proposal: a durable model path for Computer (gateway free tier vs paid credits vs OpenRouter vs Groq)

[EthanThatOneKid]

Why this exists

#77 was the incident: a revoked per-provider key took Computer offline and every dispatch posted the provider's error into a thread. #78 moved the runtime onto the Vercel AI Gateway, so provider choice and retry left this repo, and #82 stopped failure comments from echoing provider text.

What #78 did not settle is which credential and which model the deployment should keep running on. The restore on 2026-09-23 used the gateway's free tier, which is a stopgap. This issue is the follow-up that was missing: no issue tracked "replace the direct DeepSeek dependency with a free-token path", only the outage did.

Verified on 2026-09-23 (gateway key already stored on the dev host)

Model slug	Result	Notes
deepseek/deepseek-v4.1-flash (the pinned slug)	HTTP 403	"Free tier users do not have access to this model." Needs purchased credits.
google/gemini-2.5-flash	HTTP 200	Free-tier eligible. Tags: tool-use, structured-output, vision, reasoning, 1M context. A single-tool probe returned the expected tool_calls with correct arguments.
deepseek/deepseek-r1	HTTP 200	Free-tier eligible, tool-use tagged. A single-tool probe returned prose and no tool call, and it is a reasoning model, so it is the weaker pick for a tool-calling pipeline.

Production therefore runs COMPUTER_MODEL=google/gemini-2.5-flash with AI_GATEWAY_API_KEY set on Production and Preview. The credential's AI Gateway free tier tags list is short — five free-tagged slugs, none of them a general-purpose tool-caller — so "free" is not a durable answer on its own.

Options with the evidence we have

Path	Reaches the pinned DeepSeek slug	Cost	New external account
AI Gateway, paid credits	yes	$0.15/M input, $0.60/M output	no
AI Gateway, free tier	no (403)	free	no
DeepSeek direct	yes	per DeepSeek pricing	yes — a replacement key; the current one is revoked (401)
OpenRouter	no DeepSeek among the 21 :free slugs (ling-3.0-flash, nex-n2.5, qwen3.8-27b, nemotron, inkling, laguna, gemma-4, north-mini-code)	free	yes; a GOOP_OPENROUTER_API_KEY exists on the host but belongs to Goop, not Computer
Groq	no	free tier	yes; no key on the host today

Recommendation

Keep the gateway lane and buy a small AI Gateway credit balance so deepseek/deepseek-v4.1-flash resolves again. One credential then covers all six roles and there is exactly one failure mode to watch, with the free-tier COMPUTER_MODEL override as an already-proven fallback if credits run out.

Credential requirement, for the record

AI_GATEWAY_API_KEY (or Vercel's injected VERCEL_OIDC_TOKEN) is now a deployment prerequisite: without one of the two the runtime cannot reach any model, and preflight says so by name rather than returning a bare status. Whoever deploys Computer must have a gateway credential available to the project. The key in use is the account-level gateway key already stored on the Zo dev host.

Open question for a human

Buy credits for the gateway, or mint a replacement DeepSeek key and keep the free-tier gateway model as the fallback? Both restore the pinned model; the first keeps one credential, the second keeps the direct provider we just removed.

[EthanThatOneKid]

Closing as superseded by #94 and #96 (2026-09-26).

The question this issue asks no longer applies to Computer's brain. It assumes the deployment owns a model path: pick a credential, pick a slug, keep it alive. #94 took that away — the Discord brain is the Zo persona 5f58a6ba-da81-4b8e-9105-4c85685a6a93 (model zo:deepseek/deepseek-v4.1-flash, paid from Zo credits), and #95 moves the web surface onto the same persona. #96 then retires the eve engine along with COMPUTER_MODEL and AI_GATEWAY_API_KEY. The durable model path is the persona's model, and there is no provider credential in the repository or on any deployment to rotate, expire, or revoke.

So the open question is answered by not choosing: no gateway credit purchase, and no replacement DeepSeek key. Both options existed to keep a repository-side model path alive, and that path is being deleted.

What was worth keeping from this issue, and stays on the record:

• The outage class is a per-provider key held by the deployment — a revoked key took Computer offline and every dispatch posted the provider's error into the thread. Removing the key from the deployment removes the class, which is exactly what refactor(discord): move Computer's brain onto the Zo host #94 did.
• The gateway's free tier cannot reach the pinned slug (403 on deepseek/deepseek-v4.1-flash), and only a handful of free-tagged slugs exist, none a general-purpose tool-caller. "Free" was never a durable answer for a repository-side path.
• google/gemini-2.5-flash was the free-tier stopgap that kept production answering between 2026-09-23 and the refactor(discord): move Computer's brain onto the Zo host #94 move.

Reopen or re-file if the persona's model ever moves off Zo credits to a per-provider key — that is the same class of dependency, just stored in a different place.