CX Simulator
Know the token bill before the job runs.
Pre-run token estimation for agent jobs — inputs counted exactly per model, outputs predicted from measured fleet history, so spend is decided in advance instead of read off a bill afterwards.
Decide the spend before you spend it.
What it does for you
Exact input counting, per model
Tokenizers differ: the same document is 15,509 tokens to Gemini and 14,566 to GPT-4o (~6.5% apart). CX Simulator counts with each model's real tokenizer — Google's official LocalTokenizer for Gemini, tiktoken for OpenAI — never a proxy count dressed up as exact.
Component vectors, never one number
Fresh input, cache creation, cache read, output, and reasoning stay separate end to end. The median fleet job is cache-dominated, and cache reads are priced and behaved differently from fresh input — summing everything into one total destroys the economics the estimate exists to show.
Calibration labels you can audit
Every prediction carries its evidence class — measured, calibrated, borrowed, or declared estimate — and its sample size. Small samples are gated: under 5 observations a cell is suppressed, 5–9 is exploratory, 10–29 provisional, 30+ calibration-eligible. Nothing is dressed up past its evidence.
Pool-true costs
Free credits, plan-covered usage, metered spend, and customer-borne keys are costed as separate pools with separate labels. Only metered spend is ever called an actual cost; a subscription's API-equivalent value is shown alongside, never summed into it.
Capabilities
- Per-model input counting (Gemini LocalTokenizer, tiktoken, Claude API count)
- Component token vectors — fresh / cache-create / cache-read / output / reasoning
- Retrospective distributions from real fleet job history, versioned with provenance
- Sample-size gates — suppressed, exploratory, provisional, calibration-eligible
- Funding-pool cost display — actual vs API-equivalent, never summed
- Cross-model comparison of the same payload in one run
- Borrowed priors (Hugging Face corpora) pinned by revision, licence-carrying
Proof, not adjectives
- Built on 38 real jobs
- The v1 distributions come from 38 completed CAMP fleet jobs with full token components — real history, labelled completed-only, with its success bias stated rather than hidden.
- Tokenizer-exact where it counts
- Gemini inputs are counted with the model's real SentencePiece asset (SHA-256 recorded per run); the Claude adapter refuses to substitute a proxy tokenizer when the real API counter is unavailable.
- Registered L1 tool
- Registered in the Creytix naming SSOT on 2026-10-23 (id cx-simulator): one deterministic job, invoked headless, receipted — it serves the execution layer's planning before anything runs.
- Savings settle honestly
- With-vs-without comparisons settle as their own records with the baseline source named — negative savings settle negative, unedited, and estimate classes never mix with measured ones.
Related tools
Built with Creytix, on Creytix
This page itself was produced by the platform it describes — the same design tokens, module pipeline, and review gates used for every Creytix client site. Nothing here is a mockup of the product; it is the product.