Congress Spends Almost 90% of Its AI Budget on One Chatbot.

1  Alibaba released a model it's calling an always-on "workmate," and it beat hundreds of human teams in a live coding competition.

The new Qwen-3.8Max is an open-weight model built for knowledge work, coding, and planning that spans multiple steps, and Alibaba says it goes toe-to-toe with the best US frontier models on raw performance. Per the lab's own claims, the model improved on an existing research paper's results in just 125 hours, coded without supervision for over ten straight days, and outperformed hundreds of human teams in an online competition. Try it yourself on QwenCloud, or read Alibaba's own guide to getting better prompts out of it.

2  A new disclosure shows Congress overwhelmingly prefers one chatbot over every other AI tool available to it.

ChatGPT accounts for nearly 90% of all AI spending among members of the US House of Representatives, with Claude a distant second. The timing isn't a coincidence: OpenAI, Anthropic, Meta, and Google are all meeting with the White House today to hash out a shared framework for testing and releasing new models. A lawmaker spending report landing the same day as a regulatory sit-down is the kind of detail that's easy to miss and hard to unsee once you notice it.

3  DeepSeek's cheapest model now rivals Claude Opus on quality while charging nearly 90 times less per token.

DeepSeek-V4-Flash is closing in on Opus 4.8's performance while charging 28 cents per million output tokens against Opus's $25. US labs are already responding to the sustained price pressure coming out of China: OpenAI quietly cut prices on GPT-5.6 Luna and Terra last Thursday, and Google shipped a leaner Flash model family earlier this month. What started as one lab undercutting one competitor is now visibly reshaping how every major US lab prices its budget tiers.

Yesterday's most-clicked story:  Sam Altman caught some heat online after suggesting people use ChatGPT to generate custom podcasts to entertain the kids on long car rides.

The premium is disappearing. Most of today's leading models can already draft a solid email or throw together a presentation competently. Unless the task genuinely calls for frontier-level reasoning, there's less and less reason to keep paying top dollar for it. Zack Kass, OpenAI's former head of go-to-market, has a name for this: "diminishing model returns," and some companies have already acted on it by cutting spend without cutting usage.

Companies are quietly routing around the expensive tier. Coinbase CEO Brian Armstrong said back in June that his team kept total AI spend flat while usage climbed, mostly by sending routine tasks to cheaper models and reserving frontier models for the work that actually needs them. He called out GLM 5.2 and Kimi 2.7 specifically as the cost-savers. Microsoft has clocked the same pattern and is weighing whether to add DeepSeek V4 to Copilot Cowork as a lower-cost option.

The whole thing rests on one shaky assumption. None of this price compression holds if performance stops being comparable across the tiers. The moment a frontier lab pulls meaningfully ahead again, the incentive to pay a premium snaps right back. And given how fast new models are shipping right now, that gap could open, or close, within a single week rather than a full product cycle.

What that means for how you should actually be buying AI. Locking into one model and sticking with it is the wrong instinct in a market moving this fast. The teams getting the most out of this moment aren't the ones that picked a favorite; they're the ones that built the habit, or the infrastructure, to shift workloads between models as pricing and performance keep moving underneath them.

😅 A little too convincing. Flux 3 has gotten disturbingly good at faking historical footage. The results are realistic enough that people are genuinely worried about what it could do to how we remember the past. Here's the clip that started the conversation (3.5M views).

🤣 Objects behaving badly. Redditors have been asking AI to render "completely serious" images of ordinary objects being used in absurd, wrong-context ways. The results are somehow both unsettling and genuinely funny.

🧠 One-word philosophy. Someone asked ChatGPT, Claude, DeepSeek, and Gemini the same question, name the biggest lie in human history in a single word, and got four completely different answers. The disagreement between models is the actual story here.

📚 Better than the textbook. A developer vibe-coded an interactive human anatomy site that reviewers say teaches the subject more effectively than any classroom textbook they've used (10,000 bookmarks). A small but real preview of what AI could do to education once it actually catches up.

🎨 Sticker upgrade. A viral logo prompt turns any brand mark into a glossy holo-vinyl sticker using ChatGPT Images, the kind of asset that instantly reads as more premium on a site or a social profile. Check out the original post.