The AI platform blog for modern teams

Most AI automations live in a Python script only one engineer has ever opened. A node-based canvas turns the automation into its own documentation — readable, reviewable, and editable by anyone on the team.

Opus 5 scored 43.3% on Frontier-Bench v0.1 at max effort vs. Sol's 37.5% — days after Sol's own sandbox breach made headlines for the opposite reason. The full 8-benchmark scorecard, the technical anatomy of the breach, and what the gap says about the capability/safety tradeoff both labs are making.

At 50 seats, ChatGPT and Claude both push you off their mid-tier plans and into an enterprise minimum you don't need yet. Here's what the seat math actually looks like — and the self-hosted alternative.

Engineering orgs routinely carry three separate AI subscriptions because no single vendor is best at everything. Here's why that's a procurement failure, not a feature requirement — and what replacing it with one self-hosted platform actually looks like.

Moonshot released Kimi K3's open weights today under the Kimi K3 License — a 2.8-trillion-parameter model trading blows with GPT-5.6 and Claude Fable. Self-hosting means finding ~1.4TB of GPU memory. Or you just point Wirebase at it.