AI

How Windsurf split into Google Antigravity and Devin Desktop—and why Cognition's quiet SWE-2 launch makes that strange AI coding history matter again.

Why ChatGPT can feel less capable after launch: risk controls, hidden model routing, lower Juice values, and reasoning that gets cut short under load.

Explore GPT Image 2.5 Flare and Sunburst, with 12 reusable prompts, eight generated examples, and a practical look at speed, image quality, and editing.

A practical AI workflows guide for choosing models, designing agents, controlling costs, and evaluating real work without trusting benchmarks alone.

Fresh CodexRadar data shows when GPT-5.6 Sol, Terra, or Luna makes sense, separating repeatable task coverage from latest-run cost and speed.

DeepSeek V4 Flash and GPT-5.6 Luna theoretically match Sol xhigh's coverage for about $0.42, but only if a verifier can identify the correct patch.

GPT-5.6 Luna is now 80% cheaper in the OpenAI API, making Luna Max practical for bounded Codex subagent tasks while Sol handles planning and review.

GPT-5.6 Sol passed fewer repository tasks at max than xhigh while costing 46% more. This benchmark maps the full Codex effort curve across 112 tasks.

Personal GPT-5.6 Juice value observations for Sol, Terra, and Luna across reasoning effort levels in Codex, plus prompts used to reproduce them.

Build a custom LLM evaluation with hidden tests, mechanical scoring, blind review, and checks for variance, leakage, and framework confounders.

Claude Opus 5 and Claude Fable 5 tied on most of eight tests. Fable won the decision memo, while Opus delivered the same objective score at half the price.