Three research tracks. Every one with the numbers attached.
We benchmark local models on our own hardware, track the AI job market on live data, and keep an honest scorecard on the frontier APIs. The rule across all three: measured beats reported, and where a number is someone else's, we say so.
What a 48GB card really does, measured to the token.
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.
AI isn't coming for your job. The data says it's coming for the grunt work.
25,154 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.
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. These are reported numbers - vendor pricing and public leaderboards, not yet our own gauntlet - and we label them that way 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 | $0.435 / $0.87 | 80.6% | The economics-breaker: ~80% SWE-bench at ~28x cheaper output than Opus |
| MiniMax M3 | $0.60 / $2.40 | 75.0% | Budget text throughput; weak agentic scores on current comparisons |
The local fleet: extraction, formatting, TypeScript scaffolding, casual vision, bulk drafts. Free after hardware, and every capability claim benchmarked on our own rig.
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.
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.