AI News This Week: GPT-6.1 Terra, Gemini 4 Carbon, Copilot Fusion
Prefer to watch? Watch on YouTube · Subscribe for new builds · Read the transcript
In this video I rank the 25 biggest AI releases from the week of October 26 to November 1, 2026: five headliners, a rapid-fire round, and the five GitHub repos everybody was starring. For each one I tell you what it is and whether you can actually use it this week. If you build apps or AI agents and don't have time to scroll all week, this is for you.
What I cover in this video
- 0:45 GPT-6.1 Terra, OpenAI's cheaper and faster computer-use model
- 3:04 Gemini 4 Carbon takes #1 on the Artificial Analysis Intelligence Index
- 4:34 GitHub Copilot Fusion, so you stop picking a model yourself
- 6:05 Up to $500 a month in API credits for ChatGPT Pro and Business
- 7:36 Genie 4 world model, plus Tencent's open Hy-Nomas
- 9:07 Rapid fire: Claude Ultra, Claude Code compaction, Anthropic's S-1, Cursor, GitHub, Gemini API and more
- 17:30 The five trending GitHub repos of the week
- 20:34 Wrap-up: this was the week of computer use
The five headliners
GPT-6.1 Terra
GPT-6.1 Terra is OpenAI's new midsize model, built mainly for computer use. Your agent sends a screenshot, the model decides the next click or keystroke, your app does it, and the loop starts over. That loop burns a lot of tokens, so OpenAI is selling Terra as the cheap, fast option. They say it's half the API price of the last Terra (press reports say $1 in and $5 out per million tokens) and finishes tasks over three times faster on average than GPT-6 Astra and GPT-6.1 Soul. OpenAI also says that at max effort it costs about nine times less per task than Claude Opus 5.5. That's OpenAI's own number, and it's per task, not per token.
Can you use it? Yes. It's in ChatGPT Work and Codex on Plus, Pro, Enterprise, and Edu, and in the API as GPT-6.1 Terra. OpenAI still recommends Soul or Astra for the hardest coding. Terra is the one you run a thousand times a day: filling forms, QA runs, clicking through back-office apps.
Gemini 4 Carbon
Gemini 4 Carbon scored 59 on the Artificial Analysis Intelligence Index. That's #1, and nine points above Gemini 4 Argon. On their coding agent index it's second only to Claude Code with Opus 5.5 (59 vs 62), but it costs about 23% less per task. Google is aiming it at cybersecurity and long-running work. Two caveats: Artificial Analysis measured its hallucination rate going up from 4% to 11%, and you can't use it yet. It's going to cyber defenders first, then trusted partners, and Google hasn't given a date for everyone else.
GitHub Copilot Fusion
With GitHub Copilot Fusion you pick Fusion and an effort level (standard, high, or max), and Copilot decides which models handle your prompt. It has three modes: single (one model does the job), cascade (a cheaper model drafts and the work moves up to a stronger model if the draft isn't good enough), and critique (one model drafts, another reviews, and the first revises). Artificial Analysis scored it 60 on coding agents, two points behind Claude Code with Opus 5.5, at about half the cost per task. It's generally available on all Copilot plans, and paid plans get 30% off model costs through November 30th.
Up to $500 a month in API credits
ChatGPT Pro and Business subscribers can now claim up to $500 a month in API credits. In ChatGPT, go to Settings → Billing and look for the API credits benefit. The credits fund one linked API organization and expire at the end of every billing cycle, so use them or lose them. The next day OpenAI also said it's doubling usage limits on the Pro and Business premium tiers. None of this applies to Plus.
Genie 4 and Hy-Nomas
Genie 4 generates a world frame by frame as you move through it. You can type what happens next and watch it change, and for the first time it makes its own audio. Gemini acts as the brain and Genie renders the picture and sound. It's a research preview for trusted testers. The one you can try tonight is Tencent's Hy-Nomas. It's Apache 2.0, runs on a single RTX 4090 according to Tencent, and its free web demo now turns an uploaded image into a world you can walk around in.
Rapid fire highlights
- Claude Ultra: Max 20x is now called Ultra ($200), and verified Ultra accounts no longer have the 5-hour limit. Weekly limits still apply. If Settings → Usage shows no current session bar, you're verified.
- Claude Code 2.1.311: sessions that run past 200K tokens now compact and keep going instead of erroring out. Early details can get lost in the summary.
- Anthropic's S-1: a proposed Nasdaq IPO. Share count and price range aren't set yet.
