8-flash did 309m tokens for $11.17 — 6.5× fewer tokens, paid for. $11.17 against a $6.00 / 5-hour window means one good afternoon blocks the model.

2. Luna and Haiku at $3.00 / 5 hours are 4× tighter. Both are "complex task" candidates in the plan. Against a $3 five-hour ceiling they are fragile as fallbacks — they are escalation-only, never a default.

Prompt caching is the cheaper lever, and nothing used it

GLM-5.3-Flash caches reads at $0.03 vs $0.15 input — a 5× reduction. Our fleet is already at 91.6% cached input. A caching strategy buys more headroom than any model switch on this table.

3. ZEN — paid models enabled on the workspace

From Young, 2026-10-10. ZEN is $40/month pay-as-you-go credit — a float that drains, not an allowance. ZEN paid is the LAST rung, the only one that costs real money.

Rung order, now live in both switches:

GO free  →  ZEN free  →  GO paid  →  ZEN paid

Cost rung beats instrument rung. A free model on ZEN is still $0.00; a paid model on GO still costs money. The instrument-major order (GO paid before ZEN free) called qwen3.8-flash on attempt 3 while a free model sat untried until attempt 13.

4. Who is on what — Brain and Heart, measured

Brain — runs on Hermes, switches via fallback_providers

hermes fallback list, 2026-10-11, identical on all three profiles:

Four models, all free, zero paid in the chain. And it matches the V3 ladder's rung order exactly — reached independently, which is the evidence the doctrine is right.

Heart — runs on SolOSV3, switches via router.route()

cortex/registry/agents.yaml, all 11 agents:

Every agent on the one model with 769 successes behind it. Deliberate for a build phase; the ladder handles switching.

🔴 And the live defect this sheet records

core/config/models.yaml declares space-bunny-free under GO. It is not served by GO — it lives on ZEN. Measured:

go/space-bunny-free → "ModelError: Model space-bunny-free is not supported"
MISS lab-go/space-bunny-free | NOT served by go (45 models listed)

This is why the sheet exists. The registry had a model on the wrong endpoint, and nobody could see it until a probe asked the endpoint instead of reading config. Brain moved off it on 2026-10-11 — all three primaries are now longcat.

5. Selection rules — the reason each model is listed

Rule 6 in force: Qwen3.8 Flash at $11.17 has no recorded reason. The ledger now refuses such a record at construction, and audits history for rows predating the guard.

6. What this sheet cannot tell you

• Prices are Young's figures, not my probe. I verified which models exist and where, not their prices. No endpoint publishes a price list, so a price here is a quoted figure, not a measurement. The same is true of the monthly limits and the 5-hour window percentages.
• Free-model quality is unmeasured. I rank by cost and proven volume, never by quality. V2's 769-success figure proves longcat works; it does not prove it is best.
• The 5-hour percentages are from the settlement docs, not observed. V2's breaker measured real 429 provider.quota events — but against the hard configuration. Our guard assumes the soft one. That is settled by the console's "Extra usage" setting, which is still an open question.
• Nothing here is refreshed automatically. The catalogue probe (core/config/models.py::probe) can answer this file's tables on demand — this sheet is a snapshot from 2026-10-11 and it will drift.

The honest use of this document: it answers "which model should Brain or Heart use, and what happens when it blocks" on the day it was written. For anything time-sensitive, run the probe.