The X user @chetaslua posted on November 7 that a reverse-engineered build of Cognition’s Devin CLI maps a Claude-labelled codename, claude-donut-eap, to a model family called “Silicon” with five levels: low, medium, high, xhigh and max. We checked Cognition’s own npm packages: the Devin CLI build published on November 7 lists five silicon model identifiers, and the build published on October 30 lists none.
Anthropic has not announced a model called Silicon, and the identifiers do not name the lab behind it. By our inspection, the five levels carry the same names as the effort levels in Anthropic’s own documentation, and the same suffixes Devin attaches to Claude Opus 5.5 and Claude Haiku 5.5.
Business Pill · The Model and the Harness
An evergreen Business Engineer explainer on how a model and the harness it runs in work together. It teaches the general idea only and does not discuss Devin, Cognition or Silicon.
The key insight: Cognition’s Devin CLI build published on 7 November lists five silicon model identifiers, from low to max, that its 30 October build does not: by our count, 5 against 0. X users say a codename, claude-donut-eap, maps to silicon-high and read the claude- prefix as pointing to Anthropic. Anthropic has not announced a model called Silicon, and the identifiers do not name its maker.
In build 3000.12.28, which the npm registry lists as published on November 7, the identifiers silicon-low, silicon-medium, silicon-high, silicon-xhigh and silicon-max sit in a list of model identifiers directly after four obsidian-3 levels, which in turn follow five claude-haiku-5-5 levels. In build 3000.11.234, published on October 30, no silicon identifier appears. By our count, that is 5 identifiers against 0, in builds 8 days apart.
The npm package is named devin, described as “Devin for Terminal”, and lists github.com/CognitionAI/devin-cli as its repository; its maintainers include accounts with cognition.ai email addresses. We downloaded the Linux binaries of both builds and searched their text strings, without running them.
The codename claude-donut-eap does not appear in build 3000.12.28, and no identifier ending in -eap does. The string claude-onigiri appears in the October 30 build, next to claude-mythos-5-1, and not in the November 7 build. The X posts describe build 3000.12.17, which is not among the versions listed on npm, so we could not check the donut-to-silicon mapping ourselves.
Effort-level model identifiers per family in Cognition’s Devin CLI binary: Silicon has 0 in build 3000.11.234 (30 Oct 2026) and 5 in build 3000.12.28 (7 Nov 2026); Claude Opus 5.5, Claude Haiku 5.5 and Obsidian 3 have 5 in both. Data: FourWeekMBA inspection of the npm packages, 8 Nov 2026. Our count.
What Was Posted on X
🚨 HOLY SHIT… Anthropic might be Cooking an entirely Different Claude Code Family. 🚨
A mysterious model family called SILICON just showed up in the latest Devin CLI build.
And here’s what’s interesting: it has FIVE reasoning modes → Low → Medium → High → XHigh → Max
BUT WAIT… remember the mysterious DONUT codename we spotted earlier?
Well, we’ve now found a connection.
Reverse-engineering Devin CLI revealed:
→ “claude-donut-eap” → “silicon-high”
So DONUT is actually an alias for SILICON HIGH. 👀
Both names were missing from the older CLI build, which makes this discovery even more interesting.
Think about it. We might be looking at a completely new Anthropic model family, with five different reasoning configurations, quietly showing up before any official announcement.
A post by @buildwithrajath at 18:24 UTC on 7 November that calls DONUT an alias for SILICON HIGH and asks whether Anthropic is building a new model family. That reading is the poster’s; this piece reports what our own inspection of Cognition’s npm builds shows.
🚨 claude-donut-eap is silicon high
Devin labels it claude- so Anthropic
it’s an alias. reversed the Devin CLI (3000.12.17) and the parser maps the name straight to a model called “Silicon”
A post by @chetaslua at 11:08 UTC on 7 November describing a reverse-engineered Devin CLI build, 3000.12.17, that maps claude-donut-eap to silicon-high. That build is not on npm, so we could not check the mapping; the silicon identifiers themselves appear in the 7 November build we inspected.
What the Finders Posted
The codename surfaced first. @abhinavflac posted at 03:43 UTC on November 7 that claude-donut-eap “just appeared in Devin’s CLI (build 3000.12.17, Nov 6)” and was “stored right next to opus 5.5, mythos 5.1, fable 5.1, haiku 5.5 and the onigiri id”, adding: “Nothing confirmed, I’m only reading the build”.
At 11:08 UTC, @chetaslua wrote that the parser maps “claude-donut-eap → 0x63 → silicon-high” and “claude-onigiri-eap → 0x5e → obsidian-3-high”, and that “Devin labels it claude- so Anthropic”. @Mr_Salio posted at 14:36 UTC that Silicon is “listed directly after Obsidian 3”, which matches the order we found in the November 7 build.
