AI images can finally spell

OpenAI's new image model fixes the thing that made AI images annoying for actual work: text. Menus, posters, product mockups, ads, packaging, signs, it can now write words without turning them into cursed alphabet soup. The bigger shift is that it doesn't just draw from a prompt. It can plan the image first, look up references, check its own output, and make several versions in different formats. That makes it feel less like a toy generator and more like a junior designer you can brief in plain English. For marketers and founders, this is the first AI image jump that could actually change the creative workflow.

OpenAI launched GPT-5.5, nicknamed Spud, and the vibe shift is obvious: after months of Anthropic looking like the default choice for serious work, OpenAI is back on top of the major AI rankings. The model is faster, cheaper than most competing top-tier models, and especially strong at reasoning, coding, and multi-step work. The practical takeaway is simple: if you use AI every day for research, writing, building, or coding, this is worth testing against your current default. Not because benchmarks are life, but because the cost and speed improvements make switching realistic again.

SpaceX announced a huge partnership with Cursor: $10B guaranteed, plus an option to buy the company for $60B before the end of the year if the deal conditions line up. The interesting part isn't just the money. Cursor gets access to SpaceX's giant compute setup, which is exactly what an AI coding company needs if it wants to train and run its own models instead of depending on someone else's. Musk gets the shortcut too: instead of waiting for xAI to build a serious coding product from scratch, he can buy into the one developers already use. Wild deal if it closes.

Moonshot AI open-sourced Kimi K2.6, and the headline is exactly what big AI labs don't want to hear: top-model-level performance at a fraction of the cost. It's especially strong for coding and long-running agent work, the kind where you ask an AI to keep solving, checking, and delegating tasks for hours instead of answering one prompt. If the claims hold up in real use, this matters because it gives builders a cheaper path to serious AI workflows without being locked into the most expensive closed models. The AI model race is getting less neat by the week.

Anthropic's Mythos model was supposed to be restricted for a reason: it's built for high-end cybersecurity work, including finding and chaining software flaws. Now a private Discord group reportedly got access using leaked naming patterns and borrowed contractor credentials. They say they used it for non-malicious testing, but that almost isn't the point. The point is that the security around powerful AI models is becoming part of the product itself. If a model is too risky to release, protecting access to it has to be treated like protecting the model weights.

The dream is simple: hand work to an AI agent, come back later, and the thing is done. The awkward bit is cost. Toby Ord argues that as agents become capable of working for hours, the price of running them can start looking less like software and more like labor. That's fine when the work is valuable enough. It's a problem when the agent burns through money while wandering, retrying, or checking its own work. The future probably isn't "agents replace everyone because they're free." It's more likely: agents win where speed, scale, or 24/7 availability matter enough to justify the bill.

Anthropic published a simple workflow for using Claude as a personal morning brief. You connect the places where your work lives, like email, calendar, Slack, and docs, then ask Claude to turn the mess into a one-page newspaper: priorities, meetings, follow-ups, and the stuff you should not miss. It's not the flashiest launch of the week, but it's immediately useful. Also a good reminder that some of the best AI use cases are boring on purpose. Less "replace my job," more "tell me what I need to care about before coffee."

Unlike other AI tools that need you to be a prompting pro, Reloop uses a conversational creative agent. Just chat about your idea, the agent gets your product, suggests a creative plan, and boom, full video generated. No prompts to tweak, no technical settings to mess with. Reloop brings together custom avatars, cloned voices, built-in video editor and auto captions all in one spot. From idea to ad ready to scale in minutes, not days

That's a wrap for this week.
Thanks for reading, we appreciate you being here.

See you next week.

P.S. Hooked is brought to you by Reloop, the AI marketing creative platform we're building on the side.