
What happened
Cognition's Devin CLI build 3000.12.28, published November 7, lists silicon-low through silicon-max, directly after four obsidian-3 levels; build 3000.11.234, published October 30, lists none — 5 against 0 in builds 8 days apart.
Why it matters
Five new model identifiers appearing in Cognition's npm package within eight days signals an undisclosed model family is being wired into Devin before any public description; matching effort-level names show how Devin labels models, not who built them.
WHO IT HITSDevelopers and procurement teams choosing coding agents like Devin may soon see a Silicon model option they cannot yet evaluate; competing tool vendors may spot rival models in third-party packages before any vendor announcement.
Summaries like this, every morning.
The Silicon identifiers did not arrive alone. In build 3000.12.28, they sit directly after four obsidian-3 levels, which in turn follow five claude-haiku-5-5 levels — the same five-level naming pattern Devin uses for Claude Opus 5.5 and Claude Haiku 5.5. That ordering is what prompted X users to treat the new family as another entry in the same lineup, but the same document also shows how Devin labels effort levels generally.
Obsidian 3 has its own trail. An October 28 update on Cognition's FrontierCode test recalled an archived data file listing Obsidian 3 at 46.8%, above Claude Opus 5.5's 45.5%, before the entry was removed; that file gave Obsidian 3 a chart colour associated with Muse Wonder 1.0 rather than Claude models. The pattern the inspectors describe is that a third-party company's own files can surface a model name before any lab page does.
The key gap is attribution. The codename claude-donut-eap does not appear in build 3000.12.28, and no identifier ending in -eap does, so the donut-to-silicon mapping could not be checked; the original posts describe build 3000.12.17, which is not on npm. Matching effort-level names show how Devin labels models, not who built them, and neither Anthropic nor Cognition has described Silicon on any page found.
Pick your industry and the AI tools you use, and get news related to your work every day.
Free · 30 seconds with Google · unsubscribe anytimeWhat is AIToday? →
Ask AI anything about this article. The AI reads this article, earlier AIToday articles, and Wikipedia, and cites its sources. Q&As are published on this page for other readers too.
ISFnet started offering the Microsoft 365 Copilot 導入フルサポートパッケージ, bundling three trainings and AI agent require…

Goldman Sachs has limited some bankers' and traders' use of third-party AI coding agents such as Claude Code a…

Director Hironori Muraoka built the clinic's reservation system and dashboard himself using Claude Code, and r…

Mihail Eric, who teaches The Modern Software Developer at Stanford and leads AI at Monaco, brought open-source…

The author built an inventory-demand model with scikit-learn's RandomForestRegressor, saw cross-validation abo…

The author's PowerShell worker polls GitHub Issues and runs Claude Code on the DeepSeek API, with a separate G…
