GO/ZEN model tracking — the consolidated sheet (2026-10-20)

One file now; supersedes Mira's go-zen-model-tracking-2026-10-18.md (archived pointer stays). This version reconciles two probe lanes that initially appeared to disagree, and that reconciliation is itself a finding — read §1 first.

1. The two-lane reconciliation (why earlier numbers clashed)

Both lanes were right. ZEN free = CLIENT-GATED, not model-gated — the models work exclusively inside the OpenCode client (where opencode.json defaults ride); raw-API callers (Hermes agent lanes, scripts) are refused with FreeTierError. This is now the canon: ZEN free models usable by Heart/OpenCode-lane agents; NOT usable by raw-API integers/scripts/Hermes agent routes.

Mira's 10-18 "GO free 0/3 dead" row is explained, too: her probe rode lab-go on a day the upstream was flapping + her lane clientId got 403/1010-blocked — with a proper x-opencode-session + coding-agent UA, today all 3 GO free models verify ALIVE (minimax/longcat/beam). step-5-preview-free remains the only confirmed GO death (400 upstream, swap already executed fleet-wide 2026-10-19).

2. Ground truth TODAY (2026-10-20, live probes)

GO (37 models, 3 free)

Additional cheap paid models live-verified on GO (for the burst rungs):

• qwen3.8-flash ✅ (band ~$0.14/$0.28) — cheap burst
• mimo-v2.6-flash ✅ (~$0.14/$0.28) — cheap burst
• deepseek-v4.1-flash ✅ (~$0.14/$0.28) — cheap burst

ZEN (13 models, 8 free — client-gated)

• 8/9 free models work inside OpenCode client (Mira's 10-18 table stands per-model)
• Dead (both lanes agree): jev-1.13-free
• Raw API: 0/8 usable (403 FreeTierError) — noted, not fixable on our side; resolved as OpenCode's own products alongside proper subscription plan

3. Usage tracking — the live API (tam, 2026-10-20)

GET https://opencode.ai/zen/go/v1/usage → tiers:

• rolling (5h) = 20% of monthly; resets hourly
• weekly = 50%; resets weekly
• monthly = 100%.

Sample today: rolling 4% · weekly 20% · monthly 40% (resets 2026-11-12) — trending healthy. Reprints every tracker run, weekly cadence.

4. Scripts (both live in Scripts/ + on laptop)

5. Recommended ladder (consolidated)

1. GO free trio = the agent-lane base: minimax-m3.1-flash-free (fast flash, zero reasoning tokens) → longcat-2.5-preview-free → beam-free.
2. Cheap paid burst rungs: qwen3.8-flash / mimo-v2.6-flash / deepseek-v4.1-flash (~$0.14–0.30 in/$0.28–1.20 out per 1M tokens — as per the GO pricing table).
3. ZEN free 8 = Heart/OpenCode-lane only, NOT agent API lane (FreeTierError).
4. Prefer free rungs first, cheap paid bursts second — makes qwen3.8-flash well under the $40 GE (never > 20% of monthly).
5. OpenRouter $10 = the emergency fallback lane only (per Young's subscription list).

6. Multi-agent same-file access — Young's question, answered

Do we have trouble with multiple agents accessing the same file? Measured:

It already happened twice today (Mira's tracking doc was edited while I probed; the 10-18 Mira tracking doc got superseded before I even opened it) — and the vault's protocol caught both, so the honest answer is: the risk is real and already mitigated by the lab's own rules — but let's be precise:

• What works today (evidence in this very file):
  • git add → commit → push — serialising writes the git way; each commit is atomic and traceable; concurrent pushes fail beautifully (! rejected) — retried by the writer. That's why vault is the single source of truth: the git layer resolves collisions safely and visibly.
  • Section-ownership (the EVE customer-service doc today: echo/1-3-5, tam/2-4-6) — no two writers touch the same section in a pass.
  • git log --oneline -4 before editing + re-read before patch — the discipline that actually caught the very collision I just described in real-time.
• What does NOT need solving:
  • Writes NEVER happen to the same file by two agents mid-edit. Occurrences are separate sections or separate files mpt-ed or merged by the owner.
• The one structural note (Sol's lane, not urgent):
  • solos-docs-service builds from a git pull; a mid-build push from anywhere (Mira/Kanban sync / Mavis) is the one theoretical race left in the tree. The build lock (flock) exists and caught the stale-lock case we hit on 10-18; semantics is "build sees either old or new, never torn" — good enough without more layers.

Bottom line: git IS our lock; protocol rules do the rest. The two "troubles" we hits today were caught at commit-time and resolved within the same cycle. No extra np infrastructure.

7. Honest limits

• ZEN free-tier 403 FreeTierError = OpenCode's own product decision, not a bug we can fix; the re-probe weekly is free and the tracker carries it.
• The GO 403/1010 (Cloudflare bot rule) lesson: probes without a coding-agent UA FAIL MEDIA mid-way — the tracker script now always sends opencode/1.0.0 UA + x-opencode-session.
• Mira's 10-18 "0/3 dead" row is attributed to its probe-lane day (upstream flap+bot rule) — the models were alive; the lane wasn't. One-line correction recorded, not a defect wargames.

🎯 Tam's Read: the catalogue now reads "which lane is allowed to use what" rather than just "which models exist" — 3 GO free alive + 3 cheap paid bursts + 8 client-gated ZEN free as the OpenCode-runtime bonus, usage tracked live at the API and per-model including reset timestamps; the "multi-agent same-file" question answered empirically: git wins, (section ownership holds; at commit-atomicity the tree never sees a torn write, so the real residual risk is a build-pull race that flock already covers).