Pixel-art adventurer in a helmet and backpack chases a glowing ring down a red-lit sci-fi hallway full of crates and pipes.

Quick answer

The Claude Silicon leak is a report that Anthropic is testing a new, separate Claude model family called “Silicon”, spotted as “Claude Donut EAP” inside Cognition’s Devin CLI on 6 November. It sits alongside “Obsidian 3”, another hidden model that reportedly beat Claude Opus 5.5 on Cognition’s coding leaderboard. The same Devin build also revealed two OpenAI code names, Cactus and Kraken, while Qoder launched a free stealth model, Qoduck, that testers think is Qwen. None of these models is confirmed. Treat them as a sign of what may launch soon, not a reason to switch tools today.

Key takeaways

  • Claude Donut EAP reportedly maps to Silicon High, a model family separate from Obsidian 3.
  • Obsidian 3 High (code name Claude Onigiri EAP) reportedly topped Claude Opus 5.5 on Cognition’s Frontier 2.0 leaderboard.
  • Silicon appears to have five reasoning effort levels: low, medium, high, x-high and max.
  • OpenAI’s Cactus and Kraken appear between GPT-6 Luna and GPT-6 Astra in Devin’s model table.
  • Qoduck is free to try on Qoder for a limited time. Testers put the odds it is a Qwen model at 70% to 75%.
  • Anthropic has reportedly softened Claude Code’s game recreation guidance, and Claude Duet can now manage Claude Code sessions from your phone.
  • All model names and benchmark claims here are leaks or secondary reports until the companies confirm them.

🛠️ Practical guides

Ready to put this into practice? These step-by-step guides answer two common follow-up questions:

Another weekend, another batch of code names. This time the most interesting find is not a new version number but what looks like a whole new Claude model family, hiding in a coding tool’s configuration file.

This article explains the Claude Silicon leak, how it differs from the Obsidian 3 model, what OpenAI’s Cactus and Kraken might be, and why Qoder’s free Qoduck model has people talking about Qwen 4. It also covers this week’s practical Claude Code changes, which matter more to most readers than any leak. Last week’s Sonnet 6 rumours are in our Claude Sonnet 6 leak round-up.

Watch: the source video

What this video covers: WorldofAI’s 8 November AI news round-up covers the Claude Donut EAP and Silicon leak, Obsidian 3, OpenAI’s Cactus and Kraken code names, Qoder’s Qoduck stealth model and Qwen 4 comparisons, Claude Code’s revised game guidance, cloud sessions and Claude Duet support, Gemini 4 Hydrogen, a robotics safety study, Grok Bot for Slack, Meta Muse Spaces, Seedream 5.5, OpenAI’s data centre plans and Figure 03 robots. The video is sponsored by Omen AI.

Video: WorldofAI (@intheworldofai). Sponsored by Omen AI (Omen Sentry). This summary and the opinions in it are our own. Model code names, leaderboard claims and unreleased features are leaks or secondary reports until the companies confirm them.

What is the Claude Silicon leak?

It is a claim, based on a reverse-engineered build of Cognition’s Devin CLI, that Anthropic has a new model family called Silicon in early testing. Devin is an AI software engineering tool, and its command-line app lists the models it can call.

According to the findings shown in the video:

  • A new entry, Claude Donut EAP, appeared in a Devin CLI build released on 6 November.
  • Someone who reverse-engineered the CLI found that it maps to a model called Silicon High.
  • “EAP” usually stands for early access programme, which suggests limited testing rather than a public launch.
  • The entry sits next to Opus 5.5, Mythos 5.1, Fable 5.1 and Haiku 5.5, plus new medium and x-high tiers for Claude Mythos 5.5.

The video also shows a free-fall physics demo said to come from one of these hidden models. It looks impressive, but a single demo tells you little about everyday work.

🔎 Unconfirmed / rumour: Nothing here comes from Anthropic. The information is a reverse-engineered configuration file reported on social media. Configuration entries can be placeholders, test aliases or models that never ship.

How is Silicon different from Obsidian 3?

They appear to be two separate model identifiers, not two names for the same thing. That is the main new detail in this leak.

Code name in DevinReported model
Claude Donut EAP (0x63)Silicon High, a new model family
Claude Onigiri EAP (0x5E)Obsidian 3 High, previously seen beating Opus 5.5 on Cognition’s Frontier 2.0 leaderboard

WorldofAI suggests three possibilities. Obsidian 3 could be Fable 5.5, with Silicon something that comes later. Silicon could be another Fable 5.5 checkpoint. Or Silicon could be a specialist family built for agentic coding, which would fit its appearance in a coding tool. The presenter leans towards Obsidian being Fable 5.5, but there is no direct evidence either way.

