Off-map
We ignore the crowded middle. The best problems live at the edges, where incumbents never think to look.
Artificial intelligence, off the map
We build
where no one
is looking.
Untethered is a collective making AI products and software for the edges of the map — problems too strange, too small or too early for anyone else.
Descend01 — Manifesto
Everyone is building the same ten things. The same chat window, the same copilot, the same dashboard with a sparkle icon. We are not interested in the obvious. We go where the maps run out — the quiet industries, the odd workflows, the problems nobody has named yet — and build the thing that should have existed there all along.
Quiet
We don't announce things that don't exist yet. You'll hear about it when it works — not before.
AI-native
Not features bolted on. Products that could only exist now that models can reason, see, and act.
Human
The models do the heavy lifting. People make the calls. That line stays exactly where it is.
Currently in the lab
Hover to request clearance
- P—01 Classified Sector Research Not yet.
- P—02 Classified Sector Prototype Soon.
- P—03 Classified Sector Stealth Patience.
- P—04 Classified Sector Unnamed Access denied.
02 — The shift
Software used to be a place you went. Now it's something that works for you.
Traditional SaaS sold seats, screens and workflows you had to learn. AI-native software sells outcomes. The difference isn't a feature — it's the entire business model, and it is being rewritten in real time.
#
Dimension
Traditional SaaS
AI-native
01
Pricing
Per seat, per month
Per outcome delivered
02
Interface
Dashboards, forms, filters
Intent — say it, and it's done
03
Work
You operate the software
The software operates itself
04
Progress
Quarterly feature releases
Smarter with every model generation
05
Moat
Lock-in and switching costs
Speed, taste and learning loops
06
Team
Hundreds of people per product
A handful of people per product
03 — The curve
Intelligence is getting cheaper faster than anything we have ever made.
0×
Drop in the cost of GPT-3.5-level performance, Nov 2022 → Oct 2024
$20.00 Nov 2022
$0.07 Oct 2024
$0/ $60
per 1M tokens, in / out
The frontier at launch. The benchmark price everything else is measured against.
Input price vs GPT-4 launch$0/ $10
per 1M tokens, in / out
24× cheaper input than GPT-4 at launch, two and a half years later — and far more capable.
Input price vs GPT-4 launch$0/ $25
per 1M tokens, in / out
80.9% on SWE-bench Verified — resolving real GitHub issues from real codebases.
Input price vs GPT-4 launch$0/ $0.50
per 1M tokens, in / out
150× cheaper input than GPT-4 at launch, with a two-million-token context window.
Input price vs GPT-4 launchFrom solving half to solving four in five.
SWE-bench Verified — share of real software issues resolved, as reported at release
- Claude 3.5 Sonnet Oct 20240%
- Claude 3.7 Sonnet Feb 20250%
- Claude Sonnet 4 May 20250%
- GPT-5 Aug 20250%
- Claude Opus 4.5 Nov 20250%
List prices and vendor-reported benchmark scores at release. This field moves weekly — treat every number here as a snapshot, not a ceiling.
04 — Daily practice
We don't just build with AI. We live in it.
Every day, for almost everything. That is how we find the gaps — by pushing these systems until they break, and noticing where nobody has built the bridge.
The boring 80% disappears
Writing, reviewing and refactoring at machine speed, so human attention goes to the interesting 20%.
Reads everything
The 400-page report, sixty papers, the entire codebase — then answers the question you actually had.
Every tongue, one sentence
Drafting, translating and editing across dozens of languages. Hindi to Japanese without breaking stride.
Sees what you see
Screenshots, charts, handwriting and photographs understood as easily as plain text.
Does, not just says
Plans multi-step work, uses tools, checks its own output and hands back a finished result.
A partner at 3 a.m.
A tireless sparring partner for half-formed ideas. It never says “let's take this offline.”
05 — Signal