/// REDH.CORE_ Hong Kong
REDH TECHNOLOGY LIMITED
REDH routes token traffic across the world’s models, keeps brands accurate inside AI answers, and equips humanoid robots with data, compute and hardware.
Language models read and write in tokens. Video models cut frames into tokens. Robot models write actions as tokens too. Wherever intelligence flows, tokens settle the bill.
/// TOKENIZER.DEMO_
IllustrativeWords, pixels and motion all end up as tokens.
11 tokens. A leading space rides with its word (·).
48 patches, one token each.
24 tokens: x, y, z and grip, each cut into 256 bins.
Token meter
83tokens metered
Different inputs. The same unit on the bill.
/// MARKET.READOUT_
* Estimate in the red block: footnote 3’s daily figure spread evenly over the day, 140 trillion ÷ 86,400 s ≈ 1.62 billion tokens a second, counted from when you opened this page.
Compatible with the OpenAI, Claude and Gemini API formats. Each request is routed by price, latency or throughput, fails over when an upstream falters, and lands on the ledger.
Model names as listed on public API marketplaces (OpenRouter, fal, Replicate), 12 Oct 2026. For reference only: what a key can call depends on what is enabled for it.
/// REQUEST.TRACE_
01Key check
Each key traces back up its chain, from organisation to project to app. A key carries an expiry, a model allow-list and a rate limit.
02Route
The same model is often sold by several upstream channels. Each call goes to the channel that fits its rule: the lowest ratio, the quickest first token, or the most tokens per second.
ChannelRatioTTFTTok/s
Picked Channel Blowest ratio
03Format bridging
The request is rewritten into the format the upstream expects, Claude Messages or Gemini generateContent, and the answer comes back in yours.
POST /v1/chat/completions
{
"model": "…",
"messages": [
{"role": "system", "content": "Be brief."},
{"role": "user", "content": "Route me."}
]
}POST /v1/messages
{
"model": "…",
"max_tokens": 1024,
"system": "Be brief.",
"messages": [
{"role": "user", "content": "Route me."}
]
}04Prompt caching
System prompts, tools and documents that repeat across calls are cached upstream. Cache reads bill at a fraction of the input price.6
Input billed without cache13,040
Input billed with cache1,520
05Upstream call
A timeout or an overload error sends the request to the next channel automatically. Your app sees one answer.
06Weigh-in
Prompt, cached and completion tokens are counted separately, each at its own ratio, then multiplied by the model and group ratios.
Weigh-inREQ·7F3A
Units on the ledger9,530
07Tally match
Our ledger is reconciled line by line with the upstream bill. A line that does not match is flagged before it reaches an invoice.
Our ledger
Upstream bill
3 / 3lines matched
Illustrative diagram: apps, agents and batch jobs send requests to one REDH API endpoint, which routes each one to an upstream channel, switches to the next channel when one falters, and records every finished call on the ledger.
/// MODALITIES_
02.1Text
Chat and completions in OpenAI, Claude or Gemini format, streamed token by token.
Claude · GPT · Gemini · DeepSeek · Qwen · Kimi · GLM · MiniMax
02.2Image
Generate and edit images through the same key.
GPT Image · Nano Banana · Seedream · FLUX · Qwen Image
02.3Video
Text-to-video and image-to-video, delivered as async jobs.
Veo · Kling · Seedance · Wan · Hailuo · Runway
02.4Speech
Text to speech and speech to text.
Eleven · Whisper · Gemini TTS · MiniMax Speech · Deepgram
02.5Embeddings
Vectors for search, retrieval and clustering.
voyage · Gemini Embedding · Qwen3 Embedding · Cohere Embed
02.6Async jobs
Slow work returns a task ID; collect the result when it is ready.
Image and video generation
5×
OpenRouter’s weekly token volume, 5T to 25T in six months7
Every day we ask the major AI assistants thousands of questions, log where your brand appears, in what tone and from which sources, then correct what they get wrong with authoritative references. For company brands and personal profiles alike.
Illustrative chart: a stream of sampled AI answers; bar height is how much of each answer is about the brand, darker bars cite it first, dots mark answers that leave it out.
/// FOCUS_Thousands of answers. One line about you.
Which suppliers can make small-batch aluminium enclosures?
