Does OpenCode (OpenCode Models gateway (formerly OpenCode Zen)) train on your code?
Anomaly (OpenCode) · OpenCode Models gateway (formerly OpenCode Zen) · checked 2026-11-02
Paid models are ZDR at the provider; free ones may train. OpenCode's own policy is vague.
Trains on code
Not published
Default
OpenCode Models (the gateway formerly called OpenCode Zen; renamed in OpenCode's docs as of Nov 2, 2026) providers follow zero-retention and no-training, EXCEPT listed free models (as of Nov 2, 2026: Big Pickle, Flint 40B Free, MiMo-V2.6-Flash Free, plus NVIDIA free endpoints such as Nemotron 3 Ultra Free that log use to improve NVIDIA products) whose data 'may be used to improve the model'. OpenCode's own use of prompts isn't stated; its privacy policy lists prompts as collected and 'Improving the Services' as a purpose.
How to opt out
Use paid OpenCode Models (formerly Zen) models, not the free ones listed as exceptions. To hide free models entirely, set OPENCODE_DISABLE_FREE_MODELS=1 (documented on the Zen and Models docs pages).
Retention
Providers: zero retention, except some free models (e.g. Flint 40B Free and Beam Free: 30 days) and OpenAI/Anthropic APIs (30 days). OpenCode's own retention: 'as long as necessary'.
Telemetry
Not published.
Disable telemetry
Not published.
Zero data retention
partial · Provider-side ZDR on most models.
Offline / local model
no · Hosted gateway.
Self-hostable
no · n/a
Score
19/100 · training 8/40 · retention 6/25 · offline 0/20 · telemetry 5/15
Classes: training Not published · retention Not published · telemetry Not published · offline no
Not published (not found in the vendor's official sources)
OpenCode's own training use of prompts, telemetry_sent
Notes
The opencode.ai/zen landing page now says you can choose whether any model can train on your data, but OpenCode's docs don't describe that setting yet, so it isn't scored.