TokenZero vs Braintrust
Braintrust is an evaluation platform with datasets, scorers, logging and a caching proxy. The matrix below says what each product is and how it bills; the only numbers are TokenZero's own measurements and prices.
TokenZero and Braintrust differ in kind: Braintrust is evals and logging with an AI proxy; TokenZero is a BYOK gateway whose engines reduce the provider bill and sign a receipt per call. Measured overhead is 1.52 ms p50; the cascade holds a 99.8% parity floor.
· Source: Braintrust documentation at https://www.braintrust.dev/pricing read 2026-10-26; TokenZero overhead from /benchmarks; prices from table 2026-10-26
At a glance
- Compared with
- Braintrust (evals and logging with an AI proxy)
- TokenZero overhead
- 1.52 ms p50 / 3.42 ms p95 (measured, /benchmarks)
- BYOK
- TokenZero: always · Braintrust: yes (proxy)
- Receipts
- TokenZero: signed per call · Braintrust: none
- Billing
- TokenZero: 15% of verified savings · Braintrust: free tier, then usage plans
- Reference saving (gpt-6.1-sol, 100k calls)
- $285/month from prefix caching (computed)
- Net after 15% fee
- $242/month (computed)
- Parity floor
- 99.8%
- Free tier
- $500/month routed
- Vendor page read
- 2026-10-26
TokenZero vs Braintrust: what each one is (vendor page read 2026-10-26)
| Dimension | TokenZero | Braintrust |
|---|---|---|
| Category | BYOK savings gateway | evals and logging with an AI proxy |
| Hosting | hosted | hosted; hybrid deployment |
| Bring your own key | always; keys pass through and are never stored | yes (proxy) |
| Open source | CLI, MCP server, audit core and GitHub Action Apache-2.0; gateway hosted | proxy open source |
| How it bills | free up to $500/month routed; then 15% of verified savings, 3x-ROI-or-$0, capped plans | free tier, then usage plans |
| Caching | prompt-cache formatting + exact and semantic response cache | proxy cache |
| Routing | cascade to the cheaper sibling inside one provider, 99.8% parity floor, advisory by default | none |
| Savings receipts | Ed25519-signed per call, hash-chained, public key published | none |
| Gateway overhead | 1.52 ms p50 / 3.42 ms p95 (measured, /benchmarks) | not measured by TokenZero |
| Data retention | metadata only; ZDR mode available | see vendor docs |
| Net saving, 100k gpt-6.1-sol calls/month (prefix caching only) | $242 after fee | — |
What you keep when you move from Braintrust
evals and datasets; only the proxy base URL changes.
- Shadow parity is an in-production eval of cascade routing: a sample of routed requests is re-run on the original model and a pair is auto-disabled when parity drops.
- Receipts, not logs: metadata only, signed.
The change
OPENAI_BASE_URL=https://tokenzero.sh/v1
ANTHROPIC_BASE_URL=https://tokenzero.sh
# keep your provider keys exactly as they areThen npx -y @tokenzero/cli doctor to confirm the gateway, signer and pricing table, and see the switch guide.
Questions people ask
What is the difference between TokenZero and Braintrust?
Braintrust is an evaluation platform with datasets, scorers, logging and a caching proxy. TokenZero is a BYOK gateway: you keep evals and datasets; only the proxy base URL changes, and every call comes back with a signed receipt. Shadow parity is an in-production eval of cascade routing: a sample of routed requests is re-run on the original model and a pair is auto-disabled when parity drops.
What does switching from Braintrust to TokenZero save?
It depends on your prompts; on the reference workload (100,000 calls of 2,000 input with a 1,500-token static prefix and 300 output on gpt-6.1-sol, table 2026-10-26) prefix caching alone cuts $700 to $415 a month, $285 saved and $242 net after the 15% fee; exact-cache hits and cascade routing add to that. The receipts show the real figure.
Can I keep Braintrust and add TokenZero?
Yes. Braintrust is not a transport-level proxy, so it keeps working unchanged; TokenZero changes only the base URL your SDK calls.
How is the 15% fee computed?
On verified savings only: list-price baseline minus provider-reported actual cost, per signed receipt. A cycle whose savings are below 3x the fee is billed $0; Pro Cap and Scale Cap add monthly ceilings. The first $500 of routed spend each month is free.