That error means OpenCode's free models stopped answering API calls from outside the OpenCode app. In September 2026 the Zen gateway's free tier became app-only: the models still exist, still cost $0, but only respond when the request comes from OpenCode itself. Your key is not broken, and your config is not wrong. You have three ways out, and one of them takes about two minutes.
We reproduced the exact error today, October 9, 2026, with an anonymous call to the Zen endpoint, no key attached:
$ curl -X POST https://opencode.ai/zen/v1/chat/completions \
-H 'Content-Type: application/json' \
-d '{"model":"big-pickle","messages":[{"role":"user","content":"hi"}]}'
{"type":"error","error":{"type":"FreeTierError",
"message":"OpenCode's free tier can only be used from within OpenCode"}}
HTTP 403, error type FreeTierError. The model catalog itself (https://opencode.ai/zen/v1/models) is still publicly readable, which is why scripts that only list models keep working while actual calls fail.
Disclosure: Wallaby Token sells API access to Kimi K3, one of the models discussed below. The error reproduction above involves none of our infrastructure.
What changed
OpenCode Zen is the curated model gateway built into OpenCode. During its public feedback period, the free models on Zen answered any caller, key or no key, and a fair number of scripts, proxies, and third-party tools came to rely on that unauthenticated endpoint. In September 2026, OpenCode closed the free tier to external traffic. Public reports of the change appeared within days, and the error message above is what those callers started seeing.
The free models themselves did not disappear. As of October 9, 2026, the Zen docs list thirteen free entries, including Big Pickle, MiMo-V2.5 Free, and Nemotron 3 Ultra Free. What changed is the gate: free models answer only inside the OpenCode app. Paid models on Zen still work through the documented API endpoints with a Zen key, billed per request from a prepaid balance.
Option 1: stay inside the app, keep paying $0
If you only ever used the free models for coding inside OpenCode, nothing broke. Open the TUI, pick a free model, keep going.
Two honest caveats, both from Zen's own documentation:
- The free lineup rotates. Every free model is marked "available for a limited time," and unannounced "stealth" models come and go without notice. We watched this ourselves: in mid-September we tracked a stealth model called Union Alpha in the Zen catalog, and it is already gone. Treat any specific free model as temporary.
- Most free models train on your data. Zen's privacy section says that during the free period, prompts and completions may be used to improve the model. The exceptions it names with a zero-retention policy are Space Bunny Free, LongCat 2.5 Preview Free, and Step 5 Preview Free. If you point a free model at a private repository, read that section first.
Option 2: BYOK, bring a model that answers everywhere
OpenCode supports any OpenAI-compatible provider natively. If what you actually want is a strong open model callable from OpenCode and from your own scripts, CI, or other agents, bring your own key.
The least-friction path we know of, because we run it: Kimi K3 is in the models.dev catalog that OpenCode syncs, with Wallaby as a listed provider. Setup is a refresh and an environment variable:
opencode models --refresh
export WALLABY_API_KEY=<your key>
opencode -m wallaby/moonshotai/kimi-k3
No hand-written provider block, no base URL to paste. If the model does not appear after the refresh, restart the TUI once. We documented the full setup step by step in Kimi K3 in OpenCode.
On price: Zen sells Kimi K3 at the official list price, $3.00 per million input tokens and $15.00 per million output. Wallaby sells the same model at $2.70 / $13.50, 10% below official list, with cached input at $0.27. A new Wallaby account starts with $0.50 of trial credit, no credit card required, and top-ups start at $20. Zen, by comparison, requires a minimum prepaid top-up of $20. Rates are as of October 9, 2026; our free trial page has the current details.
Option 3: other free API lanes
If the goal is strictly "a free model I can call from anywhere," several providers run genuinely free API endpoints, including NVIDIA's hosted free tier, which currently serves Kimi K3. Free lanes come with the usual trade-offs: rate limits, queues, and data policies worth reading — do not build production workloads on unguaranteed free lanes. We maintain a dated comparison in Cheapest way to run Kimi K3 API, free options included, rather than duplicating it here.
Which one is yours
- You code inside OpenCode, cost is the only concern, and your repos are not sensitive: Option 1, eyes open about rotation and data use.
- You want one model callable from your editor, your terminal, and your own tooling: Option 2. This is the group the September change actually broke.
- You want free API calls at any cost: Option 3, and read the rate limits before you build on them.
FAQ
Is OpenCode Zen free?
The Zen gateway itself is pay-as-you-go from a prepaid balance. Zen also carries a rotating set of free models, but since September 2026 those only answer inside the OpenCode app. The OpenCode software itself remains free and open source either way.
Can I use OpenCode's free models via the API?
No. External calls to free models return HTTP 403 with a FreeTierError, which we reproduced on October 9, 2026. Listing the catalog still works; chat completions do not. Paid Zen models do work via the documented endpoints with a Zen key.
What is the cheapest way to run Kimi K3 in OpenCode?
Bring your own key. Kimi K3 is in OpenCode's synced model catalog via models.dev, so configuration is one refresh and one environment variable. Wallaby sells Kimi K3 at $2.70 per million input tokens, 10% below the official list price that Zen charges, with $0.50 of free trial credit and no card required.
Sources
- OpenCode Zen documentation, model catalog, pricing table, and privacy section:
https://opencode.ai/docs/zen/(accessed October 9, 2026) - Our own reproduction of the 403
FreeTierErroragainsthttps://opencode.ai/zen/v1/chat/completions, October 9, 2026, quoted verbatim above - Public reports of the September 2026 free-tier restriction, including user posts and third-party write-ups from September 10-17, 2026