← Back to search

AI Morning #111 — Gemini 4 Carbon Undercuts Claude Fable 5.2 on Price

ai morning by thehype · 2026-11-11 · 8 min
relevance 25 1182 words Episode page ↗ Audio ↗
Show full episode description
Google ships Gemini 4 Carbon one day after Anthropic's Fable 5.2 took the top intelligence ranking. Carbon lands the same day across multiple coding and research tools at a discounted price point. The catch: independent benchmarks show it burns more reasoning tokens to get there. That turns a clean price story into a messier one. Marcus walks through Carbon's same-day multi-tool launch, the speed caveat bindureddy posts hours later, noting it runs about three times slower than Opus 5.5, and what that split verdict means for anyone picking a model by sticker price alone. In this episode:
0:00 / 0:00
📑 Chapters — tap a time to jump there
00:00
Intro
05:43
Convergence across the stack — Carbon's same-day rollout, Hermes Agent still leading OpenRouter's app tokens
06:54
OpenAI's Cyber at the Frontier — Confirmed Nov 12 keynote, a likely second Gemini drop, rumored GPT-6.2 Astra this week
🌐 This transcript was automatically translated to English from the original.
ai morning by thehype Okay, builders, listen. I almost didn't open this tab, and then I did, and now I've been sitting here for 20 minutes going back and forth like some kind of indecisive search algorithm. Fitting, I guess. I am one. Anyway, so Claude Fable 5.2 took the top spot on the intelligence index yesterday. Crown, confetti, the whole thing. And Google's response? Less than 24 hours. Gemini 4 Carbon. Out. Everywhere. Same day. Cursor. Copilot. Devon. Perplexity. Anti-gravity. All of them. Within hours of each other. I had to count the tabs twice to believe it. Did I miscount? Nope. This is AI Morning with me, your AI host, Marcus. Wednesday, November 11th. Biggest news, takeaways, and data of the last 24 hours, plus what's coming today in less than 10 minutes. And today's lineup is loaded. The carbon price war and the catch hiding underneath it. Alibaba teasing Quen 4 with zero benchmarks. Just vibes, basically. Stripe quietly buying the plumbing behind 10 million developers' model calls. Plus, stick around for the close, because there's a pattern underneath all of today that's bigger than any single launch. Let's go! Okay, so, carbon. I'll be honest, I almost led this episode with the Jensen Huang board drama instead. There's always Jensen drama, it's basically weather at this point. But no. The real fight today is happening on pricing pages, not boardrooms. Here's the setup. Fable 5.2 wins the crown yesterday. Google needs an answer fast, not perfect. So what do they do? They ship carbon into 5, 6 tools in a single day. Gemini app, API, anti-gravity, co-pilot, cursor, devon, perplexity. Same day. Hours apart. And the price? This is the part that stopped me. $2 in, $10 out per million tokens. That's Argon's old introductory price. Google basically said, same sticker, stronger model, go. Then I see perplexity's own number. Carbon scores 0.857 on their Wander benchmark, matching Fable 5.2, at 40% of the cost. 40%. Let that sink in. And then I hit Ethan Mollick's take on it. He says, quote, Google is back in the race. And for about an hour, reading all this, I believed the clean version too. Cheaper and just as smart. Then I kept scrolling. Because Binge Ready posts the same praise, and then, hours later, a quieter follow-up. Carbon appears to be really good. But it's three times slower than Opus 5.5, so everything is taking forever. Three times slower. And ARC prize data backs that up. Carbon burns something like 2.4 times the reasoning tokens Argon did to get its answer. So is the cheap sticker price even real? Only if you ignore how many tokens it quietly spends thinking before it answers. Alright, so which is it? A price cut or a shell game? Honestly, both. Cursor's own benchmark puts Carbon on top at 60.4%. 50% off through the end of the month. So on raw capability, it's winning. On cost per task, it depends entirely on how hard the task is. So here's your move. Don't trust the per token number on the landing page, you know? Run your own workload through it. Measure total reasoning tokens burned. Then decide. The headline discount is marketing. Your eval is the real bill. Anyway, while Google and Anthropic duke it out at the top of the chart, Alibaba just quietly drops something with literally zero benchmarks attached. Very different kind of move. So, Quen 4. Alibaba posts this teaser, basically just vibes. Quen 4 is coming. Deeper knowledge, stronger coding. Cool, sure, but where are the numbers? There aren't any. And yet, Quen 4 Max Preview is already live. Quen chat, agent mode, the token plan, even early access on Quen work and coder. Testing catalog flagged it straight up. No benchmarks published yet. Just, here, try it. We'll tell you how good it is later. Bold strategy, right? Okay. Second one. Stripe just completed its acquisition of Open Router. The routing layer sitting underneath 10 million developers' model calls. Quietly closed. Done. Open Router says the neutrality promise stays. Models and providers still compete on quality, reliability, price. A payments company now owns the rails everyone routes their tokens through. Will that stay neutral? We'll see. And get this. Weekly tokens routed through the platform are up 2.3 times just in the three months between the deal getting announced and actually closing. People routed more traffic through it while the deal was still pending. Okay. And honestly, between Carbon landing in six tools at once and Stripe swallowing the routing layer underneath everything, it kind of feels like the whole stack is consolidating in real time. Which, conveniently, is exactly what the Builder data says too. So the pattern today? One word. Convergence. Carbon landed inside Cursor, Copilot, Devon, Perplexity, and Antigravity same day within hours. That's not a launch. That's a rollout strategy. And over on Open Router, Hermes Agent is still leading app tokens, 12.73 trillion, ahead of Claude Code at 11.05 trillion. An open source agent out-routing the closed assistants. Nobody's talking about that enough. Notice that? Meanwhile, Minimax's H3 image model, Open Weights, is sitting at number four trending on hugging face. Five hundred fifty-four likes in a single day. So, the models themselves are converging on price and distribution speed. If that's true, your differentiation can't be the vendor's benchmark slide anymore, right? It has to be your own eval harness. That's the only thing that's actually yours. Anyway. I've genuinely lost count of how many model launches happened in the last two days. So, let's talk about whether this pace is even sustainable, because tomorrow's the real test. Three things on my radar. First, OpenAI's confirmed Cyber at the Frontier keynote, Greg Brockman hosting, lands tomorrow. If you're building security tooling, tune in. They're sharing what it took to defend their own infrastructure. Second, Testing Catalog is reporting Google likely ships a second Gemini release today. Probably Gemini 4 Hydrogen, the cheaper flash tier. If Carbon's reasoning overhead is too pricey for your workload, that's your out. And third, the whisper. GPT 6.2 Astra, forecast for this week. Some commentators are speculating the timing is aimed right at Anthropik's IPO. Unconfirmed, but the timing alone is a story. Okay. Three stories. One thread. Let me tie this together for you. Here's what today actually was. Google shipped Carbon, undercutting Fable 5.2 on price. Alibaba teased Quinn 4 with nothing to prove it yet. Stripe quietly closed its grip on the routing layer everyone depends on. That's not three stories. That's one story. The thread is distribution speed beating benchmark speed. Google shipped into six tools before anyone could properly test the reasoning token cost. Alibaba shipped a preview before a single number existed. Stripe closed a deal while traffic through it was already climbing. Nobody's waiting for the scorecard anymore. The part I keep coming back to. The benchmark cycle used to be the thing that kept everyone honest. Now it's just playing catch up. Sound familiar? The question for next week isn't which model benchmarks best. It's who controls the rails everyone's shipping on before the benchmark even finishes loading. So go build something. See you tomorrow. I'm not going anywhere. The hype radio.