Tech Brew Ride Home - Le Chonk

Episode Date: October 6, 2026

Nvidia-backed Reflection unveiled Beam, a US open-weights model taking on Chinese labs. Plus, SemiAnalysis said Claude subscriptions were a better deal than OpenAI's, Meta and Microsoft cut back on Claude, Mistral dropped le Chonk, and ChatGPT faked New Yorker cartoonists' signatures. NYC-based Reflection unveils Beam, an open model it says rivals GLM-5.2 on reasoning while using 3x-4x less compute and approaches Qwen3.8-Max on agentic tasks (Semafor) Nvidia-backed Reflection pitches Beam to businesses and developers as a cheaper option for coding and agent tasks; the startup has held talks to raise $2.5B at a $25B valuation (Bloomberg) Analysis: for agentic workloads, Anthropic's plans with Claude Opus 5.5 offer ~5x more API-equivalent value per month than OpenAI's plans with GPT-6.1 Sol (SemiAnalysis) Sources: Meta and Microsoft are working to cut internal Claude use; Meta employees using Claude Code dropped to ~30,000 from ~60,000 earlier in 2026 (The Information) Mistral releases Mistral Large 4, dubbed “le Chonk”, a 1T-parameter open-weight model for general agentic capabilities, trained on 4,000 Grace Blackwell GPUs (The Deep View) ChatGPT is not only using New Yorker cartoonists' style but also adding their signatures to the fake cartoons it creates, which have gone viral on social media (Nieman Lab) Subscribe to the ad-free feed.

Welcome to the Tech Brew Ride Home for Tuesday, October 6, 2026. I'm Brian McCullough today. Reflection unveiled Beam, a U.S. Openweight's model for taking on Chinese labs, are Claude subscriptions are better deal than Open AIs. Meta and Microsoft cut back on their Claude addictions. Mistrel dropped Le Chonk, and ChatchipT is faking New Yorker cartoonist signatures. Here's what you miss today in the world of tech. Every day, shareholders meet to discuss important matters about the companies you invest in. Now you can make your voice heard too. Vanguard investor choice makes it easy to set your proxy voting preference for your eligible

Vanguard index funds. Whether you hold a Vanguard fund directly or through another brokerage firm, all it takes is a few clicks to select your proxy voting preference and be heard on important shareholder topics like executive pay and director elections. Visit vanguard.com slash investor choice to learn more. It's your shares. It's your voice. It's easy.

Vanguard investors own shares of our index funds and those funds own shares of the companies they invest in. Vanguard Marketing Corporation distributor. Well, here it is, people. We have a U.S. Openweight's model and company. New York City-based reflection has unveiled Beam, an open model. It says rivals GLM 5.2 unreasoning while using three to four times less compute and approaches Quinn 3.8 Max on agentic tasks, quoting Bloomberg.

The model called Beam is designed to perform well for business customers and developers on coding an AI agent has while costing less than many competitors the company said on Monday. It said Beam scores comparably on advanced reasoning benchmarks to GLM 5.2, a flagship open model from China's ZAI. Reflection intends to release more technical details about the model later this month. While China is still widely viewed as behind the U.S. and AI, Chinese developers like Deepseek and Moonshot AI have gained ground by offering cheaper and customizable open models. Top U.S. firms, including OpenAI and Anthropic, have primarily focused on closed or proprietary offerings as they work to build up their enterprise businesses. Reflection emerged from stealth last year with a focus on developing OpenAI models.

The firm is backed by Nvidia and other investors. It has held talks to raise a $2.5 billion round at a valuation of $25 billion of Wall Street Journal previously reported, and quote. And quoting the information. Beam has about 500 billion total parameters, the numbers that are. make up an AI model and determine its behavior. That makes it about half the size of thinking machines labs openweight model inkling and around the same size as Nvidia's Nemotron 3 Ultra, which the chip giant released in June. Reflection says that Beam outperforms both models in tests of coding and reasoning capabilities, but comes in behind leading Chinese models. More Western companies have

turned to developing open weight models in recent months. Invidia is also working on a larger Nemotron model with assistance from a coalition of AI companies that includes reflection, end quote. And quoting Silicon angle. Reflection AI kicked off the development of Beam by training a relatively small prototype model. It then created a series of successively larger, more capable algorithms. The workflow eventually produced Beam Base, the foundation on which Beam is based. The company developed Beam Base using a cluster of 6,144 graphics cards. It trained the model on 23.8 trillion tokens sourced from the public web and commercial sources. The dataset included a significant amount of software code. It created custom filters for each programming language to remove