- Cursor Annotate: draw on your live page and send those marks to the agent as a visual prompt.
- GitHub: the Copilot app added Projects and computer use (public preview), and Agentic Workflows is now GA. You write repo automations in plain Markdown.
- Google: Gemini API Computer Use is GA. Prompt injection detection is off by default, so build in a confirmation step. The Gemini API CLI is in beta. Project Suncatcher got an AI response, a haiku, from a TPU in orbit.
- Also: Mistral Large 4 open weights, Kimi 3D Make, Idol (one photo to a 3D Gaussian splat), Pika H3 Flex, Runway App Studio, and Clicky drawing real-time graphics on your screen.
Trending GitHub repos
- agent-desktop (Vercel Labs): computer use that reads the app's accessibility tree instead of screenshots, and doesn't move your mouse. It's one npm install, and you can give it to your agent as a skill.
- Hallmark (Together AI): a design skill for Claude Code, Codex, Cursor, and OpenCode that checks pages at eight widths for signs of AI slop.
- DeepTutor (University of Hong Kong): drop in your PDFs and notes to get tutoring, quizzes, and now narrated video lessons. You bring your own model key.
- ThinkRail (JetBrains): an open-source IDE where each agent task gets its own git worktree. Basic usage events are always sent.
- OpenMausBot: a chat app where each bot is a real agent (Claude, Codex, or Grok) using the CLIs and logins you already have.
Key takeaways
- Computer use was the theme. OpenAI, Google, GitHub, and Vercel all shipped ways for AI to drive your apps in the same week.
- GPT-6.1 Terra is for cheap, repetitive agent work. Keep a stronger model for hard coding.
- Gemini 4 Carbon tops the index, but you can't use it yet, and its hallucination rate went up.
- Copilot Fusion trades two points on the coding index for about half the cost per task.
- If you pay for ChatGPT Pro or Business, claim your API credits. They don't roll over.
FAQ
What is GPT-6.1 Terra used for?
It's OpenAI's midsize model built mainly for computer use: high-volume tasks like filling forms, QA runs, and clicking through back-office apps. OpenAI still recommends GPT-6.1 Soul or GPT-6 Astra for the hardest coding work.
Can I use Gemini 4 Carbon yet?
No. It's going to cyber defenders first, then trusted partners, and Google hasn't given a date for general access.
How do I claim the ChatGPT Pro API credits?
In ChatGPT, go to Settings, then Billing, and look for the API credits benefit. The credits fund one linked API organization and expire at the end of each billing cycle.
What changed with the Claude Ultra plan?
Max 20x was renamed Ultra and is still $200 a month. Verified Ultra accounts no longer have the 5-hour limit, but weekly limits still apply.
Chapters
- 0:00This week in AI
- 0:48GPT-6.1 Terra (OpenAI)
- 2:40Shipping Skool
- 3:00Gemini 4 Carbon (Google DeepMind)
- 4:49GitHub Copilot Fusion
- 6:37Up to $500/month in API credits (ChatGPT Pro and Business)
- 8:02Genie 4 (+ Tencent Hy-Nomas)
- 9:37Rapid fire: Claude Ultra plan
- 10:13Rapid fire: Claude Code compacts past 200K
- 10:43Rapid fire: Anthropic files its S-1
- 11:08Rapid fire: Cursor design mode Annotate
- 11:38Rapid fire: Copilot app Projects and computer use
- 12:06Rapid fire: GitHub Agentic Workflows
- 12:38Rapid fire: Gemini API Computer Use
- 13:07Rapid fire: Gemini API CLI
- 13:43Rapid fire: Project Suncatcher
- 14:23Rapid fire: Mistral Large 4 open weights
- 15:01Rapid fire: Kimi 3D Make
- 15:38Rapid fire: Idol
- 16:13Rapid fire: Pika H3 Flex
- 16:43Rapid fire: Runway App Studio
- 17:15Rapid fire: Clicky real-time graphics
- 17:45Trending repo #5: OpenMausBot
- 18:21Trending repo #4: ThinkRail
- 18:56Trending repo #3: DeepTutor
- 19:26Trending repo #2: Hallmark
- 20:06Trending repo #1: agent-desktop
- 20:46Wrap-up
Transcript
Read the full transcript
0:00OpenAI just cut the API price of its computer use model in half. And the same week, Google took the number one spot on the artificial analysis intelligence index with a model you can't even use yet. Come on now. And that's just two of them. GitHub stopped making you pick a model. Open AAI is handing out up to $500 a month in API credits. Anthropic filed to go public. And somebody put an AI model in orbit and had it write a poem. So, in this video, we're going to be breaking down the 25 biggest AI releases from this past week, ranked five headliners, then a rapidfire round, then the five GitHub repos everybody was starring. And for every single one, I'm going to tell you what it is and if you
0:45can actually use it this week. Let's jump right into it. So, number one is GPT 6.1 Terra from OpenAI. It's their new midsize model and it's built mainly for one thing, computer use. So, what is computer use exactly? Your agent takes a screenshot and sends it to the model. The model looks at the screen and decides the next move. Open this, type that, click here. Your app runs that action, takes a new screenshot, and the loop starts over. Simple. And the catch has always been the cost because that loop eats a ton of tokens. So, OpenAI's pitch is that Terara is the cheap, fast model for running that loop all day long. They say it has half the standard API price of the last Terara. Press
