Solo developers do not leave because price is unclear. They leave because they cannot tell if cheaper output is good enough — or if the endpoint will hold under real CLI load. This page publishes the operational facts. The eval tool shows the quality diffs.
A page that is always green is worse than none. Components below are marked for what is live today versus what opens with the first cohort.
Operational
Waitlist API
Cohort closed
Inference endpoint
Available
Pair-diff eval
Last checked client-side against waitlist health. Inference status updates when the first GPU cohort goes live — we will not fake uptime.
User-facing CLIs die on the tail, not the mean. Figures below are design targets for the opening cohort (shared RTX 5090 class, ~8 concurrent seats). We replace targets with measured p50/p95/p99 after seven days of production traffic.
What we publish next: a rolling 7-day histogram for time-to-first-token and end-to-end completion, plus temperature-0 run-to-run variance as a measured distribution — not a single score.
Seat billing removes the token meter. Concurrency still exists because cards are finite. Forecast the seat, not the token.
| Limit | Opening cohort | Notes |
|---|---|---|
| Seats per shared card | ~8 concurrent | Soft concurrency. Dedicated cards available for larger teams. |
| Requests per seat | Unlimited under contract | No weekly token cap. Abuse / DoS ceilings still apply. |
| Spend predictability | Fixed $/seat/month | No per-request invoice. Usage export available for capacity planning. |
| Retries | Idempotent where the client sends an idempotency key | Safe retry without duplicate side effects on tool calls when supported. |
OpenAI- and Anthropic-shaped endpoints rarely hold past simple chat. Below is the honest list for Claude Code / Codex style clients — pairs, not marketing claims. Ranked by how often it breaks real workflows.
Strict schema adherence; rare invented parameter names on mature models.
Open-weight models can invent or omit tool args. We publish failing traces in the pair-diff eval. Prefer JSON-mode prompts until parity is measured on your suite.
SSE event shapes and stop reasons match vendor docs.
Chat completions streaming works for Claude Code connect. Anthropic-native message stream fields may differ; we document event mapping per client.
200k+ context with stable instruction following deep in the window.
Context window follows the chosen open-weight model. Quality drop-off is visible in pair-diff — keep frontier models for the hard slice of work.
Image inputs in the same CLI session.
Text-first for v1 seats. Multimodal is a separate model pick, not silently claimed.
Prompts go to GPUs rented for your team. We do not train on your traffic. We do not keep prompt or completion bodies after the request completes.
| Data | Retention | Purpose |
|---|---|---|
| Prompt & completion bodies | Zero — not retained | Inference only. Dropped after response. |
| Request metadata | 30 days | Latency, status codes, token counts for capacity and billing integrity. |
| Waitlist email | Until you ask to delete, or cohort ends | Contract outreach in list order. |
| Training on customer traffic | Never | Not used to improve models. |