Free models accessible through OpenCode — Console, OpenRouter, NVIDIA, and other configured providers — are convenient: you experiment without a credit card, parallelize without watching the meter, and switch model families with one line of configuration. The problem is that "free" is a temporary entitlement, not a property of the model.
Two free models with similar names can differ in privacy, regional availability, request limits, quality, and how long they will exist. One declares zero retention, another retains prompts for safety, a third uses them for model improvement. One is reachable from your account, another answers not available in your country. One disappears next month.
This post walks through a practical workflow: classify the work, read the catalog, probe reachability, classify privacy, assign roles with fallbacks, and record the choice so it can be re-verified.
The starting rule
Treat "free" as a temporary entitlement to verify per role and re-verify before every long or sensitive run. The catalog is a dated snapshot, not a contract.
Scope: what this workflow covers
We are talking about free models accessible through OpenCode, referenced as <provider>/<model-id>:
- OpenCode Console:
opencode/<model-id>, with an optional variant (e.g.opencode/deepseek-v4.1-flash-free#max). - OpenRouter:
openrouter/<model-id>(e.g.openrouter/qwen/qwen-2.5-coder-32b-instruct:free— prefer current coding-capable:freeendpoints with tool support; old 7B/8B instruct models often fail multi-step tool schemas). - NVIDIA:
nvidia/<model-id>(e.g.nvidia/nemotron-3-ultra). - Other providers: any provider configured in OpenCode that offers free models.
Paid Console models are out of scope. The catalog changes without notice. Documentation and runtime can both be stale. That is why every choice must be probed, not just read.
1. Classify the work before the model
Before opening the model picker, capture six facts. They narrow down the candidates, not the other way around.
- Confidentiality: can sources, prompts, or outputs leave the trust boundary? This decides whether logged or training models are allowed.
- Region: some models are geo-blocked. A catalog row does not prove your account can reach them.
- Volume: free tiers cap aggressively — often 20–40 RPM or tens of requests per day. See Rate limits by provider before planning fan-out.
- Role: architecture and review need reasoning; implementation needs code quality and throughput; exploration needs latency.
- Independence: when a genuine second opinion matters, the reviewer should not share the implementer's model family.
- Lifetime: free models are time-limited. Plan a fallback in the same privacy class.
Example: reviewing a concurrent interface on proprietary code calls for zero retention + strong reasoning + an independent reviewer. High-volume exploration over synthetic code can settle for a fast model in the trial class.
2. Read the current catalog
One runtime, several documentations — each with a different job:
- OpenCode Console: opencode.ai/docs/console/models/ lists the free set, variants, pricing, privacy notes, and deprecations. Use it for contractual terms.
- OpenRouter: openrouter.ai/models?max_price=0 filters free models. Check each model's Context and Architecture tabs plus the upstream provider's policy for privacy and retention.
- NVIDIA: build.nvidia.com/explore/discover lists Nemotron and other models; free tiers are marked. Review NVIDIA's data usage policy per model.
- Runtime:
opencode models, or/modelsin the TUI, lists what this account can actually select across all configured providers. Use it for entitlement.
Reference a model as <provider>/<model-id>, such as opencode/deepseek-v4.1-flash-free#max or openrouter/qwen/qwen-2.5-coder-32b-instruct:free. The #variant suffix is OpenCode 2 syntax — on 1.x use --variant instead.
When docs and runtime disagree, trust the docs for terms and the runtime for entitlement. In both cases, probe before trusting either.
For quality, use an independent tie-breaker: LMArena (formerly LMSYS Chatbot Arena) provides Elo ratings — prefer the Coding leaderboard for agentic work, and complement it with tool-assisted benchmarks like SWE-bench or LiveCodeBench where available. It only makes sense at equal privacy class and reachability — a higher Elo never compensates for unacceptable retention.
3. Probe reachability
Catalog status is not reachability. Probe the exact provider, model, and variant. Examples target OpenCode 2 — on 1.x replace #max with --variant max. --archived is recent; upgrade if your CLI reports an unknown-flag error.
# OpenCode Console, with variant
opencode run --archived --model opencode/deepseek-v4.1-flash-free#max "Reply with exactly: OK"
# OpenRouter (:free has no variant — tool support varies by upstream model)
opencode run --archived --model openrouter/qwen/qwen-2.5-coder-32b-instruct:free "Reply with exactly: OK"
# NVIDIA
opencode run --archived --model nvidia/nemotron-3-ultra "Reply with exactly: OK"
# Tool-use probe for Implementer/Explorer candidates (chat OK ≠ agent OK)
opencode run --archived --model opencode/glm-5.3-flash-free#max "Read package.json and list its dependencies without editing anything"
A plain OK ping only proves chat completion. Free endpoints — especially via OpenRouter — can pass it and still fail the first real agent turn with a 400 or malformed output when the upstream model lacks tool/function-calling support. Probe at least one read-only tool call for any role that will use tools.
