Gemini 4 Krypton, GPT‑6.2 Astra, Claude Code upgrades and more
AI development isn’t slowing down. This week brought a wave of model upgrades, new tools for developers, and even an AI-assisted rocket milestone. From Google’s new Gemini 4 Krypton checkpoint to leaked GPT‑6.2 Astra samples and major Claude Code improvements, here’s a breakdown of what actually matters and how it might affect your work.
Gemini 4 Krypton: Google’s next big model upgrade
Google is quietly testing a powerful new Gemini 4 variant called Krypton. Internally, it appears in Arena and the API under the “Gemini 4 Hydrogen” name, with a fresh checkpoint first spotted on November 13.
Early benchmarks and community tests suggest this new Krypton checkpoint is a major leap over earlier Gemini 4 Argon and Carbon models. It reportedly:
• Surpasses Gemini 4 Carbon on most domain tasks
• Rivals Claude Opus 5.5 in several areas
• Comes surprisingly close to – and sometimes matches or slightly beats – Fable 5.2 on specific coding and visualization tasks, while being significantly cheaper
What Krypton is already doing well
Developers testing the model are seeing strong results in real-world coding scenarios, especially for web development and visual programming.
Examples shared so far include:
• A basic but fully functional Counter‑Strike-style FPS clone: Krypton generated the core gameplay, UI, loadouts, map, health bar, and minimap logic. It’s not AAA quality, but it’s far beyond what older Gemini models typically managed in one go.
• A 3D rocket engine simulation: Inspired by Ursa Major, Krypton produced a full 3D visualization with thermal flow effects and heat visualization, all in a single HTML file using pure JavaScript and no external assets.
• Page‑flip animation: A book page‑flipping effect coded entirely in JavaScript, with realistic physics and smooth animation, only minor visual glitches.
• SVG controllers: Highly accurate SVG renderings of a PlayStation 4 controller and a GameCube controller, including correct dimensions and small UI elements that many models usually get wrong.
The big theme: Krypton seems much better at following instructions, handling UI details, and reducing “lazy” or incomplete outputs — long‑standing complaints about earlier Gemini versions.
How Krypton compares to Fable 5.2 and Opus 5.5
Some testers claim Krypton is “Fable 5.2 level but cheaper,” but the reality looks more nuanced:
• For web development and general coding, Krypton feels closer to Claude Sonnet 5.5 or Opus 5.5 than to Fable 5.2’s absolute peak performance.
• In certain domains, especially visual coding and simulations, Krypton appears to match or beat Opus 5.5 and gets surprisingly close to Fable 5.2 quality.
• Given expected Gemini 4 pricing, Krypton could become a highly cost‑effective “near‑frontier” model for everyday dev work.
Google seems to be iterating quickly on Krypton checkpoints, tightening up hallucinations, instruction following, and output completeness. If it launches at the same introductory price tier as other Gemini 4 models, it could be one of the best value options on the market.
Claude Code: custom themes and smarter projects
Anthropic has shipped a big quality‑of‑life update for Claude Code, especially inside the desktop app.
Custom themes with mods
You can now create fully custom themes using the new “mods” system in the Claude desktop app. This isn’t just color tweaks — you can:
• Change the full appearance of the interface
• Add animated backgrounds
• Make the editor feel like a personalized workspace
You don’t need to hand‑code the theme either. You can describe the look you want, optionally provide reference images, and have Claude help generate the theme. Anthropic also offers a theme example plugin to get you started. For now, this works in local sessions, with cloud session support coming later.
Big upgrades to Projects and parallel work
The more impactful update is to Claude Code’s Projects feature, which is getting closer to a real multi‑agent coding environment:
• A more proactive project coordinator: Claude can now coordinate work across multiple parallel threads, answer questions in a shared project chat, and keep you updated on what each thread is doing.
• New commands:
– /bytheway: Ask an unrelated question without derailing your main thread. Claude spins up a separate side thread automatically.
– /fork: Branch a thread to try a different solution while keeping your original progress intact — perfect for testing alternative implementations.
Other improvements include:
• Thread reminders: Ask Claude to remind you about a thread later; it temporarily hides, then resurfaces when due.
• Read‑only thread sharing: Share individual conversations without requiring the recipient to sign into Claude (threads that touch your local machine remain private).
• Better navigation: The updated threads panel shows recent, updated, and pinned threads with thumbnails and artifacts, making it easier to find past work.
Projects is currently in public beta for Claude Pro and Max users, with a waitlist for others.
How Anthropic says you should pick Claude models
Anthropic also shared new guidance on when to use each Claude 3.5 model — and the takeaway is that the most expensive model isn’t always best.
According to their internal tests (52 coding tasks in Claude Code):
• Claude Opus 5.5 should be your default for daily coding and general development, at medium reasoning effort. It solved 84.6% of tasks.
• Claude Fable 5.2 is for the hardest problems: major refactors, big architecture changes, deep research, or long‑running autonomous workflows. It solved 82.7% of the same tasks but cost ~2.8× more per solved task than Opus 5.5.