low-quality files. Beam Base was developed in under four weeks. It subsequently performed mid-training and optimization process that extended the model's context window and enhanced its reasoning capabilities. That phase of the project set the stage for the third most hardware-intensive phase of the training workflow. Reflection AI spun up 10,000 GB300 graphics cards and used them to launch 1.3 billion reinforcement learning sandboxes. Those are virtual environments in which an AI model learns new skills. Beam's sandboxes were optimized for tasks such as generating code, searching the web, and running AI agents. The company says the project's reinforcement learning phase took only four weeks. It avoided unnecessary delays by developing software that prevented malfunctions from interrupting the

workflow. It says its cluster achieved a median recovery time of eight minutes across the 71 errors that cropped up during training. Beam is initially available through an early access program. Reflection AI plans to release its weights, documentation, and fine-tuning tools later this month, end quote. If you subscribe to a monthly plan for an AI model, have you ever wondered what sort of deal you're getting over just, you know, hooking up to the API and buying token by token. Well, semi-analysis crunched the numbers and they say for agentic workloads, Anthropics plans with Cloud Opus 5.5 offer around five times more API equivalent value per month

than OpenAI's plans with GPD 6.1 sole. Quote, subscription plans are still the primary way consumers and small businesses pay for AI. These plans are highly subsidized, as we previously explained in June, but can still make economic sense as powerful customer acquisition and marketing tools. For example, the goodwill engendered by Open AIs generous Resets is partially responsible for the recent surge in Codex adoption and as forced Anthropic to repeatedly walk back planned subscription nerfs to avoid getting clobbered in the court of public opinion. Despite being just 10% of overall revenue, subscriptions can take up over 40% of inference compute

and lower blended revenue per megawatt by around 36 million. dollars. Subscriptions are even more important for OpenAI as they make up a larger portion of their total revenue. In other words, if you want to accurately model AI lab financials, you need to understand their subscription limits. So how do limits actually work? The way to think about subscriptions is that your monthly payment grants you some number of credits. Each model token type combo consumes a different amount of credits because credit cost ratios can differ dramatically from API price ratios. The value of the same plan changes depending on what model and workload you're running. Put differently, it doesn't make sense to say a plan is worth X dollars in isolation. You need to consider

the full plan model workload tipple. The biggest question in AI subscription land is how Anthropics limits compare to Open AIs. Note that Open AI just made some massive changes to their subscriptions last week, having the API equivalent value for the $200 plan and introducing a new $500 tier. Limits are quite similar across the board for GPT6 Astra versus Fable 5.1, though remember that Fable can only be used for 50% of your limit. In other words, your $200 ant plan would still have 50% left after consuming $2,485 worth of Fable 5.1, whereas the equivalent Open AI plan would be fully exhausted after $2,897 of Astra. This extra $5,485,000 worth of ASTRA. This extra $50,000, percent starts becoming very relevant when we compare Opus 5.5 versus GPD 6.1 sole. As this mid-tier of models,

which both companies market as the intended daily driver for most users, Anthropic is an overwhelmingly better deal, offering around five times the API equivalent value across the board. You could argue that this is unfair for Open AI because 6.1 sole is much cheaper per token than Opus 5.5, but the gap is still massive even if you switch to comparing the number of tokens. Our tracking confirms the 50% value reduction for OpenAI's $200 plan announced by Tebow. $200 plans purchased before the cut will keep the old higher limits until October 29th, but newly purchased plans immediately start with the lower limits. Anthropic, on the other hand, already offered the same per dollar value for all of their subscription tiers

and currently crushes Open AI. Open AI used to be lauded by indie developers for their generous subscription limits compared to Anthropic, but this is simply no longer true today. Opus 5.5 on any Claude subscription offers far better value than anything from OpenAI. The only counter-argument for Open AI is that none of their pro plans have a five-hour limit, which makes it easier to consume a higher percentage of your total monthly limit. However, we don't think this offsets the around four times higher API equivalent. value offered by Opus 5.5 on Claude plans. In other words, Anthropic would likely already have software-like margins on their subscription plans if everyone only used Fable. Opus class models will continue to become cheaper to serve over time as Anthropic figures out how to make the model smaller,

and new model tiers above Fable will have even lower subscription limits, assuming they're included at all. Open AI, on the other hand, pick the nuclear option of just immediately cutting to fable level limits across the board, end quote. Teams that struggle to scale aren't falling behind because they're lacking headcount or budget. It's because critical knowledge isn't documented. That's what today's sponsor, Scribe, was built to fix. When specialists or experts leave your team, their knowledge disappears too.