A later post from @buildwithrajath at 18:24 UTC, calls DONUT “an alias for SILICON HIGH” and asks whether Anthropic is building “a completely new Anthropic model family”. That attribution is the posters’ reading of a codename prefix, not a statement from Anthropic or Cognition.
Where Obsidian 3 Fits
Obsidian 3 has appeared before. Our October 28 update on Cognition’s FrontierCode test reported that Cognition’s archived data file listed a model called Obsidian 3 at 46.8%, above Claude Opus 5.5’s 45.5%, under the claude-code harness, before the entry was removed. That file gave Obsidian 3 the chart colour Cognition uses for Muse Wonder 1.0, not the one used for Claude models, and its maker was not confirmed.
The Business Lever: Model Names in Third-Party Files
Devin’s own binary names Claude models directly, with Claude Opus 5.5 and Claude Haiku 5.5 identifiers carrying the same five level suffixes, by our inspection. As we read it, a third-party company’s own files can show a model name before any lab page does: Obsidian 3 first surfaced in Cognition’s benchmark data file, and Silicon in Cognition’s CLI build. Our earlier coverage of Anthropic’s named compute deals and Claude’s overnight memory work tracks what Anthropic itself has published.
The Structural Read
As we read the two builds, Silicon was added to Devin’s model list within an 8-day window, next to the Obsidian 3 entries that had already appeared in Cognition’s benchmark data.
As we read the X posts, the link to Anthropic rests on a codename prefix and on matching effort-level names, neither of which Anthropic or Cognition has confirmed.
As we read Anthropic’s documentation, low, medium, high, xhigh and max are its effort levels, so the matching names show how Devin labels models, not who built them.
@abhinavflac, who first posted the codename, on X
“Nothing confirmed, I’m only reading the build”
Three Implications
FOR COGNITION Its 7 November Devin CLI build lists five silicon identifiers that its 30 October build did not, by our inspection; Cognition has not described the model.
FOR ANTHROPIC Its documentation lists low, medium, high, xhigh and max as effort levels and names no model called Silicon or Obsidian 3.
FOR THE NUMBERS TO WATCH Our earlier piece reported that Cognition’s archived FrontierCode file listed Obsidian 3 at 46.8%, above Claude Opus 5.5’s 45.5%, before the entry was removed.
The essay says “Anthropic is betting on many harnesses, embedded where work already lives”.
As we read it, a third-party harness such as Devin listing models by name is one of those surfaces. The essay does not discuss Devin, Cognition or Silicon.
What Is Not Established
No source we read says which lab built Silicon or Obsidian 3. Anthropic’s documentation lists no model by either name, and Cognition has not described either on any page we found. We could not confirm the donut-to-silicon mapping, because build 3000.12.17 is not on npm and claude-donut-eap is absent from the November 7 build we inspected. Matching effort-level names do not show who made a model.
This Week In The AI Supercycle
The Great Convergence of Capital, Power, and Geopolitics
This week’s issue: for two years the AI build-out was priced as a supply story, and last week as a credit story. This week it was priced as a permission story, and the answer arrived as a cost, not a ban.
Cognition’s Devin CLI build of November 7 lists five silicon model identifiers, named low to max, that its October 30 build does not; by our count, 5 against 0. X users say a codename, claude-donut-eap, maps to silicon-high, and read the claude- prefix as pointing to Anthropic. Anthropic has not announced a model called Silicon, and the identifiers do not name its maker.
A note on sourcing. The build findings are from our own inspection of Cognition’s Devin CLI packages on npm on 8 November 2026: we downloaded the Linux binaries of builds 3000.11.234 and 3000.12.28 and searched their text strings without running them. The codename mapping and attributions are from the X posts linked above. Effort-level names are from Anthropic’s documentation. The 5-against-0 count is ours.
Gennaro Cuofano is a CRO and tech executive who has worked in AI since late 2015, starting with bringing NLP, natural-language generation, voice and chatbot products to the marketing industry. His background is in law and finance: he holds a Master's degree in Law and an International MBA with an emphasis on corporate finance (LUISS Business School and the University of San Diego, 2012), and worked as a financial analyst at a real-estate investment firm in San Diego and as an assistant controller. His work focuses on business model strategy, business engineering and, more broadly, structural analysis: how companies make money, read from their own filings. He created FourWeekMBA and leads research there and at The Business Engineer, his newsletter on AI and business strategy, with over 95,000 subscribers and more than 1,000 published analyses. His writing has also appeared on Entrepreneur (for example, How to Digitize a Family Business in Europe) and HackerNoon. Find him on LinkedIn and Substack. See how we source.