• Claude Sonnet 5.5 is recommended for smaller, well‑scoped tasks like bug fixes, small features, and local edits.
• Claude Haiku 5.5 is ideal for lightweight utility work such as classification, searching files, or acting as a sub‑agent.
All four models support a 1 million token context window, so you don’t lose context length by choosing a cheaper option. The main lesson: match model choice to task difficulty instead of always defaulting to the most powerful (and most expensive) option. For more background on how Claude stacks up against other leaders, you can compare this guidance with our earlier coverage of Claude Code and competing agent platforms.
GPT‑6.2 Astra: early leaks look wild
OpenAI’s upcoming GPT‑6.2 Astra model has started to leak through early testers, and the sample outputs are raising eyebrows.
From what’s been shared so far, Astra shows:
• Extremely detailed, creative visual environments generated from scratch
• High‑quality, production‑ready web experiences in a single pass
• Strong performance on agentic tasks, where the model plans and executes multi‑step workflows autonomously
The most striking change is how often Astra seems to nail a polished, visually impressive result on the first attempt — something that previously required several iterations with other models.
Gemini 4 models help crack long‑standing math problems
After OpenAI’s recent work on AI‑assisted mathematical discoveries, Google has announced its own breakthroughs using the Gemini 4 family (Argon, Carbon, Hydrogen/Krypton) in multi‑agent setups.
Using Gemini‑driven systems in combination with human mathematicians, Google reports:
• The first machine‑checked formalization for the 3D Birkhoff conjecture, with AI identifying and fixing a 50‑year‑old gap in the proof.
• A resolution of a 20‑year‑old C¹ regularity problem.
• Verified proofs for the 40‑year‑old “caterpillar conjecture.”
• An autonomous formalization of the Weil conjecture via their Teamwork/Leap system, generating around 1.8 million lines of Lean proof code across 29,000 sub‑agent calls.
This continues a trend where frontier language models are becoming powerful research tools for formal mathematics and proof verification, not just natural language.
OpenAI shutters its short‑lived robotics effort
Earlier this year, OpenAI quietly launched a robotics initiative. Just five months later, that team is being disbanded.
The reason: OpenAI is choosing to focus its resources on scaling core models and reinforcement learning systems rather than spreading itself across hardware‑heavy robotics projects. In practice, that means more compute and budget going into GPT‑series training and agent capabilities instead of physical robots.
AI wallets: ChatGPT and Gemini move into payments
Both OpenAI and Google are moving aggressively into AI‑powered payments and shopping.
ChatGPT Wallet
OpenAI is rolling out “Wallet” for ChatGPT Pro users in the US. It’s designed to let AI agents make purchases on your behalf without exposing your card details.
Key points:
• Works with American Express, Mastercard, and Visa
• Every transaction still requires a security check
• Eligible purchases maintain existing card rewards and dispute protections
In practice, this could let you do things like: “Order my usual weekly groceries,” have the agent plan the cart, and then approve the payment via Wallet.
Gemini integrates with Google Wallet
Google is taking a similar route. Gemini users in the US can now connect Google Wallet directly to Gemini, enabling AI‑driven purchases and autonomous workflows like ordering supplies, groceries, or other recurring items.
Between ChatGPT Wallet and Gemini + Google Wallet, AI assistants are moving from “advice only” into actually transacting in the real world — with all the convenience and risk that implies.
GLM 5.5 rumors: a cheaper Opus‑level model?
There’s growing chatter out of China about GLM 5.5, based on an internal Zhipu AI discussion that appears to have leaked. A benchmark chart circulating online claims that GLM 5.5 surpasses Claude Opus 5.5.
Important caveats:
• The leak is unverified and could be fake.
• No official release date or pricing has been announced.
If even directionally true, this would mean an Opus‑class model at a lower cost — something many developers have been asking for. Given how GLM 5.2 already fits into some workflows (including via Claude Code, as we explored in our guide to using GLM 5.2 alongside Opus), GLM 5.5 is worth watching closely.
Gemini Notebook and comic artifacts
Gemini Notebook is also seeing small but interesting upgrades. The configuration page for its comic artifact generator has been updated, making it easier to create different kinds of comic‑style content directly from within the notebook interface.
While this isn’t a headline feature, it shows Google’s continued push to turn Gemini into a creative, multimodal workspace rather than just a chat box.
Google Anti‑gravity: dynamic workflows for agents
Google’s Anti‑gravity framework for AI agents is getting more powerful with “dynamic workflows.”
Using the /run-workflow command, you can now have an agent automatically generate a Python script that:
• Orchestrates multiple sub‑agents
• Runs tasks in parallel
• Hands off work between agents in a controlled way
Instead of relying purely on prompts — which can cause agents to drift or behave unpredictably — dynamic workflows let you:
• Inspect and edit the generated Python script before running it
• Save workflows for reuse
• Achieve more deterministic, repeatable agent behavior
This is a big step toward production‑grade agent systems where you want both flexibility and control.
DeepSeek: new app‑like API experience and possible paid plans
DeepSeek is expanding beyond its popular desktop app and making that experience available to developers through its API.
