Skip to content
Houtini.
Contact
Research

Local models, the AI job market, and the frontier AI APIs at work - our current research tracks:

One benchmarked on our own hardware, one read from live data, one priced live from the routing market. Each links through to the page that owns the depth, and we're explicit about which numbers are ours and which we're reporting.

01 · Local models Measured on our hardware · updated weekly

What a 48GB card can do with some popular open-weight models on your workstation.

We run a local model fleet as a delegation sidekick - Claude orchestrates, local models execute - and every model in it earns its place through the same benchmark harness: decode speed, prefill, cached time-to-first-token, a structured tool-calling gate, and a real TypeScript task judged on whether the code compiles. The findings that matter came from the failures as much as the wins: the quantisation format trap that costs consumer cards 40% of their speed, a mixture-of-experts model twice as fast as a dense model half its size, and a headline technique that never served a single token on our rig.

18.8 → 66
tok/s, same model, after tuning (3.2x)
133 tok/s
35B mixture-of-experts daily driver
176 tok/s
fastest model in the fleet (8B, 1B active)
26 runs
recorded benchmarks, every config logged
02 · AI and the job market Live · YubHub

AI isn't coming for your job. The data says it's coming for the grunt work.

25,173 live job ads from the YubHub index, joined to Anthropic's Economic Index and cross-checked against LinkedIn's Work Change Report. Three findings hold up everywhere we look: the most AI-exposed roles are still being hired in volume (augmentation in the data, not displacement); the work that stays is judgement, strategy and integration; and the market is rotating rather than retreating - 1.3 million new AI-related roles in two years, with job titles appearing weekly that didn't exist as categories three years ago.

Skills where AI shows up in the work, augmentation at scale
50 skills · Anthropic Economic Index methodology · /stats/skill-ai-exposure
Live
Amplified · humans + AI Native · work the model does directly
0.0 0.1 0.2 0.3 0.4 0.5 50 100 250 500 1,000 2,000 AI exposure (Anthropic Economic Index, March 2026) listings (log) Python Communication Project Management Physical Therapy Data Analysis Artificial Intelligence SQL Communication Skills C++ Machine Learning Leadership Java Sales Problem-solving Go Kubernetes Distributed Systems Analytical Skills Marketing Software Engineering JavaScript AWS Treatment Planning Rust Patient Assessment Teamwork Electrical Engineering TypeScript Product Management Excel Patient Care React Customer Service Large Language Models Mechanical Engineering Automation Computer Science Salesforce Data Science AI Tools Software Development Linux Networking Team Management Collaboration English SaaS PyTorch Google Cloud Time Management
Weekly job-ad volume, twenty-week window
24,722 ads · 32 weeks · /stats/hiring-velocity
Live
0 1,000 2,000 3,000 peak 3,195 2025-49 2026-01 2026-05 2026-10 2026-17 2026-21 2026-25 2026-29
03 · Frontier models Prices live from OpenRouter · benchmarks as reported

The frontier scorecard, and where each model earns its tokens.

The local fleet exists to take work off paid APIs, and that decision needs the other half of the equation: what the frontier APIs cost and deliver. The prices below are live - pulled from OpenRouter's public model list when this page loads - while the benchmark scores and one-line reads are reported from public leaderboards, not yet our own gauntlet, and we label the split on purpose. The routing conclusion at the bottom is ours.

Model $/Mtok in / out SWE-bench Verified The one-line read
Claude Fable 5 $10 / $50 95.0% The long-horizon king; the tool-calling reliability benchmark
Claude Opus 4.8 $5 / $25 88.6% The reliable workhorse at half Fable's price
Kimi K3 $3 / $15 93.4% Vision in the agentic loop - fixes CSS from screenshots
DeepSeek V4 Pro $1.17 / $2.34 80.6% The economics-breaker: ~80% SWE-bench at ~28x cheaper output than Opus
Qwen Coder (latest) $0.3 / $1 - The open-weight coder line we run locally, hosted for when the rig is busy

Prices live from OpenRouter's public model list · * = curated snapshot where unlisted · benchmarks and reads as reported

70% 80% 90% 100% $0.1 $0.25 $0.5 $1 $2 $5 $10 $25 input $/Mtok · log scale · live SWE-bench Verified 4x cheaper input Claude Fable 5 $10 Kimi K3 $3 Claude Opus 4.8 $5 DeepSeek V4 Pro $1.17 MiniMax M3 $0.3
Capability is a reported benchmark; price is what the routing market charges right now. The gap is the routing argument. Routed prices are uncommitted list prices and may reflect quantised deployments; benchmark scores were measured on the original weights.
Tier 1 · free, measured

The local fleet: extraction, formatting, TypeScript scaffolding, casual vision, bulk drafts. Free after hardware, and every capability claim benchmarked on our own rig.

Tier 2 · candidate, unverified

DeepSeek V4 Pro as the middle tier for high-volume agent loops that outgrow local capability - once it passes the same correctness gauntlet the local models face. Reported numbers don't route real work here; ours will.

Tier 3 · expensive, proven

Claude for orchestration, architecture, hard debugging - anywhere being wrong costs more than the tokens. Tool-calling reliability is where the frontier lead is least contested, and our whole delegation model depends on it.