@barnaby_lloyd Thank you 🙏
To respond to on the layer underneath the user-facing controls — for the reason that charge tale is partially about routing, however most commonly about what we run in the back of the routing.
Aggregated LLM credit score pool. Karma negotiates quantity charges throughout Anthropic / OpenAI / Gemini / Qwen / DeepSeek and swimming pools utilization throughout our person base. That is 20–40% higher unit economics than somebody
person will get on direct API. Variance will get absorbed through the pool — you do not really feel the spikes, and the power unit remains strong throughout seller pricing adjustments.
Aggregated cloud execution pool. Sandbox runtimes (E2B), VPS-class compute, GPU assets for symbol/video — all warm-pooled with second-scale provisioning. No cold-start tax billed for your process, no per-VM
condominium to your account. When an avatar enthusiasts out to a sandbox or spins up an extended container, the capability is already there.
That is what we could routing keep quality-neutral. If we had been a skinny passthrough, each and every “Sonnet no longer Opus” name could be a quality-vs-cost tradeoff for the person. Since the backend is pooled, routing optimizes
in opposition to pool economics — you continue to get the fitting type for the duty; the financial savings come from infrastructure leverage, no longer from downgrading your output.