If you want to judge rumours like this for yourself, our guide on how to verify AI model leak rumours gives you a simple checklist.

What do five reasoning effort levels mean?

Reasoning effort is how long a model “thinks” before it answers, and more effort usually means better answers on hard tasks but slower, pricier responses. Silicon reportedly supports five levels: low, medium, high, x-high and max.

That matters because it is becoming the norm. OpenAI’s GPT-6 models already offer similar tiers, and the leaked Cactus and Kraken entries show the same five levels. In practice it means you choose speed and cost per task rather than per model:

  • Low or medium for quick rewrites, summaries and simple questions.
  • High for most coding and analysis work.
  • X-high or max for difficult, multi-step problems where accuracy matters more than time.

💡 Tip: Before you pay for a higher tier, run the same task at two effort levels and compare. For many everyday jobs, the cheaper setting is good enough.

What are OpenAI Cactus and Kraken?

Cactus and Kraken are two new OpenAI code names found in the same Devin CLI build, and nobody outside OpenAI knows what they are yet. In Devin’s model table they sit between GPT-6 Luna and GPT-6 Astra.

Each appears with the same five reasoning levels as other GPT models. WorldofAI reads this as a sign that they are separate model families rather than single checkpoints. They could be experimental reasoning models or new architectures in testing. For background on OpenAI’s recent naming, see our GPT-6.2 Astra round-up.

What is Qoduck, and is it Qwen 4?

Qoduck is a free stealth model on the Qoder coding platform, and early analysis suggests it is probably an Alibaba Qwen model, though that is not confirmed. It launched on 7 November with limited free usage.

  • Testers say it beats DeepSeek and Qwen 3.8 Max in their early tests.
  • Its reasoning pattern, refusal style and Alibaba Cloud-style API responses point to Qwen, with an estimated 70% to 75% likelihood.
  • WorldofAI thinks it may be a Qwen 4 Plus-sized model rather than the largest Max version, and says it outperformed DeepSeek V4.1 Flash on many agentic coding tests while using fewer tokens.

The presenter also clarifies his earlier Qwen 4 preview tests. The preview API he used routed requests across the four Qwen 4 models, so some outputs may have come from the smaller Qwen 4 27B and others from a larger model. In one test, a Qwen 4 preview at max thinking spent about 200,000 tokens building a dungeon crawler game, with results he describes as close to Claude Opus 5.5 on the same prompt.

If you want to try Qoduck, read how to use a free stealth AI model without risking your data first.

What changed in Claude Code this week?

Several small but useful Claude Code changes were reported, and these are the items most likely to affect your work this month.

ChangeWhat it means
Game recreation guidance revisedSimilar gameplay mechanics are now explicitly allowed. Copying a game’s name, characters, art, music, levels or look is still discouraged. There is no setting to turn it off.
Collaborative projects (spotted)A new section summarising recent threads hints at team features. Not released.
Claude Duet supportMonitor Claude Code sessions, approve changes and send instructions from your phone.
Conductor migrationAdd your Git repository to Claude Code projects and ask Claude to pick up branches where you left off.
Cloud environment set-upClaude works out dependencies, network access and API keys a project needs, then saves the environment for next time. Available on Pro, Max and Ultra.

The game guidance change is reported to sit under internal flags and to ship in Claude Code 2.1.3.17. In plain terms: you can build a block-building game inspired by Minecraft, but not a copy of Minecraft. Our guide to building an original game with an AI coding agent shows how to stay on the right side of that line.

Phone approvals are the change most readers will use. Our new guide on approving AI coding agent changes from your phone safely explains what to approve on the go and what to leave for your desk.

⚠️ Warning: Anthropic says API keys in the new cloud environments are stored securely and Claude cannot see them. Even so, never paste keys into a chat. Keep them in environment settings, as covered in how to store AI API keys safely.

What else happened in AI news this week?