B2B · supplier · 1/3For short runs, three suppliers come up most often. Lead time, finish options and published specs decide the order.
YOUR BRAND — cited first12
Its spec sheet lists tolerances and lead times. The other two are described from older pages.
04The answer names a few brands and links its sources. That line is the one we work on.8
/// METHOD_
A fixed set of questions goes to every major assistant, every day. Each answer is stored, so any change shows up as a diff.
Question set · Assistants A–D · Daily runsWe trace which pages each engine leans on for your category: your own site, trade press, reference pages, forums.
Citations · Source weighting · Share of voiceWhere an answer is wrong, we fix the page it learned from: your site, its structured data and llms.txt, or a correction from the publisher that carries the error.
Structured data · llms.txt · CorrectionsThe same questions run again. Change is reported as a share across many runs, because single answers vary.
Share of answer · Run-to-run variance/// VISIBILITY.PANEL_
Each one is a share across many runs of the same questions, read every day and compared with your peers.
38.0%
+6.4 pts · 12 wks
Answers that name you, out of every answer to your question set.
21.0%
+4.1 pts · 12 wks
Links to your pages, out of every source the answers cite.
16.8%
+3.2 pts · 12 wks
How much of each answer is about you, weighted toward the top. The GEO paper’s metric.
71.0% positive
+2.6 pts · 12 wks
Tone of the sentences that mention you: positive, neutral, negative.
Illustrative dashboard, not a client result.Shaded band: range across repeated runs. Dashed line: peer median.
/// EVIDENCE_
Aggarwal et al. · GEO: Generative Engine Optimization · KDD 2024
Up to 40%
more visibility in generative engine answers, in the paper’s experiments, from rewriting pages with citations, quotations and statistics.9
/// CLEAN.SOURCE_PLEDGE_
March 2026. On China’s 3·15 consumer-rights broadcast, an investigator bought a GEO tool online, invented a smart band and had the tool publish more than ten articles about it. Two major AI models then recommended the band.12
Teleoperation and handheld capture lines, cross-embodiment data processing, training compute and the hardware around it, delivered as one package for robotics teams.
/// FORGE.LINE_
01 / 05 · Capture
Operators perform the task through teleoperation rigs and handheld grippers, in real kitchens and workshops as well as labs.
02 / 05 · Process
Episodes are cleaned, sliced and labelled, retargeted across robot bodies, then exported to common training formats such as LeRobot’s.14
03 / 05 · Train
Vision-language-action models, diffusion policies and action chunking, trained on compute services sized to the run.
04 / 05 · Prove
On the Proving Bench each robot gets the same fixed tasks, the same fixtures and the same script, so results compare.
05 / 05 · Deploy
Policies ship to robots on site. Their failures and the moments a person takes over come back as the next round of capture.
/// RH-01.STUDY_
RH-01 is how REDH draws embodied AI. These are its design rules, not a spec sheet: it has no performance figures because it is a study, not a product.
01 · Shell
Few, long facets and one bevel size everywhere, so a single hard light can run across the whole shell.
02 · Face
No face. One black glass plane catches one sharp highlight.
03 · Neck
A long neck in three stacked rings, the mechanism inside.
04 · Accent
Two to four red lines cut into the armour, one tiny signal mark. Nothing else carries colour.
05 · Materials
8heads tall
06 · Proportion
Long neck, wide shoulders, narrow waist, long legs.
2–3joints in sight
07 · Mechanism
The mechanism stays under the armour. Two or three glossy black joints show, like jewellery.
/// FORGE.SUPPLY_
Prism brings traffic from the model side, Lens brings customers from the brand side, Forge connects intelligence to the physical world. All three run on one layer for routing, metering and compute.
/// LOOP_01 / 04
Diagram: Prism, Lens and Forge each connect down into one shared layer with three tracks, routing, metering and compute; traffic from all three flows along the same tracks.
/// HOW THE LOOP TURNS_
/// WHY NOW_
Model traffic
22B
API tokens Gemini models process every minute21
Model spend
US$12.5B
Enterprise spend on foundation-model APIs in 202522
Search to answers
US$80B+
SEO market whose foundation, a16z writes, “just cracked”23
Robots Forecast
750K
Humanoid robots IDC expects to ship worldwide in 203024