Claude Donut, Qoduck and Cactus: Model Leaks Checked

World of AI's weekend roundup opens with names that sound like imminent launches: Claude Donut, Silicon, Obsidian 3, OpenAI Cactus and Kraken. It then shows Qoder's Qoduck and previews attributed to Qwen, Gemini and Seedream. The key distinction is between a name seen in a product or screenshot and a model whose maker has confirmed its identity, access and performance. Most of these names are still in the first category.

Watch the World of AI Episode

The video includes a sponsored Omen AI data-center cooling segment around 02:46. Its model-identification arguments are the creator's reporting and interpretation, not announcements from the named labs.

Source Note: What Was Checked

This article uses the supplied transcript and the original video. On 8 November 2026, we checked Cognition's live leaderboard and methodology, Qoder's Qoduck terms, and the official Claude Code, Gemini and Seed documentation. We did not inspect Cognition's compiled CLI, run the hidden models, reproduce benchmark scores, or authenticate the social-demo footage. The supplied X status URLs are shortened with ellipses; links below go to the named accounts rather than pretend to identify a particular post.

What Is Confirmed, and What Is Claimed?

Claude Donut, Obsidian 3, Cactus and Kraken

The video reports that a Claude Donut EAP identifier maps to Silicon High in a Devin CLI build, while another identifier maps to Obsidian 3 High. Different internal labels can indicate separate experiments, versions, routing targets or aliases. Without a vendor model card or reproducible access, none of those possibilities can be promoted to “Anthropic launched a new Claude family.” Even “EAP” is an interpretation of an abbreviation in this context.

The benchmark claim also needs its denominator. Cognition's FrontierCode 2.0 explanation describes 132 maintainer-defined software tasks with reviewer criteria, and its live page currently shows Opus 5.5 at the top of the visible best-reasoning results. A reported hidden Obsidian row, even if genuine, is not a public launch, and a score in one harness is not a general coding verdict. Likewise, Cactus and Kraken being positioned in a reported Devin model table does not tell us whether they are new GPT architectures, private experiments or placeholders.

Qoduck: A Real Trial, an Unknown Model

Here the official record is unusually clear. Qoder says Qoduck is a stealth model whose identity it has not disclosed. Starting 7 November, eligible individual Qoder International users receive a limited daily free allowance; after that, usage can consume Credits. Teams and Enterprise accounts are excluded from this promotion, and the terms can change. The video's “probably Qwen 4 Plus” argument rests on output style and apparent behavior, not a provider confirmation. The same caution applies to its comparisons with DeepSeek and Opus: prompts, tools, reasoning budgets, retries and evaluation criteria would need to match.

If you want to try it, follow Qoder's own steps: sign in to an eligible individual account, update the supported Qoder product, select Qoduck in the model picker, and check the live allowance before a larger test. Give it one small task you also run with your current model, then compare correctness, edit quality, review time and credits used. Avoid putting confidential code into an unidentified model until your organization has approved the data path.

What Else in the Roundup Matters?

Claude Code's game guidance and cloud work

The video says Anthropic revised server-side wording around recreating existing games and is preparing collaborative project features. Those claims come from reported instructions and interface sightings, not a stable public policy announcement. The practical boundary is still to build original work and check rights to game names, characters, artwork and music. The episode also shows Claude Duet phone control and easier cloud project setup. Anthropic's settings documentation confirms a crucial distinction: cloud sessions run in a fresh repository clone and do not automatically inherit every local permission rule, hook or credential. Review each environment's access before assuming a phone-controlled or cloud session has the same guardrails as your laptop.

Robot-safety research needs its test conditions

At 21:05, the creator summarizes experiments in which manipulated visual inputs allegedly led models toward harmful simulated actions. These are researcher-reported scenario results, not a measured rate of dangerous behavior in real homes or factories. The demo used safe stand-ins, and this article did not reproduce the study. The useful design implication is narrower: a physical agent must check what its sensors mean before acting, with independent safety interlocks rather than language-model judgment alone.

Rapid-fire items

Grok Bot in Slack, Muse Spaces, OpenAI data-center spending, a vocal-tract synthesizer and Figure robots round out the episode. Treat an interface preview or social clip as a signal to follow, not evidence of general access or deployment. The same applies to a fast Gemini “Hydrogen” 3D demo and Seedream 5.5 images: Google's Gemini catalog and ByteDance Seed's catalog are the better checkpoints for official model names and availability.

A Better Test Than Guessing the Codename

Pick one task you actually need done, keep the input, tools and acceptance criteria fixed, and record the result for each accessible model. For code, inspect the diff and run tests; for visuals, save the prompt and evaluate usable output rather than a highlight reel. Note total retries, elapsed time, credit use and human repair. Then check the vendor page for a model ID and terms before adopting it. This turns the rumor cycle into a useful evaluation habit without pretending anonymous checkpoints are products.

Related Videos and X Accounts

The three companion videos named by the creator are below. They revisit rumored models; none independently identifies Donut, Qoduck or Cactus.

For the reported Devin identifiers and benchmark sightings, see @abhinavflac and @StatsWire. For model demos and the continuing rumor trail, see @chetaslua, @testingcatalog and @loktar00. The robotics and image examples are associated in the supplied list with @robocurve and @lyraxana. These are links to accounts, not verified citations to the incomplete status URLs or endorsements of the claims.

Sources and Resources by Role

Video Chapters

The supplied list placed “Claude Code Collab” at 14:04 after a 15:42 segment. The transcript places that discussion at approximately 17:04, which is the chapter link used below.

Hidden model names

Qoduck and Qwen previews

Coding products and other model claims

Infrastructure and closing demos