1:30reports put it at $1 in and $5 out per million tokens. OpenAI says it finishes tasks over three times faster on average than GPT6 Astra and GPT6.1 Soul. And on one benchmark, Agents Last exam, they say it matches Astra at onetenth the price per token. And here's the one that got everybody talking. Open AAI says at max effort, Terra costs about nine times less per task than Claude Opus 5.5. Now, that's per task, not per token, and it's OpenAI's own number. Take that for what you will. So, can you use it this week? Yes. It's in ChatGpt Work and Codeex on Plus, Pro, Enterprise, and EDU, and it's in the API as GPT 6.1 Terra. Here's
2:17something to pay attention to. OpenAI still recommends Soul or Astra for the hardest coding work. So keep Soul for your hard coding. Terra is the one you run a thousand times a day. Form filling, QA runs, clicking through back office apps. For my resale business, that's the show numbers we type in by hand after every single stream. That's exactly the kind of job I'd hand to something cheap like this. Real quick, every video I make is built in code with Claude Code. Inside Shipping School Premium, I show you how step by step to turn your product into videos that get views, plus new courses every day, 500 plus hours of replays, live calls every week, and a Q&A with me every Saturday. Join free link below. So, number two is Gemini 4 Carbon from Google DeepMind.
3:04And this one is the new number one model on the artificial analysis intelligence index at 59. That's nine points higher than Gemini 4 Argon, which came out just 2 weeks before it. And Artificial Analysis checked it themselves. On their coding agent index, it's second only to Claude Opus 5.5. Carbon in the anti-gravity CLI scored 59. Claude Code with Opus 5.5 on Max scored 62. So, Opus still wins on coding agents, but Carbon costs about 23% less per task on their intelligence index. 776 a task versus 1004. So what is it actually built for? Google is pitching it hard at cyber security and long work. Find a vulnerability, check if it can actually
3:49be exploited, then write the fix. Google says it scored 88.2% on SEC Bench Pro. And they say their own teams have watched it keep working for days on hard problems. Those are Google's numbers, not independent ones. And they say successful cheating on Terminal Bench went down to 0%. Attempts to cheat were still 4%. So, it still tries sometimes. It just doesn't get away with it. Now, here's the deal breaker. You can't use it. It's going to a group of cyber defenders first, then trusted partners, and everybody else gets it as soon as possible. No date. When it does open up, it's $2 in and $10 out per million tokens to start. Then, 4 and 20. And one more thing, artificial analysis measured
4:34its hallucination rate going from 4% on argon up to 11% on carbon. They say that's still the second lowest of any model, but it went the wrong direction. I mean, number one on the index, and none of us can touch it. Google, come on. So, number three is GitHub Copilot Fusion. And I think this one matters more than people realize. You stop picking a model. You pick Fusion. You pick an effort level. And you can see them right here. Standard, high, or max. And co-pilot decides which models handle your prompt. So, how does it decide? It has three modes. Single, where one model just does the job. Cascade, where a cheaper model writes a draft, a quality check looks at it, and if it's not good
5:19enough, it gets sent up to a stronger model. And critique, where one model drafts, another model reviews it, and the first one revises. basically a code review between two AIs before you ever see it. And artificial analysis tested it. Fusion scored 60 on their coding agent index. Claude Code with Opus 5.5 on Max scored 62. So two points behind, but Fusion cost $559 per task versus $11.42 for Claude Code with Opus. That's about half the cost per task, which is insane. And per artificial analysis, it picks from models like Claude Opus 5.5, Sonnet 5.5, GPT6 Astra, and GPT6 Soul. So, can you
6:05use it? Yes. It's generally available on all C-pilot plans in VS Code, plus the C-Pilot app, the Copilot CLI, and Jet Brains. You pay each model's standard rate for whatever it uses, and paid plans get 30% off model costs through November 30th. Here's the catch. It does lag on one of their tests, SWE Atlas Q&A, where it scored 57, so it's not winning everywhere. But if you're tired of guessing which model to pick for every task, this is the easiest way to stop. So number four is money. If you pay for Chat GPT Pro or Business, OpenAI says you can now claim up to $500 a month in API credits to build apps and agents. And notice it says up to reports
