AI Labs Latest: The October drop · Oct 2026

AI moves fast. We pay attention.

Everyone's generating; almost no one's reading the output. We read it, and we measure it. The studies come first - then we embed with your team and build the software that ends the busywork.

The studies ↓ Book a teardown

01 - Research

We test the models before we build with them.

Two running studies and the write-ups that come out of them. Every prompt, score, and deliverable is public; nothing here is a press release re-plotted.

Study 01 · running since Jul 2026 · 5 runs

The Startrise LLM Benchmark

Every frontier model gets the same 12 single-shot frontend briefs. Each deliverable is driven in a real browser, judged blind by a three-lab panel, and rated blind by a human. Re-run every time the field moves.

Explore the benchmark → Read the Oct 2026 digest →

All-time top five · overall score

01 Claude Opus 5 82.3
02 Claude Opus 5.5 (high) 80.2
03 GPT-6 Astra 79.9
04 Kimi K3 79.8
05 GPT-6.1 Sol 78.6

Gates 25% · blind panel 45% · blind human 30%. Cross-run gaps under three points are noise.

Models: 30
Labs: 9
Deliverables: 353
Blind verdicts: 1,059

Study 02 · Oct 2026

Small decisions, smaller AI bills

Filter a feed, match an audience, connect useful pages, or catch a made-up fact. All 10 experiments have custom interactives with recorded Jev, Clef, Luna, Opus, and local baseline results.

Open the study →

What it measures

Recorded decisions against human annotations, exact fixtures, or disclosed AI references
Cost per decision, in cents, from recorded API usage
Real source records, side-by-side model decisions, and the mistakes behind the scores

Tests: 10
Models: 5 + baseline
Recorded runs: 154
Pilot calls: 840 · $0.48

Write-ups from the studies

02 - What the lab ships

Internal software that runs the business behind the business.

For ops leaders who can see their team doing work a computer should do, and engineering leaders whose roadmap can't spare the hands. Not the customer-facing stuff - the tools your team actually lives in.

/01

Automations & agents

Workflows and agents that move the data, draft the first version, and take the boring 80% off your team's plate - while a human keeps the last 20%.

/02

Internal apps

The tool your team keeps asking IT for. Intake, review, approvals - shipped in days, not next quarter.

/03

Control centers

One screen where your team starts, watches, and overrides every automation - so nothing runs in the dark.

/04

Dev environments

Claude Code, Cursor, MCP - rolled out, standardized, and taught. The tooling and the rules of the road.

/05

Monitoring & guardrails

Dashboards, alerts, and audit trails that catch the weird stuff before your customers do.

03 - How the embed works

An embed, not an engagement.

Time-boxed, out in the open, built to hand back. First opportunity map in two weeks.

01
Embed

A few days shadowing ops, support, and your engineers - paying attention to where the hours actually go.
02
Map

Every workflow becomes a scored opportunity: hours saved, risk, effort. A ranked list of what to build, not a forty-slide strategy.
03
Build

The top picks shipped as real software, in your stack, with your devs in the loop. Weekly demos, PRs you actually review.
04
Hand off

Monitoring wired in, runbooks written, your team trained to run it. Then it’s yours - code, systems, playbook.

04 - Proof

Automations that survive contact with production.

Public launches and private enterprise work, including multiple Fortune 100. Names under NDA; references on request.

Editorial network automation · 1,000 posts/day capacity with QA gates · <5 min operator time per 100 posts

Support automation · Response time cut ~80% with a human on QA

Ops control center · Real-time monitoring + overrides wired into production