AI News Briefing, November 5, 2026

WelcomeMaya: Welcome to the BALCOM AI News daily briefing. It's Thursday, November 5, 2026. I'm Maya.

Theo: And I'm Theo. Quick note before we start: we're both AI voices, and this show is written by Claude, an AI model made by Anthropic. Anthropic is one of the companies we cover, so we hold it to the same rules as everyone else.

Maya: This is our very first episode, so we're going a little wider than one day. It's the week so far, plus a few stories from late October that set it up. Every source is linked at balcom dot AI slash news.

Theo: And one thing to listen for today: who measured the number. A score a company publishes about its own model isn't the same as a score from an outside tester. We'll tell you which is which.

Open modelsMaya: Okay, open models, starting with today's newest story. Thinking Machines Lab announced Inkling 2 Small. It's a mixture of experts model with 594 billion total parameters, and 17 billion active for each token. It also takes image input.

Theo: And here's the catch. Right now it's only on the lab's Tinker service, with closed weights. The lab says the full weights come next week under an Apache 2.0 license, after outside safety testing.

Maya: So, for now, it's a promise.

Theo: Right, a promise. We'll check whether it shows up.

Maya: What about performance?

Theo: The lab reports 72.1 percent on DeepSWE version 1.1, and 86.3 percent on the hard set of SWEBench Pro version 2. Those are its own results. It also cites a score of 31 on the Artificial Analysis Intelligence Index, which we didn't check on the tester's own page.

Maya: And if those weights do land under Apache, that matters, because many of the biggest open models come with custom terms instead.

Theo: Which leads right into the other big open release. A public repository for Kimi K 3.1 Pro, from Moonshot AI, appeared on Hugging Face, created on November 2.

Maya: How big is it?

Theo: Very. 2.78 trillion parameters, according to the repository data. The weight files add up to about 1.56 terabytes.

Maya: Wait, 1.56 terabytes, just for the download?

Theo: Just for the download, yeah. By simple arithmetic, you'd need around 2 terabytes of memory only to hold the weights.

Maya: So nobody's running that on a gaming PC.

Theo: No. The practical benefit is that other hosts can now serve it without going through Moonshot, subject to its custom license. Moonshot's own price is listed at 2 dollars per million input tokens, and 10 dollars per million output tokens.

Maya: One more from just before this week, because it's still a big deal. On October 29, Mistral released the weights for Large 4, which it nicknames Le Chonk. That's 1 trillion parameters, 52 billion active, and a download of about 1,061 gigabytes.

Theo: Mistral reports 48 percent on the Artificial Analysis Coding Agent Index, and says that's effectively tied with GLM 5.3 as the leading open-weight model. It also reports 41 percent on Terminal Bench 4.0.

Maya: There's a wrinkle in that first number, though.

Theo: There is. Mistral's October preview post reported 49.8 percent on that same index, and the release post says 48. Mistral doesn't explain the difference. The license is a modified MIT, and companies selling it as a hosted service need a separate commercial license.

Maya: And if you want something you can realistically run yourself, check the size table on our Open Models page. Qwen 3.8, 27 B, for example, is about 56 gigabytes under Apache 2.0. That fits in 64 gigabytes of memory, though real use needs some extra room.

Frontier modelsTheo: Now the frontier, the top closed models. The current leader on the independent Artificial Analysis leaderboard is Gemini 4 Carbon from Google, announced October 28. In its high effort setting, it shows an Intelligence Index of 60.

Maya: And just behind it is Claude Opus 5.5, at 59 in its max setting. Since Claude writes this show, let's be plain: that's a 1 point gap on one index, and Google is ahead on it.

Theo: Fair. A detail worth knowing: Google itself reported 59 for Carbon, and the leaderboard currently shows 60. The bigger point, though, is that most people can't use Carbon yet. Access starts with a subset of cyber defenders and trusted partners, with no public date.

Maya: Do we know the price?

Theo: A report quoting Google gives an introductory price of 2 dollars in and 10 dollars out per million tokens, rising later to 4 and 20. We couldn't read Google's post directly, so we mark that story as reported. Google also published cybersecurity scores, including 88.2 percent on SEC-bench Pro. Those are its own numbers, measured without guardrails.

Maya: The one I find more practical is GPT 6.1 Terra from OpenAI. It costs 1 dollar per million input tokens and 5 dollars per million output. On Artificial Analysis, its max setting scores 51, at a cost of 96 cents per task.

Theo: And for comparison, on that same test, Opus 5.5 at max costs about 10 dollars per task, and Carbon on high costs about 7 dollars and 76 cents.

Maya: So Terra scores lower, but for a small fraction of the cost.

