[Zen Free Tier] Free models connection frequently unstable (upstream failures/timeouts); request stable endpoints + more strong coding models

Problem

As a user of the OpenCode Zen free tier, I'm experiencing frequent, intermittent connection failures with the free models on a daily basis (Windows 11, Desktop v1.18.15 and recent CLI builds):

• Requests randomly fail with connection-related errors: Upstream request failed, connection resets, TLS/handshake timeouts, and frequent network stalls that appear provider-side rather than local — the same requests succeed on retry seconds later, then fail again minutes later.
• The failures are not model-specific: deepseek-v4-flash-free, longcat-2.0-free, nemotron-3-ultra-free, and others all exhibit the same flakiness pattern at different times of day. Retrying usually works, but mid-coding-session drops are disruptive.
• Errors appear both as visible stream error notices and silent stalls (no output for 60s+ until timeout), making work unreliable without a fallback.

(This complements but is separate from the existing reports: #37270 "model not supported on full model list", #38028 hy3-free/nemotron failing, #36506 paid models failing — my report here is specifically about connection stability for free models, not model-list metadata or paid tiers.)

Request

1. Make the free-tier connection stable: add retry/backoff/reconnect at the gateway level, improve load balancing, and reduce connection drops for the -free models (at least deepseek-v4-flash-free, currently the best free option for coding).
2. Provide a stable/fallback base URL for free models — e.g. a documented endpoint with better routing or automatic failover between entry points.
3. Add more free models with strong coding abilities — the free catalog is heavily skewed toward chat/agentic-general models. Users would benefit from more free options strong at coding tasks (e.g. a free Sonnet-class/GPT-class code model, or a coding-strong reasoning model with long context). Currently the reliable free coding option is effectively one or two models, which makes the connection instability even more painful.

Environment

• OS: Windows 11 x64
• App: OpenCode Desktop 1.18.15 (also reproducible in CLI)
• Provider: OpenCode Zen free tier
• Models tested: deepseek-v4-flash-free, nemotron-3-ultra-free, longcat-2.0-free, others

Related

• #37270, #38028, #40732 (free model failure reports), #36506 (paid failures)