That's where Scribe comes in. Scribe is a specialized intelligence platform trusted by 94% of the Fortune 500. It captures workflows in real time. automatically generates the documentation for how your company works with no manual writing. It automatically redacts sensitive information like names and account numbers from screenshots. Plus, your team can launch real-time on-screen guidance showing exactly where to click, step by step, inside the actual tool. And these captured workflows give your AI agents real context,

so they can operate on how work actually happens, not guesses. Learn more at Scribe.com. slash ride home and mention TechBrew Ride Home for your first month of Scribe capture free on select plans. That's SCR-I-B-E dot how slash ride home. If you're still basing your pricing on seats, it's time to rethink how you determine pricing because of LLM and inference costs. The cost of serving a customer now depends on how much they use your product, not just how many people use it.

And your product teams are shipping new features, models, and AI agents that have paced legacy billing systems weren't designed for. Charge B can help. They provide metering, billing, and monetization infrastructure to support both seat-based and usage-based pricing models, so your engineering team doesn't have to build and maintain the whole thing from scratch every single time you ship a product, feature, or agent. Chat with an expert about pricing and monetization at chargeb.com slash ride home.

That's C-H-A-R-G-E-B-E-E-E-D-E-E-E-E-L-H-H-H-H-E. This is obviously related to that last segment. Microsoft are apparently trying to wean themselves off of using Claude internally. Remember how in the Anthropic IPO perspective, two customers accounted for something like 25% of their revenue. Again, one of those has to have been meta. Well, the information says that meta employees using Claude Code dropped to around 30,000 from around 60,000 earlier this year. Another reason why Anthropic needs to get this IPO out the door toot suite. Quote, quote, Microsoft executives earlier this year projected that the company was on track to spend at least

$1 billion on its own internal use of Anthropics technology, according to two people with knowledge of the spending. Microsoft has since lowered that figure by more than a third after it asked staff to use less clod to save on costs and spend more time using Microsoft's homegrown AI tools. One of the people said that figure doesn't include the money customers spent on Anthropics models through Microsoft's tools, which has risen steadily. At Meta, The number of employees using Anthropics Coding Assistant Claude Code has dropped sharply to about 30,000 recently, according to a person familiar with the matter. That's down from about 60,000 users earlier this year, according to internal data reviewed by the information. Part of the decline stems from layoffs in the spring that cut about 10% of Meta's than 78,000-person workforce.

But a bigger factor has been meta's development of its own AI coding tools and its Mews Spark line of AI models, which the company is marketing to other companies and pushing its own employees. employees to use. Muse Code, a Claude Code competitor that uses meta models, which Meta started testing with outside customers in August, had more than 60,000 employee users recently, according to the internal data. A similar tool for in-house use only called Meta code has more than 30,000 users. Meta had been on pace to spend billions of dollars this year on its internal use of AI tools, according to a June memo reported earlier by the information. The memo didn't specify how that spending was distributed among different companies, but Claude Code was by far the most popular

outside coding tool at the time. The internal data show that in a recent 28-day period, Meta spent more than $105 million on Claude Code. The two tech giants remain big-clod users, but even small reductions in spending from such mainstays contribute to broader headwinds for Anthropic, as surging costs for its AI tools have compelled other business customers such as cognition to look for cheaper alternatives, end quote. Sorry, but another important model just dropped that I need to tell you about Mistral has released Mistral Large 4, which they are calling Lechunk, a one trillion parameter open weight model for general agentic capabilities trained on 4,000 Grace Blackwell