6:51say pro gets 100,20 or 500 a month depending on your tier. Business gets $20 per standard seat or 100 per premium seat pulled up to 500 per workspace. Those tier numbers come from testing catalog not an open AI page. So take them with a grain of salt. So how do you get it in chat GPT? Go to settings then billing and look for the API credits benefit. The credits fund one linked API organization and they get spent before any credits you bought. And reportedly, your subscription has to be paid and at least 7 days old. Here's something most people don't think about. They expire at the end of every billing cycle. They don't roll over. So, if you don't use them, you lose them. And then the next
7:36day, OpenAI's Tibo Sotio posted that they're doubling all usage limits on the Pro 100, 200, and 500 tiers and on business premium. They also walked back a planned cut to Pro 200. Now, nobody said if the doubling is permanent, and if you're on Plus, none of this is for you. But like I said, if you already pay for Pro, this is free money to go build something. Go check if it's showing up for you and claim it. So number five is the best looking thing of the week. Genie 4 from Google DeepMind. Google calls it their most realistic and interactive world model yet. So what is a world model? You're walking through a world that's being generated in real time frame by frame as you move. And with Genie 4, you can type while it's
8:22running. Type what happens next and you can see the world change right in front of you. And for the first time from DeepMind, a real-time world model makes its own audio, sound effects, and characters that actually talk. And here's how it works. Your typed request goes to Gemini, which Google says acts as the brain of the simulation. Gemini turns your request into actions, plans what happens next, and watches the frames to keep track of things. Genie renders the picture and the sound. Techrunch says it can keep a scene consistent for up to 2 minutes. Now, can you use it? No. It's a research preview for trusted testers. It's coming soon to Google AI Ultra subscribers and to early access partners in the Gemini API and AI studio. And DeepMind says details can
9:07still drift over time. But here's the one you can touch today. 10 cents open world model highs. This week it got its technical report, a video recreating Genie4's own demos and on Halloween image to world in the free web demo. Upload a picture then walk around inside it. It's Apache 2.0 and Tencent says it runs on a single RTX 4090. Now, that comparison video is 10 cents owned, so it's not a fair fight yet, but it's free and you can try it tonight. All right, rapid fire round. Let's go. First up, Anthropic renamed its plans. Max 5X is now just Max, still 100 a month. Max 20X is now Ultra, still 200. And the big one, verified Ultra accounts no longer
9:53have the 5h hour limit. Starting with most existing subscribers, your weekly limit is still there, though. Ultra is 10 times pro every week, plus a separate weekly limit for Fable. To check, go to settings, then usage. If there's no current session bar, you're verified. And yes, I'm on the $200 plan. So, that's me. Next, a small Claude code fix that I actually care about. Before version 2.1.311, a standard context session that ran past 200,000 tokens could just stop with an error saying you need usage credits for 1 million context. Now, Claude Code compacts the conversation and keeps going. Now, compacting means it summarizes the older stuff, so some early details can get lost. It's not a
10:38bigger window, but no more dead sessions in the middle of a build. Update cloud code. and Anthropic publicly filed its S1 with the SEC. That's the paperwork for a proposed IPO on the NASDAQ Global Select Market under the ticker NT. And the number of shares and the price range haven't been set yet. So, anybody telling you a date or a price right now is guessing. There's nothing to use here. But if your whole stack runs on Claude like mine does, this is a big one to watch. Next, Cursor added annotate to its design mode. You can now draw right on your live page, edit elements, and check your design at different screen sizes. And those marks go to the agent as a visual prompt. So instead of typing, move that button a little to the
11:24left for the fifth time, you just circle it. Design mode itself isn't new. Annotate is the new part. Update cursor and try it in the browser inside the agents window. If you build frontends, this is going to save you a ton of back and forth. GitHub also shipped two things in the Copilot app. first projects. You give it a goal plus context like your repos and notes and Copilot runs several agents and brings back results like pull requests for you to review. Second, computer use. Copilot can now drive the apps on your Mac or Windows machine and it asks for your approval before it takes control of an app. Both are in public preview, so expect some rough edges. GitHub didn't say which plans get it. And one more from GitHub. Agentic Workflows is now