--archived keeps probes out of normal history. To probe a set of models, automate the same command in a script that iterates over the candidates for your provider list.
| Result | Meaning | Action |
|---|---|---|
prints OK |
reachable | keep it |
Model unavailable |
not in the runtime catalog or not entitled | replace it |
not available in your country |
geo-block | replace it; do not retry |
| timeout | endpoint or network stall | retry once, then replace |
| 429 or quota text | rate or daily cap | move the role or reduce fan-out; do not retry in burst — honor reset / Retry-After |
A model can show as active in the catalog and still fail. Always record the probe date.
4. Classify privacy and retention
Before using a free model on non-public code, read the current privacy and retention policy for each provider and model. Sort candidates into four classes:
| Class | Meaning | Typical use |
|---|---|---|
| Zero-retention, no training | provider states no retention and no training | confidential code, subject to policy |
| Retained, no training | short safety retention only | internal but not secret |
| May improve the model | inputs may be used for improvement | public or disposable code |
| Trial or training | logged trial data or training use | public or synthetic code only |
Never infer privacy from the name
Providers change terms, and free endpoints often differ from the paid model of the same family. Read the class from the current terms, not the name. "Zero retention" on a free promotional endpoint is a contractual promise, not an enterprise BAA/DPA — regulated teams often treat any non-contracted free tier as off-limits for crown-jewel IP regardless of retention claims.
Provider-specific notes:
- OpenCode Console: privacy class documented per model in the Console docs.
- OpenRouter: varies by upstream provider. Check the OpenRouter model card and the original provider's policy. Many free models log prompts for moderation.
- NVIDIA: Nemotron free endpoints are trials; data may be logged. Do not submit confidential code.
- Other providers: check each provider's terms; free tiers often have weaker privacy than paid tiers.
5. Assign roles and fallbacks
Think of the team as a pyramid: one architect at the apex making few high-reasoning calls, implementer and reviewer in the middle, and high-fan-out explorers at the base. "Pyramid suitability" in the next sections means which level — and which request volume — a provider can sustain without hitting its cap.
Different roles ask for different capabilities. A reasonable starting split:
| Role | Prefer | Why |
|---|---|---|
| Architect, planner | strongest reasoning, high variant | design quality drives everything downstream |
| Orchestrator | reliable instruction-follower, mid or high variant | coordinates, does not edit |
| Implementer | strong coding model, high variant | correctness and iteration speed |
| Reviewer | different family from the implementer, high variant | independent failure modes |
| Explorer | fast, low variant | high volume, low stakes |
At equal privacy class and reachability, use LMArena Elo for the role's primary capability: reasoning and instruction-following for architect/planner, coding Elo for the implementer (Coding leaderboard, plus SWE-bench / LiveCodeBench where available), a different family with the highest Elo for the reviewer, throughput for orchestrator/explorer.
Build a fallback chain inside the same privacy class. Prefer same-provider fallbacks for simpler config, cross-provider for independence. Never downgrade from a zero-retention model to a trial-logged one just to keep working.
| Privacy class | Example chain |
|---|---|
| Zero-retention | opencode/deepseek-v4.1-flash-free → opencode/glm-5.3-flash-free → opencode/minimax-m3.1-flash-free |
| May improve | openrouter/qwen/qwen-2.5-coder-32b-instruct:free → openrouter/meta-llama/llama-3.3-70b-instruct:free (filter for current coding-capable :free endpoints with tool support) |
| Trial / logged | nvidia/nemotron-3-ultra → opencode/nemotron-3.5-super-vl-free |
Watch out for variants: support differs per model. Use only variants advertised in the picker or provider metadata. A wrong variant fails at request time, not at config load. When a subagent pins model:, it overrides parent variant inheritance and silently falls back to the provider default — always pin the variant alongside the model.
6. Record and re-verify
Record provider, model ID, variant, role, privacy class, region result, probe date, rate-limit snapshot, and fallback in the project or agent config. opencode.json/jsonc and agent frontmatter do not validate custom fields — store probe metadata as comments or in a companion manifest next to the config.
- before a long or expensive run;
- after any
Model unavailable, geo-block, or quota error; - when provider documentation changes;
- when a free model is withdrawn or deprecated.
