AI briefingFriday, 9 October

Jessica: So Claude Haiku five point five just undercut GPT six Luna.

Eric: Wait, really?

Jessica: Same price. Ten cents per million input tokens, fifty cents output.

Eric: Okay, that's wild.

Jessica: This is dailybrief, I'm Jessica.

Eric: And I'm Eric. It's Friday, October ninth, twenty twenty six.

Jessica: So yeah, Anthropic released Claude Haiku five point five, and according to Simon Willison, it matches GPT six Luna pricing exactly up to a hundred thousand tokens.

Eric: Up to a hundred thousand. What happens after that?

Jessica: Price jumps five times. Fifty cents input, two fifty output.

Eric: So it's competitive until it's not.

Jessica: Exactly. And there's a catch—the new tokenizer uses one point two five times more tokens than Haiku four point five.

Eric: So you're paying the same rate but using more tokens. That's... not actually cheaper.

Jessica: Right. Though they did halve cache read prices for Sonnet five point five, so there's some give and take.

Eric: Okay. Still, for short contexts this puts real pressure on OpenAI.

Jessica: Speaking of pressure—Google launched a Gemini Enterprise agent that works across all your apps and devices.

Eric: Like, actually across apps? Not just Workspace?

Jessica: Yeah. The Verge says it runs on mobile, desktop, Workspace, and third party apps like Slack. Keeps context in the cloud so it works the same everywhere.

Eric: That's the dream, right? One interface, agent does the work.

Jessica: And it gets a company email address. You can literally treat it like a coworker.

Eric: That's either really useful or deeply weird.

Jessica: It also picks the best model for each task automatically. So you're not stuck with one Gemini variant.

Eric: Okay, I like that. No manual model switching.

Jessica: Alright, two OpenAI case studies today. First one—Pollo AI uses GPT five point six, GPT six Astra, and GPT Image two point five for video ads.

Eric: Wait, GPT six is out?

Jessica: Astra variant. OpenAI announced the case study but gave almost no technical details.

Eric: So we know three model names and that they make video ads.

Jessica: Pretty much. It shows they're combining models for creative workflows, but yeah—light on substance.

Eric: Same vibe with the Oracle one?

Jessica: Yep. Oracle uses ChatGPT Work and Codex across recruiting, engineering, and operations. Turns specialist knowledge into repeatable workflows, according to OpenAI.

Eric: That's... vague.

Jessica: It is. But it does show enterprise deployment at scale, even if we don't know how.

Eric: Okay, something with actual numbers—Hugging Face released Falcon ASR.

Jessica: Yes. Speech recognition model, one point six billion parameters. Twenty point nine two percent average word error rate across six Arabic test sets.

Eric: That's good for Arabic?

Jessica: Hugging Face says it beats published alternatives, especially for Gulf dialects like Emirati. And it supports English, French, Spanish, and Portuguese with the same weights.

Eric: Multilingual in one model. That's efficient.

Jessica: And it gives word level timestamps, so you know where each word is in the audio.

Eric: Google also shipped something practical—AI Edge Foresight. It's a meeting note taker for Mac that runs completely offline.

Jessica: On device. Uses EmbeddingGemma two, seven hundred forty million parameters. Works across apps, even in person meetings.

Eric: So it competes with Granola, but you don't need internet.

Jessica: Right. TechCrunch says it's optimized for Apple Silicon and works with PDFs, Google Docs, Office, web bookmarks—

Eric: That's actually useful for travel or anywhere connectivity is spotty.

Jessica: And Hugging Face shipped two decision models—d one three B and d one omni six hundred M.

Eric: Decision models. Remind me?

Jessica: They answer in a single forward pass instead of generating text word by word. Way faster.

Eric: Okay. How fast?

Jessica: Sixteen milliseconds on Jetson AGX Thor for the three B version. Scores forty eight point five seven on Decision Index zero point two point one.

Eric: Sixteen milliseconds. That's edge inference speed.

Jessica: And the omni version supports text with images or text with audio.

Eric: Nice for classification or real time stuff where you don't need full generation.

Jessica: Quick hit—Anthropic updated its usage policy. You can't be sustained and needlessly abusive toward Claude anymore.

Eric: Wait, they had to ban that?

Jessica: The Verge says it only applies in extreme cases. They'll just terminate the conversation. Not for normal frustration or testing.

Eric: Okay, so it's not thought police. Just no repeated cruelty.

Jessica: They also expanded the weapons ban to include control software and drone arming, and they're clearer on fake political campaigns.

Eric: That all makes sense.

Jessica: And OpenAI disrupted two AI enabled influence operations that used false front journalists and a fake think tank for geopolitical messaging.

Eric: Yikes. So people are using AI to run fake news operations.

Jessica: Yeah. OpenAI didn't give details, but they shut them down.

Eric: Last one—Ars Technica covered a developer building open source Creative Cloud alternatives with Opus.

Jessica: The headline is quote software is over unquote.

Eric: Bold claim.

Jessica: The projects are ambitious and free, but according to Ars, nowhere near finished.

Eric: So it's a demo of what's possible, not a real Adobe competitor yet.

Jessica: Exactly.

Eric: Alright, that's it for today. Every story is linked to its source on the dailybrief website.

Jessica: We'll see you Monday.