12:10generally available. You write a repo automation in plain markdown, triage new issues, explain why CI failed, update the docs, and it compiles into a normal GitHub actions workflow that runs a coding agent. It's open source from GitHub next, and it's available on GitHub free team and enterprise. But here's the catch. The default engine is the C-pilot CLI and that needs a C-pilot license or you can bring your own coding agent. Over to Google. Remember that computer use loop from Terra? Gemini's computer use tool is now generally available in the Gemini API. It works in a browser, on desktop, and on Android. And Google calls Gemini 3.8 Flash the recommended model for it. Here's something to pay attention to. Prompt
12:56injection detection is off by default. And when the model wants to do something risky, your app has to ask the user to confirm. So build that step in. Don't skip it. Google also put out the Gemini API CLI in beta. Google calls it a stateless zero dependency command line client for AI coding agents, scripts, and CI pipelines. So from your terminal or from cloud code, you can make images, speech, music, and video, transcribe files, and start Gemini agents. It's a oneline install and it uses your Gemini API key, so you pay normal API rates. And this is not the old Gemini CLI coding agent, different tool. If your agent needs a video or an image in the middle of a task, this is how you'd wire
13:42it in. And this one is just cool. Google's Project Suncatcher got the first ever AI response from a data center class TPU in orbit. The prototype satellite was built with Planet and launched October 1st on a SpaceX ride share with four TPUs on board. They ran Google's Open Gemma 4 model, asked it for a ha coup about being in space, and the poem came back down as the satellite passed by Swalbard. Blue marble below data streams in silent grace. Starlight feeds my core. It's not a data center yet. Next up is two linked satellites in 2027. I mean, this is a wild time to be alive. Mistl released the open weights of Mistl large 4. It's a one trillion
14:28parameter model, but it's a mixture of experts. So, for every token, a router picks four of its 128 experts plus one shared expert, and only about 49 billion parameters do the work. Mistl says it's effectively tied with GLM 5.3 on coding. Those are Mistl's numbers. You'll need a big multiGPU machine to run it yourself or use their API at 70 cents in and 210 out per million tokens. And Artificial Analysis says it's on the wordy side, so check your cost per task. Moonshot launched Kimmy 3D make running on Kimmy K 3.1 Pro. It connects to tools like Blender, Rhino, Frecad, Unreal, Unity, GDAU, and Roblox Studio to build 3D
15:14models, scenes, web games, and print files. You can see it here. Things with exact measurements like a desk lamp get built from code inside the tool. Organic stuff like people or animals goes to mesh generators like Tripo and Meshy. It's on Kimmy.ai now, and local apps need Kimmy work. And no, it doesn't print it for you. You still check the file and slice it yourself. Speaking of 3D, IDLE from Sparid Labs takes one picture of an object and turns it into a 3D Gaussian splat you can spin all the way around, including the sides the photo never showed. And just so you know, those hidden sides are generated. It's a guess. It's open Apache 2.0. It uses about 7.2 gigs of video memory, and
15:59there's a free hugging face space. Their own benchmark shows 3.4 4 seconds on a top data center GPU and about 19 seconds on an RTX 4090. Product photos to 3D. Very cool. Pico released H3 Flex. These are downloadable adapters on top of Miniax H3 that make 5 to 15sec videos with synced stereo sound in just six sampling steps. You can steer it with text, start and end frames, or up to 12 reference files. Now, it was tested on an H100 with 80 gigs. The output is 480p or 768P, and it's under Miniax's community license, so read the terms before you use it for client work. Runway launched App Studio. You describe
16:45the creative tool you need in plain words, and it builds that tool with Runway's models, your brand, and your files, and then hosts it for you. 3 days later, they added character swap. You can swap the characters or the location in a video and keep the original motion. And Runway Ads reads your best performing ads on Meta, Google, and Tik Tok and makes new versions for every placement. It's all on paid plans. If you run an agency, this is a custom in-house tool without writing code. And last in the rapid fire, Clickie. It's a Mac assistant that sits right next to your cursor and can see your screen. And now, Farza says, "Using Soul 6.1 Ultra Fast, it draws custom overlays, diagrams, and even interactive 3D scenes
