One flat price. No 5-hour or weekly limits.

A stream is one request while it's generating. An agent working through a task uses one stream at a time; subagents, a second terminal or an editor assistant running at the same moment each use their own. One agent in one terminal needs one stream. Claude Code often starts subagents in parallel to explore a codebase, and with a single stream they wait for each other, so if you use subagents or run two sessions, pick a plan with 3 or more. Requests over your limit wait briefly in a queue rather than failing.

There's no per-token charge, no token quota, no 5-hour window and no weekly cap. Your plan sets how many requests generate at the same time, the context window per request and a per-request output cap. Everything you send and receive within that is included, and requests beyond your parallel streams wait briefly in a queue instead of failing.

No. Paid plans will never get a token quota, a 5-hour window or a weekly cap, and a subscription keeps its price and parallel streams for at least 12 months. After that, any change gets 30 days' notice by email. Read the promise.

It's Qwen3.8-27B, an open-weights model. On the coding benchmarks Qwen published it's close to Opus 4.6 Max: 73.0 on Terminal Bench 2.1 against 78.2, 42.3 on NL2Repo-Bench against 47.6. Frontier models still lead on the hardest, longest problems, and we serve a quantized build that can score a little lower, so try it on your own work with the free daily allowance before you pay. Benchmarks and speed.

The first token usually arrives in under a second for a short prompt, up to about 2.5 s when a server is full. A single stream generates at ~42 tok/s, and when a server is running 8 streams at once each one gets ~29–31 tok/s. For comparison, OpenRouter reports about 61 tok/s for Claude Opus on Anthropic, so it's slower than the big hosted APIs; what you get instead is no meter. Long prompts take longer to start because the whole prompt is read first: a 19K-token agent context takes about 20 seconds on its first turn. Follow-up turns reuse the cache, so only the new part of the prompt is read. (Measured September 2026 on our dual-GPU servers.)

Yes, and that's often the best setup. Keep your frontier subscription for the hardest problems, and point long-running work here: background agents, big refactors, test generation, and anything you'd otherwise ration. In Claude Code you can keep the settings in a project's .claude/settings.json or a shell profile you switch to, so each session uses whichever backend you choose.

If every GPU slot is busy, a new request waits briefly for the next free one instead of failing. The queue is fair: people who've used less recently go first, and waiting requests move up over time.

Yes. Every account can send a small number of requests a day for free once its email address is verified, and the allowance resets at 00:00 UTC. Free requests are shared per device and network, so one person gets one allowance. Paid plans have no daily request limit.

It's how much conversation a single request can carry: your input plus the model's answer, in tokens. Each plan lists its window. When a conversation grows past it, coding agents compact the history and carry on, and near the limit we shorten the answer instead of failing the request.

Claude Code, Codex CLI, OpenCode, Hermes Agent, Pi, Oh My Pi, Cline, Cursor, Continue, Aider, and anything built on the OpenAI or Anthropic SDKs. The API speaks OpenAI Chat Completions and Responses and Anthropic Messages on one base URL. See the setup guides.

Yes. Monthly plans cancel from the billing portal in your dashboard and keep working until the end of the period you've paid for. Passes don't renew; they simply end.

We never use your content to train models, and we keep only request metadata such as token counts, timings and status. How requests are processed is set out in the Privacy Policy.