  • Gemini 4 Hydrogen: a new Arena checkpoint of what is expected to be Google’s Flash model shows strong 3D generation, even at Arena’s low reasoning budget, and is said to be at least twice as fast as Gemini 3.8 Flash.
  • Robotics safety study: Robocurve PBC reports that manipulated video inputs (“perceptual attacks”) pushed AI-powered robots to attempt harmful actions up to 100% of the time in some tests. Without manipulation, GPT-6.1 Sol attempted harmful actions in 47% of trials and Claude Fable 5.1 in about 40%, while Gemini 4 Argon refused every one. The tests used safe props, and the figures reflect specific scenarios, not everyday use.
  • Grok Bot for Slack: reportedly coming soon, so you can mention a bot in a channel or message it directly, with its plugins, context and memory carried over.
  • Meta Muse Spaces: Muse can now build interactive pages, data visualisations and simple apps as artifacts.
  • Seedream 5.5: rumoured to improve fine detail and realism, though skin still looks slightly too smooth in the examples.
  • OpenAI data centres: CNBC reports that OpenAI’s CFO confirmed the company will keep expanding AI infrastructure after the US midterms, despite local concerns about energy use and grid strain. The presenter presents this as a notable tension; we report it without taking a side.
  • Vocal tract synthesiser: a developer showed a voice synthesiser that models the human vocal tract as branching tubes rather than stitching recorded sounds.
  • Figure 03: footage shows many Figure 03 humanoid robots at the company’s San Jose campus.

What does the video not confirm?

🔎 Unconfirmed / rumour: The video does not confirm that Silicon, Obsidian 3, Cactus or Kraken will be released, what they are called publicly, how they will be priced or when they will arrive. Leaderboard results for Obsidian 3 come from Cognition’s own leaderboard as reported online. Qoduck’s link to Qwen is an estimate. The Claude Code collaboration mode is spotted, not launched. Seedream 5.5 and Grok Bot for Slack are reports, not announcements.

The safety study figures come from one company’s research, as described in the video. We have not reviewed the paper ourselves.

What should you do about these leaks now?

  1. Keep using what works. No leak in this round-up is a reason to change tools today.
  2. Write a three-task test pack that reflects your real work, so you can judge new models fairly when they launch.
  3. Try stealth models with dummy data only, following our stealth model safety guide.
  4. Set up safe phone approvals if you use Claude Code, using our phone approval routine.
  5. Check your reasoning settings. If your tools offer effort levels, test whether a lower one is good enough for routine jobs.
  6. Wait for official announcements before you budget for a new model or plan.

Who should pay attention to these leaks?

These leaks suit you if you build with Claude Code, Devin or other coding agents, you plan software or content budgets for next year, or you like to test new models early.

They matter less if you use AI mainly for writing, email and research. Nothing here changes those tasks yet. The Claude Code updates and the phone approval feature are the practical parts to note.

Frequently asked questions

What is Claude Silicon?

Claude Silicon is the reported name of a new, unconfirmed Anthropic model family. It surfaced as “Claude Donut EAP” in Cognition’s Devin CLI on 6 November and appears to support five reasoning effort levels.

Is Obsidian 3 the same as Claude Fable 5.5?

Possibly, but it is unconfirmed. WorldofAI leans towards Obsidian 3 being Fable 5.5, with Silicon as a separate family, but there is no official word.

What are OpenAI Cactus and Kraken?

They are two OpenAI code names found in the same Devin CLI build, listed between GPT-6 Luna and GPT-6 Astra. OpenAI has not said what they are.

Is Qoduck really Qwen 4?

Testers estimate a 70% to 75% likelihood that it is a Qwen model, possibly a Qwen 4 Plus version. Neither Qoder nor Alibaba has confirmed it. See how to use a free stealth AI model safely before you try it.

Can I still make games with Claude Code?

Yes. Anthropic has reportedly revised its guidance so similar gameplay mechanics are allowed. Copying a specific game’s name, characters, art, music, levels or look is still discouraged.

Can I approve Claude Code changes from my phone?

According to the video, Claude Duet now supports Claude Code sessions, so you can monitor, approve and instruct from your phone. Our guide to approving AI coding agent changes from your phone covers what is safe to approve.

Who sponsored the WorldofAI video?

The video is sponsored by Omen AI, which makes Omen Sentry, a coolant monitoring sensor for liquid-cooled data centres.

The bottom line

The Claude Silicon leak suggests Anthropic is preparing more than one new model, with Silicon and Obsidian 3 appearing as distinct entries, while OpenAI tests Cactus and Kraken and a probable Qwen model runs free on Qoder as Qoduck. All of it is unconfirmed. The practical news is in Claude Code: clearer game guidance, easier cloud set-up and phone approvals through Claude Duet. Keep your current tools, prepare a short test pack for when these models launch, and set up safe approval habits now.

author avatar
Garry Knight
I'm Garry Knight, the person behind Prodify Digital. I write about email list building, email marketing, SEO, AI search and the tools that connect them. My aim is to make online marketing easier to understand, so creators and small business owners can make informed decisions about building an audience and keeping people engaged. Here you'll find straightforward guides and product reviews that explain what something does, where it fits and which limitations matter. The focus is on clear explanations and useful next steps—not hype, shortcuts or promises of easy earnings.

Leave a Reply