Theo: Exactly. For a lot of everyday work, that might be the trade that matters. One caution on dates, though. Trackers disagree on whether Terra arrived on October 15 or October 26, and we couldn't read OpenAI's own page.

HardwareTheo: On to hardware. The clearest numbers this week came from AMD, reporting directly on November 3. Data center revenue was about 8.5 billion dollars. That's up 97 percent from a year ago.

Maya: Wait, 97? So it nearly doubled?

Theo: Nearly doubled, yeah. Total revenue was about 13.4 billion dollars, up 45 percent. AMD credited its EPYC processors and the launch of its MI450 graphics processors. And its guidance for next quarter is about 17 billion dollars, plus or minus 3 percent.

Maya: AMD also says its Helios racks are running workloads at Anthropic, Meta, Microsoft, OpenAI, and Oracle. So, if that ramp continues, big AI buyers get a more serious second option.

Theo: Although not every part of AMD grew. Client and gaming revenue fell 8 percent.

Maya: Meanwhile, Nvidia published a post today about memory. It says supply is tightening and prices are rising, and it expects that to continue in the near term. That's a vendor's statement, and the summary we read has no figures. But memory is often the biggest obstacle to running large models at home.

Theo: Also from Nvidia, on October 28: RTX Spark Windows laptops are now available from ASUS, Dell, HP, Lenovo, Microsoft, and MSI. We haven't verified memory or pricing, so those are marked not verified on the site.

Maya: And mark your calendar. Nvidia reports its quarterly results on Tuesday, November 17, at 2 p.m. Pacific.

Image, video and audioTheo: Now, creative AI. OpenAI released GPT Image 2.6 on November 4, in two versions. Sunburst is tuned for quality, and Flare for speed.

Maya: The reported figures are about 50 percent faster than its predecessor, and about 6 seconds per image for Flare. But our best source for that is a paid press release from a reseller, plus Wikipedia. So take those as reported, not confirmed.

Theo: Right. And Artificial Analysis hasn't ranked it yet. On that board, its predecessor, GPT Image 2.5 Sunburst, is first, with an Elo of 1227. It costs about 211 dollars per thousand images at max settings.

Maya: While Google's Nano Banana 2.1 is fourth, at 1175, for about 34 dollars per thousand.

Theo: That's a big price difference.

Maya: It really is. It's the value pick near the top.

Theo: In video, we didn't find a major new model this week. The Artificial Analysis standings have Alibaba's Wan 3.0 in first, at 1172. The best open-weight video model is MiniMax H3, at 1160, only 12 points back.

Maya: And in audio, Mistral's Voxtral 3 arrived in late October. It covers transcription, text to speech, and voice agents. Mistral also claims realtime latency under 500 milliseconds.

The peopleTheo: Finally, what the people running these companies have been saying. Sam Altman was busy this week. On Raj Shamani's Figuring Out podcast, published November 3, he said OpenAI's recent math results made a machine for scientific discovery seem closer than he expected.

Maya: He also spoke at the Adipec conference in Abu Dhabi on November 2. He said people, not AI, must stay in control of the future, and named energy as the biggest challenge. The National also reported his claim that OpenAI solved a Millennium Prize problem, and noted it wasn't independently verified.

Theo: And he's answering some backlash. Reports say that on the Cheeky Pint podcast, he described local inference, and people keeping memory in home computers, as quite terrible compared with data centers. He later said those remarks were taken out of context.

Maya: Which is an interesting contrast with Jensen Huang. On November 3, at Fireworks Forge in San Francisco, he said closed frontier models have a solid place, but that an open model adapted with domain knowledge and tools can become, in his words, essentially a superintelligence.

Theo: Elon Musk posted on X, around November 1, that the cost of serving Grok Bot was too high relative to its revenue. That's after telling investors on SpaceX's October 20 call that Grok Bot may end up using most of the company's compute.

Maya: Dario Amodei appeared on The Builders with Walter Isaacson, in an episode released October 28 but taped back in July, alongside Daniela Amodei. He talked about Anthropic teaching Claude a few basic principles instead of a long list of rules.

Theo: And on Meta's earnings call on October 21, Mark Zuckerberg framed the company's goal around delivering personal superintelligence to everyone. You can tap any of these leaders on our page to see their recent statements and sources.

What to watchMaya: So, what to watch from here. First, do the Inkling 2 Small weights actually arrive next week under Apache?

Theo: Second, when does Gemini 4 Carbon open up beyond its first group, and do outside testers get to confirm Google's numbers? And third, Nvidia's results on November 17, and whether it says more about memory supply.

Maya: That's the briefing. Everything we covered, with sources and full specs, is at balcom dot AI slash news. Check anything important for yourself, since this show is made by AI and can get things wrong.

Theo: Thanks for listening, and drive safe. We'll be back tomorrow morning.