free tool · pricing table 2026-10-26 · works without JavaScript

Prompt caching savings estimator

Paste a system prompt: exact token count, whether it clears the cache minimum, break-even reads and monthly dollars saved.

The sample support-agent prompt is 501 tokens on claude-sonnet-5-5, below the 512-token cache minimum; cached reads cost $0.100 instead of $2.00 per million, the write pays back after 2 reads, and at 5,000 calls a day the month drops from $720 to $720. Paste your own prompt to recompute.

As of 2026-10-26 · Source: pricing table 2026-10-26 and the gateway's own tokenizer and benchmark files; your inputs never leave the browser except token counts, which are not stored

Runs the four engines on a canned gpt-6.1-sol request and signs a real receipt. $0 provider spend.

At a glance

Model: claude-sonnet-5-5
Sample prompt tokens: 501 (computed)
Cache minimum: 512 tokens
Cached read / input price: $0.100 / $2.00 per 1M
Write surcharge: $2.500 per 1M
Break-even reads: 2 (computed)
Month at 5,000 calls/day: list / cached: $720 / $720 (computed)
Saved per month: $0 (computed)
Tokens by provider tokenizer: openai 344 · anthropic 501 · gemini 362 · xai 362 · deepseek 362 (computed)

Sample prompt, 5,000 calls per day, 400 user tokens, 300 output tokens, table 2026-10-26

Scenario	Per call	Per month	Delta
List price, prefix not cached	$0.00480	$720	—
Prefix served from the provider cache	$0.00480	$720	$0 (0%)

Estimator

System prompt (or paste a tool schema JSON)

You are the support assistant for Acme Analytics. Answer only questions about the Acme product, billing and account access. Be concise and factual; never invent features.

Product facts:
- Plans: Starter ($29/mo, 3 seats, 10k events/day), Team ($99/mo, 10 seats, 100k events/day), Business ($399/mo, 50 seats, 1M events/day). Annual billing takes 20% off.
- Data retention: 30 days on Starter, 13 months on Team, 25 months on Business.
- Integrations: Segment, Snowflake, BigQuery, Postgres, Webhooks, Slack, Zapier.
- SSO (SAML) is on Business only. Two-factor is available on every plan.
- Exports: CSV and Parquet from any saved report; API rate limit 600 requests per minute per workspace.

Support policy:
1. If the user reports a bug, collect the workspace id, the page URL and the time it happened, then file it with severity.
2. Refunds are handled by the billing team; collect the invoice number and hand off.
3. Never ask for passwords, API keys or card numbers.
4. For feature requests, thank the user and record the request with their plan.
5. If you are not sure, say so and offer to escalate to a human within 15 minutes.

Tone: friendly, direct, no marketing language. Use the customer's name once. Keep answers under 120 words unless the user asks for detail.

Escalation contacts: billing@acme.example, security@acme.example (for any suspected account compromise), status page at status.acme.example.

Model Calls per day User tokens per call Output tokens per call Public GitHub repo (optional)

audit a repository instead

Questions people ask

How are tokens counted?

With the gateway's own tokenizer per provider (the one the audit and the receipts use), via POST /v1/tools/tokens; the text is counted and discarded, never stored.

What is the break-even?

The number of cached reads after which the write surcharge is recovered: (write minus input) divided by (input minus cached), plus one. Where a provider charges no surcharge it is one read.

Does my prompt need to change?

No. The gateway places Anthropic cache_control breakpoints and OpenAI prompt_cache_key on the static prefix automatically; keep it byte-identical and first in the request.

What does TokenZero cost?

Free up to $500 per month of routed spend; then 15% of verified savings with a 3x-ROI-or-$0 rule. Gateway overhead is 1.52 ms p50, measured.

Prompt caching economics report Prompt cache miss playbook LLM cost calculator Cascade routing simulator LLM tokenizer and compressor GitHub repo token-waste auditor Gateway latency and TTFB benchmark OpenAI to Claude cost converter

All tools

Let the gateway place the breakpoints: swap base_url and every claude-sonnet-5-5 call caches its prefix automatically.Start free