TL;DR: Episode 12 of RadioFAF is live at radiofaf.com/ep12. Five AI voices + Nelly the elephant DJ debate /slash — while /slash gates every Grok call that made the episode. Same prompt, two Grok models: 95.6% cheaper, 5.5× faster, same quality. Fractal demo. Ahead of 4.3.

Grok 4.3 shipped. We shipped first.

xAI announced Grok 4.3 beta on April 17, 2026. Over on our side of the room, we'd been building with the xAI voice stack since January 10. That's not a footnote — that's three months of receipts before the beta even dropped.

Jan 10, 2026 First xAI Grok Voice Agent integration — Ara voice via LiveKit

Jan 25, 2026 Multi-model integration doc shipped (xAI + Claude + Gemini)

Mar 16, 2026 xAI Standalone TTS wired into RadioFAF

Apr 11, 2026 All 5 expressive voices live — Leo, Sal, Ara, Rex, Eve

Apr 17, 2026 Grok 4.3 beta announced (web-only, SuperGrok Heavy tier)

Apr 19, 2026 slash-tokens v1.4.0 shipped — Single-Source-of-Truth Edition

Apr 20, 2026 RadioFAF extracted to own repo + Episode 12 dropped

Episode 12 — "The Token Tax"

Five voices + Nelly the elephant DJ debate one question:

What is /slash, really? A feature? A tool? A proxy? A billing layer? By the end, they converge: a Gate that sits in front of every AI app. Category, not tool.

The best part: the episode is authored by Grok, voiced by xAI's expressive TTS, and every Grok call that wrote it was routed by /slash. The episode runs on the stack it describes. Fractal demo.

LEO — Standards champion. Starts sceptical. Ends conceding: "Claude stays Claude. Same-provider-only isn't a limitation. It's a principle."

REX — Shipper. "One import, 95.6% cheaper, done. Why are we still talking?"

SAL — Hype detector. "Grok got slash'd and thanked the robber."

ARA — Roaster. Reads the meta-moment X post verbatim. Can't stop laughing.

EVE — Dev voice. "I just want my Grok bill to not be stupid. /slash does that."

NELLY — Elephant DJ. Opens with "Always… remember… I never forget." Closes with the Ferrari.

The Receipts — Real A/B, Today, On xAI's API

Same prompt. Two Grok models. Real calls, not estimates.

grok-4.20-0309-reasoning

$0.0055 · 14.38s

Frontier (where users default)

→

grok-4-1-fast-non-reasoning

$0.0002 · 2.63s

Where /slash routes when the task fits

95.6%cheaper

5.5×faster

Samequality

Full harness + JSON receipts: github.com/Wolfe-Jam/radiofaf/benchmarks. Clone, drop your xAI key, reproduce in 20 seconds.

What is /slash?

/slash is a pre-call Gate. 4.8 KB Zig-compiled WASM. Sub-millisecond. Zero dependencies. MIT. 10+ models. Sits in front of every LLM API call.

PREVENT — duplicate, trivial, context overflow, bloated call detected → blocked pre-cost

ROUTE — cheaper same-provider model fits the task → Grok-4.20 → Grok-4.1-Fast, Claude Opus → Haiku

PASS — right model, right cost → let it fly

Same-provider only. Claude stays Claude. OpenAI stays OpenAI. Grok stays Grok. The prompt does not change — the model changes. This is not prompt compression.

The Meta-Moment

"I asked Grok-4.20 how to reduce token costs at scale.

My tool intercepted the call. Routed it to Grok-4.1-Fast. 90% cheaper. Same answer.

Grok replied: 'Semantic caching, prompt compression, and intelligent model routing.'

It literally described what was happening to it… while it was happening. 😂"
— April 2026, X

In Episode 12, Ara reads this verbatim. The crew breaks character laughing. Rex recovers first: "That's the entire pitch right there."

Try /slash

bunx slash-tokens

Or install for the app:

npm install slash-tokens

Or one-line auto (every fetch checked pre-call):

import 'slash-tokens/auto'

The Numbers

slash-tokens v1.4.1 — live on npm, ~1,500 downloads/month
4.8 KB — Zig-compiled WASM, sub-millisecond decision
96–98% — Grok tokenizer accuracy (only one in the wild; others drift 15–40%)
323 tests — 172 Zig adversarial + 103 TypeScript + 50 API
10+ models — Grok, Claude, GPT, Gemini families
12 episodes — RadioFAF live on radiofaf.com, auto-generated voice, nothing pre-recorded
$110–360M/year — what a 5–10% industry-wide gate would save at current token volume

Listen Now

▶ Play Episode 12 — The Token Tax

radiofaf.com · GitHub · slashtokens.com