Minimal record per role: provider + model + variant, privacy class, reachability + date, RPM/RPD snapshot + quota check method, 429 strategy, pyramid suitability (e.g. OpenRouter :free for explorer only, NVIDIA for fallback/reviewer, Console zero-retention for architect/implementer), and the same-class fallback.
// opencode.jsonc — roles with probe/privacy recorded in comments
{
// architect: opencode/deepseek-v4.1-flash-free#max
// privacy: zero-retention | probe: 2026-11-17 OK + tool-read OK
// fallback (manual): opencode/glm-5.3-flash-free#max
"architect": "opencode/deepseek-v4.1-flash-free#max",
// reviewer: different family from implementer
// privacy: zero-retention | probe: 2026-11-17 OK
// fallback (manual): opencode/minimax-m3.1-flash-free
"reviewer": "opencode/glm-5.3-flash-free#max",
// explorer: low-volume only — OpenRouter :free, 20 RPM / small daily bucket
// privacy: may-improve (logged for moderation) | probe: 2026-11-17 OK
"explorer": "openrouter/qwen/qwen-2.5-coder-32b-instruct:free"
}
// Fallback is manual (or wrapper-scripted), not automatic:
// on 429/unavailable, switch the role string above to its
// same-privacy-class fallback. Exception: OpenRouter's own
// models array can route within OpenRouter.
Rate limits by provider
Free tiers fail on throughput before they fail on quality. The three providers cap differently, and the cap decides which roles each provider can hold.
| Provider | Free tier limits | Pyramid guidance |
|---|---|---|
| OpenCode Console | No published general RPM/RPD; only explicit cap is LongCat 2.5 Free: 100 req/day. Others are "free for a limited time" without declared caps. Shared per account/network/IP bucket, reset 00:00 UTC, dynamic with capacity vs demand. |
Good for architect/implementer on zero-retention models. On Free usage limit reached, move the role or reduce parallelism; do not retry in burst. |
| OpenRouter | :free models: 20 RPM always; 50 RPD (<10 lifetime credits) / 1,000 RPD (≥10 credits). Shared across all :free models, UTC day. Negative balance → 402. Upstream 429s apply independently. |
Explorer / low-volume only. Do not use for fan-out or architect with many subagents. |
| NVIDIA | ~40 RPM global, shared across all models at key level, not per model. No official daily. Requires Developer Program + phone verification, geo-blocked in ~10 countries. Trial only: no production, no confidential data. | Fallback / reviewer / secondary vision worker only. On 429 with Retry-After, back off and move sustained work to opencode/* zero-retention. |
Check quota before a large run:
# OpenRouter: free daily counter
curl -s https://openrouter.ai/api/v1/key \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
| jq '.data.free_model_daily_requests'
# OpenCode Console: no API — reset time is in the error message, reset 00:00 UTC
# NVIDIA: account page on build.nvidia.com — ~40 RPM shared, no official daily
429 handling differs per provider
OpenCode: wait for the reset time in the error, reduce parallelism. OpenRouter: honor X-RateLimit-Reset; if the source is upstream_provider_shared_pool, switch model or add a paid fallback. The free_model_daily_requests counter can lag ~1 minute in bursts, and a missing :free suffix can silently route to a paid endpoint — double-check the suffix before a large run. NVIDIA: honor Retry-After and move sustained parallelism to Console zero-retention.
Failure handling
- Model unavailable: switch the role to its recorded same-class fallback. Do not silently downgrade privacy.
- Geo-blocked: replace it. Retrying cannot change the region.
- Rate limited: treat as a routing signal, not a retry signal — move the role or narrow fan-out per the provider table above.
- Deprecated: migrate before the removal date; check the deprecation table for both the family and the exact version.
- Unexpected quality drop: re-check the selected variant and whether the free alias now points to a different backend.
Anti-patterns to avoid
- Treating a catalog row as proof of entitlement.
- Sending confidential code to trial-logged or training models — including NVIDIA trials.
- Assigning every role to a single provider family.
- Using a daily-capped model for fan-out work — notably OpenRouter
:freefor architect with many subagents. - Using NVIDIA trial endpoints for sustained parallelism or production.
- Hardcoding a free model with no fallback.
- Confusing a deprecated base model with its current free variant.
Further reading
- Console catalog: opencode.ai/docs/console/models/ — free set, variants, privacy, deprecations.
- OpenRouter free models: openrouter.ai/models?max_price=0 — filter
:free, check Context/Architecture tabs. - NVIDIA models: build.nvidia.com/explore/discover — free tiers marked, trial terms apply.
- Independent ranking: LMArena — Coding leaderboard for Elo tie-breaker.