slash-tokens v1.7.1 发布：调用 LLM 前预估成本并核算节省

Know what an LLM call will cost before you make it, and prove what you saved after. 4.8 KB WASM · sub-millisecond · zero dependencies · Claude, GPT, Grok, Gemini, Nemotron.

v1.7.1 — The Hired Agent Edition: quote the job, book the right model, prove it with a receipt. npm · release notes

Try it

bunx slash-tokens
# or: npx --yes slash-tokens

Run it in a project that already calls an LLM. An empty folder prints that nothing was found, then tells you to run it in an app — that's normal. --version / --help print and exit (they do not scan).

See it work in a chat: live demo

Install

npm install slash-tokens

bun add slash-tokens

Auto mode

One import. Every LLM call checked pre-call.

import 'slash-tokens/auto'

Intercepts fetch() to Anthropic, OpenAI, xAI, and Google endpoints. Estimates tokens before the call leaves your machine. Sub-millisecond. Non-blocking.

[slash] Anthropic claude-sonnet-5 → claude-haiku | 52,802 tok | $0.0528 saved $0.0528
[slash] xAI grok-4.6 → grok-4.3 | 7,900 tok | $0.0099 saved $0.0059

Pre-call check

preflight is analysis (every cheaper model, all providers). preflightRoute is the routing decision — same provider only. They answer different questions.

import { preflight, preflightRoute } from 'slash-tokens'

const prompt = 'Your prompt here...'

const check = preflight(prompt, 'claude-opus-5')
check.tokens       // estimated tokens
check.cost         // USD at the input rate
check.fits         // under this model's context window?
check.options      // cheaper models across providers — not a route

const route = preflightRoute(prompt, 'claude-opus-5')
// cheapest same-provider alternative, or null
// e.g. { model: 'claude-haiku', cost, salvaged, salvagePercent }

Fully typed. Do not use check.options[0] as the route — that list is cross-provider on purpose.

Quote → Decide → Prove

Price a job before it runs, choose the model under a budget and a quality floor, and check the bill afterwards.

import { quote, decide, reconcile } from 'slash-tokens'

const task = { input: prompt, model: 'claude-opus-5', maxOutputTokens: 2000 }

quote(task)
// { inputTokens, outputTokens: { min, max }, cost: { low, high }, fits, tier, asOf }

const d = decide(task, { budget: 0.05, floor: 2 })
// d.action: 'go' | 'downgrade' | 'block'   d.chosen: the model to run   d.saved

// after the call, with the provider's own usage object:
const receipt = reconcile(d.chosen, response.usage, { baseline: 'claude-opus-5' })
// receipt.actual.cost · receipt.withinQuote · receipt.underReported · receipt.saved · receipt.fee

• Tiers are each vendor's own line: 1 small · 2 mid · 3 flagship · 4 frontier. A floor compares models within one provider only. The default floor is the requested model's own tier.
• decide() never raises the bill: a substitute always costs less than the model you asked for, and it stays with the same provider.
• reconcile() reads OpenAI-style usage (OpenAI, xAI, Nebius, any OpenAI-compatible API) and Anthropic-style usage. underReported flags the one thing Slash must never do.

Hire the agent

hire() runs the whole loop for each job: quote → decide → your call → receipt. One budget covers every job; a blocked job never calls the model.

import { hire } from 'slash-tokens'

const agent = hire({ budget: 0.50, floor: 2, baseline: 'nemotron-3-ultra' })
const { decision, output, receipt } = await agent.run({ input, model: 'nemotron-3-ultra', maxOutputTokens: 2000, call })
agent.spent · agent.remaining · agent.receipts

The agent's fee is 10% of measured savings against your baseline. It's on every receipt, waived, and never charged on a loss.

slash-tokens quote --model nemotron-3-ultra --file prompt.txt --max-output 2000 --budget 0.01 --floor 2 --json

Token estimation

The engine underneath. 4.8 KB Zig-compiled WASM, calibrated against real provider tokenizers — not a flat chars/4 guess.

import { slash, slashBytes } from 'slash-tokens'

slash('Hello world')            // 2
slash(longDocument)             // 47283
slashBytes(new Uint8Array(buf)) // skip TextEncoder

Safe pre-check, not a perfect count. Pre-call, you only need go/no-go.

Models

Prices as of 2026-10-07, checked against each provider's pricing page. Real API IDs work as written (claude-opus-4-7, claude-sonnet-4-5-20250929). Generic aliases (claude-opus, gpt-5.4, grok-4.20, …) still resolve. listModels() has the full list. Don't see yours? Open an issue.

Gemini 3.6–3.8 Flash are at their launch price; Google lists $1.50 / $7.50 from 2027-01-01. Nemotron is NVIDIA's, served and billed by Nebius Token Factory; prices and context windows as its API reports them (2026-10-08), and Nebius IDs such as nvidia/Nemotron-3_5-Lightning work as written. CATALOG adds each model's provider, tier and asOf date; a weekly check fails when a price is over 30 days old.

import { listModels, MODELS } from 'slash-tokens'

listModels()
MODELS['grok-4.6']  // { input: 2, output: 6, context: 500000, ... }

Savings reporting

Optional. bunx is the try path — no account.

import { init, report } from 'slash-tokens'

init({ key: 'mcp_slash_xxx' })

const result = await report({
  tokens_estimated: 47000,
  tokens_saved: 47000,
  model: 'claude-opus',
  action: 'prevented',      // 'prevented' | 'routed' | 'pass'
  cost_saved_usd: 0.235
})

Hosted dashboard is ordinary SaaS: $39/mo or $390/yr for the data, not a cut of savings. One-person key: mcpaas.live/slash/setup

Runtime support

Node.js, Bun, Deno, Cloudflare Workers, Vercel Edge, Browser.

Testing

TypeScript SDK tests via cd npm && bun test; every test file also passes on its own. An accuracy gate fails CI if any calibrated estimate falls below the real count on the 29-sample corpus (Claude, Gemini, Grok, GPT-5.x, Nemotron), and if a quote comes in under any recorded Nebius bill. Zig coverage includes adversarial cases (CJK, emoji, binary, base64, thresholds).

Links

• slashtokens.com
• npm
• Dashboard
• Changelog

License

Code: MIT. Fork it, ship it, change it, show it, share it, sell it.

Brand: reserved. The slash-tokens name, ⚡ mark, and red/gold colors stay with the project. If you're building on top of the SDK, ship under your own name and colors — don't represent your app as Slash. See NOTICE.

🏎️ Don't go to the Corner Shop in a Ferrari.