$29/ mo
The same models, a smaller allowance.Enough to put one real project through the loop and judge it for yourself.
Pricing
And after the year?One plan, every model.
Each plan buys about as much use as the Claude Code or Codex subscription at the same price. The difference is what it reaches: all three agents, and every mainstream model they can call, on one plan instead of one vendor's.
$89/ mo
As much use as a $100/mo Claude Code or Codex plan.Where most people settle: agents running most of the day, a few sessions open at once.
$199/ mo
As much use as a $200/mo Claude Code or Codex plan.For days that don't end, with headroom left when a build runs long.
- Claude Code, Codex and Mirasim's own agent, all on one plan
- Every model they can reach, including the open Chinese ones
- Billed in credits, at a published rate per model you can check before you spend
- No five-hour window; the allowance runs on a weekly cycle
One thing it does not do
This works inside Mirasim, and only inside Mirasim.
It is not someone else's subscription handed over to you. You cannot use it in their web app, their phone app, or your own scripts. If you need an account that works everywhere, buy theirs; this will not replace it.
No substitutions
You get the model you picked.
- We do not move you onto a cheaper model without telling you, and we do not lower quality when load is high.
- When a turn runs on the quota fallback instead of your own subscription, the transcript marks that turn. A session cannot change route without you seeing it.
Verify it yourself
Don't take our word for it. Fingerprint the model.
Every relay says it isn't substituting models, which is why saying it proves nothing. Here is how to check, and what we commit to not doing while you check.
Run it side by side
Send the same prompts through us and through a subscription you pay for directly, then compare the output. Same model, same settings, same seed where the model accepts one.
Use the published probes
Model families reveal themselves on inputs that have nothing to do with quality: tokenizer boundaries, Unicode edge cases, refusal wording, formatting habits. Public probe sets test exactly that. Point one at us.
Watch the route, turn by turn
The transcript marks any turn that ran on the quota fallback rather than your own subscription. When you audit, that tells you which turns to compare and which to set aside.
We don't watch you do it
We do not detect probe traffic, flag it, or route it differently from ordinary work. A test request and a real one take the same path.