17:30right on your screen as it explains things. No chat window." You can see it here. He's learning how hardware works, and the diagrams just show up on top of everything. Is this where AI interfaces go after the chat box? I think it might be. All right, the five GitHub repos everybody was starring this week. Counting down. Number five is Open Mousebot. The README calls it the open-source Grockbot alternative. It's a chat app where every bot in your sidebar is a real agent. Claude, Codeex, or Grock running on the CL is already installed on your machine using your existing login. No new accounts. Your transcripts and keys stay in a folder on your computer. It picked up about 4,800 stars this week, and it's sitting at almost 11,000. The readme itself says
18:17it's early but real. So treat it like early software. Number four is Think Rail, an open- source IDE from Jet Brains. It's built around the Pi coding agent, and every task gets its own workspace as a git work tree. So you can run a few agents side by side without touching your main checkout. Merge the good branch, delete the rest. And it keeps spec.md files next to your code that the agent keeps up to date as it learns your project. About 5700 stars this week. Here's something to pay attention to. Basic usage events and error reports are always sent. You can only turn off the extra sharing. Number three is Deept Tutor from the University of Hong Kong's data intelligence lab. It's a learning workspace where you drop
19:02in your PDFs and notes and get tutoring, quizzes, and learning paths. And version 1.7 added narrated video lessons that play right in the browser, a new notebook, and then flashcard export to an Docker command to run it, but you bring your own model key, so it's not free to run. It's at about 54,000 stars, and about 11,000 of those came this week. Number two is Hallmark from Together AI. It was the number one repo on GitHub's weekly trending list with over 11 12,000 stars this week. It's a design skill for claude code, codeex, cursor, and open code. And the readme says it refuses to look AI generated. It deals you three themes from a catalog of 50, builds the page, and then checks it
19:47in a real browser at eight different widths against the tells of AI slop, and then a fresh reviewer gets the last word. If you want to fix pages you already have, pair it with Impeccable, which is sitting at almost 100,000 stars. I'm tired of seeing AI slop websites. This is how you stop shipping them. And number one is Agent Desktop from Verscell Labs. It hit version 1.0 this week. And as I'm making this, it's the number one repo on GitHub's daily trending list with over 2,200 stars today. So remember all the computer use from this video. Terra and Gemini look at screenshots. This one doesn't. It reads the app's accessibility tree, the actual structure of every button and field on Mac, Windows, and partly Linux,
20:34and it works without moving your mouse. You can see codeex here playing the GarageBand piano with it. It's one npm install, and you can hand it to your agent as a skill. If you're not trying this one this week, I don't know what you're doing. So, that's the week. And if you noticed a pattern, it's computer use. OpenAI, Google, GitHub, and Verscell all shipped away for AI to drive your apps in the same seven days. So, which one of these 25 are you actually going to try first? Let me know in the comments. And if this saved you a week of scrolling, a like helps more people find it. We'll do this again next Monday. Have a blessed day.
Ready to start building with AI?
Join the Shipping Skool community free. Paid members get live bootcamp calls every week and weekly-updated courses.
Join Shipping SkoolKeep reading
AI News This Week: GPT-6 Cyber, ChatGPT GIFs, Claude Personal Finance
The big AI releases this week: OpenAI's GPT-6 Cyber wrote working root exploits for 948 Linux kernel bugs in a single run, Claude got read-only personal finance, and ChatGPT Images 2.6 makes GIFs. Gemini Omni 1.2 Flash and Meta's Muse Spaces also shipped, plus 15 smaller releases and 5 trending repos.
GPT-6.1 Terra vs Claude Sonnet 5.5: Coding Tests, Cost, and Speed
GPT-6.1 Terra costs half as much as Claude Sonnet 5.5 and finished each of my tests in about a minute. It tied Sonnet on a hard bug hunt, but Sonnet clearly won both visual builds. Use Terra for fast, cheap everyday jobs and computer use, and Sonnet when the result has to look great.
Claude Code GitHub Repos: Manim, Remotion, Blender MCP and More
Five free repos take Claude Code past writing code. Manim and Remotion Skills make videos, Lightpanda is a browser built for agents, MCP for Blender builds 3D scenes, and agent-desktop lets an agent use your whole computer. I'd install MCP for Blender first, and I'd only run agent-desktop in a VM or a separate user account.