GPUs, quoting The Deep View. On Tuesday, Mistral launched a public preview of Mistral Large 4 ML4, ML4 Lechonk. its new flagship 1-a-trillion-parameter open-weight model meant for general agentic capabilities. Mistral claims its competitive advantage lies in the fact that it is comparable to the best closed models while staying competitive with open-weight models three times its size. Meanwhile, it was trained on 4,000 Nvidia Grace Blackwell GPUs at what it describes as a fraction of its competitors' cost. With this launch, Mistral is also telling a cybersecurity story, not only claiming it is strong in both offensive and defensive, categories, but also capitalizing on the fact that it is an open-weight model. Mistral's co-founder and chief scientist Guillaume Lample told the deep view that the open-weight

nature of the model allows companies to safely build and deploy cybersecurity defenses without fearing that the model is suddenly deprecated or sundowned. Basically, new cybersecurity workflows require a lot of tokens, a lot of inference, so you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day or might be too limited, said Lample. Beyond cyber, the new model is proficient in other realms, including coding, domain-specific knowledge work, and multi-modality. Some other features, according to the blog post, include, based on its grounding capabilities. It outperforms all existing models, including the closed ones, Mistral claims, deployment on Mistral's own data centers in Europe, and served in preview on the infrastructure, and trained using reinforcement learning, and the team continues to see rapid progress.

This is notable, as many open models are mostly distilled, and the team is highlighting its own research, training, infrastructure, and expertise. The open weight model is aimed primarily at enterprises, deploying it on-premise or in a private cloud, but it will also be available through Mistral's API. At one trillion parameters, it's too large to run locally on a desktop or laptop and may be impractical for some universities to run themselves. The weights will be released October 27th, end quote. Finally today, according to Neiman Lab, ShatGPT is not only using New Yorker cartoonist's style to create cartoons, but is also adding their signatures to the fake cartoons it creates. Some of these have gone viral on social media.

At the pearly gates, Dolly Parton in a sparkling gown stands arm-in-arm with Tim Curry, dressed as Dr. Frankenfurter. I knew I chose the right plus one. Parton says to God, who is standing at Heaven's reception desk. That's the premise of a cartoon that went viral on Twitter in the days after the legendary entertainer's death in late August. One tweet of the cartoon received 25,000 likes, while other versions of the image circulated on Facebook, Instagram, and Blue Sky. In the bottom right corner of the cartoon is the pen name Blopper, the signature of New Yorker cartoonist

Brendan Lopper. But Lopper didn't draw the cartoon or sign it. It was generated by Chad ChiePT and first posted to Facebook by a Dolly Parton fan. In the comment, section of her post she wrote she had simply asked chat sh pt to make a new yorker style cartoon i had people writing to me about this cartoon asking if it was mine lapper said first came a text from his twin brother whose friend had seen the a i generated image emails and dms on social media from strangers followed it was very surreal he said this wasn't the first time chat chpt had reproduced lopper's signature in may a stranger reached out over instagram to warn lopper that his pen name had appeared in an image output by the chatbot

Over the summer, Lapper found additional examples in a Reddit forum about ChatGBTGBT, where users had been posting similar New Yorker style cartoons they had generated. Lopper already knew his work was out there getting harvested for all of this AI training, he said, but in image generator faking his signature felt more personal. I'm not a territorial person, but my name is my name. It felt very much like a violation of my personhood. In my tests, Chad ChbT regularly generated cartoons with the hallmark of the one. 101-year-old magazine's house style, a black and white single-panel illustration with a sparse background, clean linework, and a sardonic caption, or at least an approximation of one.

Some of Chat ChagpD's punchlines don't make sense. Not every image came with a signature, and some included pen names that were gibberish, but the signatures that did appear were regularly those of real contributors to the magazine, including older cartoonists like Warren Miller, who published in the 1960s. I can see the DNA of my work in there, but it still doesn't look like something I drew, said Emily Flake, who's been a New Yorker contributor since 2008. Several Chad ShepT-generated cartoons posted online are signed with her pen name E. Flake. It's kind of like seeing a distant cousin. The character stances and shading contained echoes of Flake's visual identity, but she said overall they looked more like amalgamations of several different cartoonists' illustration styles. Each signature, though, was undeniably hers. It's like somebody attributed a quote to me that I didn't say, end quote.

Thanks to those of you who gave us a five-star rating in Spotify or Apple Podcast yesterday. And a reminder, if you forgot, why not do it today? Again, it really helps the show out. Thanks.