The new capabilities include:
• Built‑in web search
• File handling and link processing
• Safety tooling
• Voice features
• Separate APIs for OCR, speech recognition, and text‑to‑speech
Developers can now recreate much of DeepSeek’s own app behavior inside their products with just a few parameters. There’s also an open‑source front‑end demo that even supports importing existing conversation history.
On the business side, new code on DeepSeek’s site references quota limit notifications, hinting that paid subscriptions and stricter rate limits may be coming. That would mark a shift from its current, very generous free‑access model.
Muse Spark 1.4 and Meta’s quiet model upgrades
Meta appears to be preparing a new model called Muse Spark 1.4, spotted in an internal pull request. This has led to speculation that the “Futon Alpha” model some users recently tested — which was notably fast and capable — may actually have been an early Muse Spark 1.4 variant.
Meta has been steadily iterating on its generative models with less fanfare than some competitors, but Spark 1.4 suggests they’re still pushing on speed and responsiveness, not just raw benchmark scores.
Qwen 4.0 Max: stealth image quality boost
Alibaba’s Qwen 4.0 Max preview model seems to have received a quiet upgrade, particularly in image generation.
Recent tests show:
• More coherent, detailed images
• Better adherence to prompts
• Clear visual gains versus earlier preview outputs
For instance, a prompt for “a pelican riding a bike” now produces a much more faithful, visually appealing result than an earlier version of the same model on the same prompt.
Alibaba’s AI‑native laptop
Alibaba is also pushing Qwen beyond the cloud with a new AI‑centric laptop powered by its “next‑generation Qwen technology.” The device runs a native “Agent OS” that lets AI agents interact directly with installed apps and system services.
It’s one of the first mainstream attempts at a laptop designed from the ground up around AI agents rather than traditional apps and operating systems.
Anthropic eyes an IPO to buy more compute
Anthropic CEO Dario Amodei recently spoke about the company’s plan to go public. The main motivation is straightforward: compute.
He framed the IPO as a way to build a financial “buffer” between a low‑compute and high‑compute future, ensuring Anthropic can keep investing heavily in:
• Larger and more capable models
• Applications in biology, robotics, cybersecurity, and knowledge work
• Long‑term research that requires massive GPU and TPU budgets
Amodei also downplayed the exact timing, saying they’re not obsessed with picking a perfect market window so much as securing the resources needed for the next phase of model scaling.
SpaceX catches Starship with AI assistance
Stepping briefly outside pure software: SpaceX has reportedly achieved a historic feat by catching the full Starship vehicle at the launch tower for the first time.
Previously, we’d seen the company catch the Super Heavy booster with giant mechanical arms on the launch tower. Now they’ve done it with the entire Starship stack — roughly akin to catching a multi‑story building after it’s flown around the planet and bringing it back to its original pad.
Elon Musk has confirmed that xAI’s Grok was used in the development of this mechanism, including calculations and other support tasks. It’s a striking example of how frontier AI models are starting to play a role in high‑stakes engineering and control systems.
Perplexity’s “teammate” comes to Mac
Perplexity is extending its “teammate” feature — essentially a proactive personal AI agent — to its Mac app for Perplexity Max subscribers.
Teammate can:
• Run tasks on your behalf using Perplexity’s models
• Use plugins and skills to interact with external services
• Go beyond Q&A to handle ongoing research and workflow automation
Functionally, it’s Perplexity’s answer to agentic tools like OpenAI’s agents and other multi‑step automation assistants.
Fable 5.2 keeps pushing what code‑only animation can do
Despite all the new contenders, Fable 5.2 continues to impress with its ability to generate complex visuals purely through code.
Recent community projects include:
• A multi‑hour, fully coded 3D animation of a computer being built, generated via Blender scripts. One user reported spending about 7 hours and $138 in API costs to drive Fable 5.2 through the entire pipeline.
• An anime‑style music video recreated entirely through Blender code, with Fable 5.2 scripting the scene layout, camera moves, and effects.
The key point: we’re reaching a stage where high‑end visuals can be produced by models that write and refine code, not just by direct image or video generation. That’s a huge deal for developers who want full control over assets, pipelines, and reproducibility.
What this all means for builders right now
Across all these updates, a few trends stand out:
• Near‑frontier models are getting cheaper. Gemini 4 Krypton and rumored GLM 5.5 suggest you won’t always need to pay Fable‑ or Opus‑level prices for “good enough” frontier performance.
• Agents and workflows are maturing. Claude Projects, Anti‑gravity dynamic workflows, DeepSeek’s app‑like API, and Perplexity’s teammate all point toward more structured, multi‑agent development patterns.
• AI is moving into transactions and hardware. ChatGPT Wallet, Gemini + Google Wallet, Qwen‑powered laptops, and Grok‑assisted Starship catches show AI stepping out of the browser and into the real world.
If you’re building with AI today, this is a great moment to:
• Re‑evaluate your default model choices for cost vs. quality
• Start experimenting with agentic workflows rather than single‑prompt interactions
• Think about where AI can safely take actions — not just make suggestions — in your product or stack
Comments
No comments yet. Be the first to share your thoughts!