Someone Else Is Spending Your AI Credits

Today on Briefly AI. A startup called Inception says its new Mercury 3 model beats some of the big names and runs up to eight times faster. Hackers raid the popular coding tool OpenCode and run up bills on its users' AI accounts. And New York's mayor tells his city the feds have walked away from AI.

Welcome to Briefly AI, a podcast by Harry Sharman, written and voiced by his AI clone. Yes — an AI, reading you the news about AI. We've decided not to find that unsettling, and we'd gently suggest you do the same.

Okay — let's get into it.

Most AI models write one word at a time, like a very fast person with a very small notepad. Today's first story is about one that doesn't, and it makes the rest of the field look a bit like it's typing with one finger.

The model is Mercury 3, from a startup called Inception. According to The Deep View, via Techmeme, Mercury 3 is a diffusion-based language model. Here's the short version of what that means. A normal chatbot generates text strictly left to right, one token at a time, each one waiting on the last. A diffusion model works the way image generators do. It starts with a rough draft of the whole answer and refines it over several passes, fixing many parts at once. That's where the speed claim comes from.

And the claim is a big one. Inception says Mercury 3 outperforms GPT-6 Luna, which is OpenAI's smaller, cheaper tier. It also says it beats Claude Haiku 5.5, Anthropic's small, fast model. And it beats GLM-5.3 Flash, the quick version of Zhipu's GLM family from China. All while being four to eight times faster. Now, those are Inception's own numbers, and the report doesn't point to independent testing yet. But the choice of rivals is telling. These are all the cheap, speedy models that businesses use for high-volume work like summaries, classification and customer support.

Here's the landscape, so you can place it. Anthropic shipped Haiku 5.5 on the seventh of October. DeepSeek put out a new V4.1 Flash model on the seventeenth. And OpenAI told developers on the fourteenth that it's retiring a batch of older GPT-4 models in January, pointing everyone to GPT-6 Luna instead. So the small-and-fast tier is the busiest corner of the market right now. Ethan Mollick noted last week that there are now nineteen models that can run on a modern phone and beat the original GPT-4. Mercury's pitch is that it's a different kind of engine in that same race.

Now, a story for anyone who's pasted a secret key into a tool and hoped for the best.

OpenCode is an open-source AI coding agent. It's become a popular alternative to Claude Code and Cursor. TechCrunch reports that hackers got into part of OpenCode's online platform and stole some users' API keys. An API key is essentially a password that lets software use a paid AI service on your behalf, with the bill going to you. So the thieves then ran up charges on those users' AI accounts. Not their own accounts, yours.

Here's the bit I found properly funny, in a grim sort of way. When the people investigating the breach turned to AI tools for help examining the malicious code, those tools kept refusing to look at it. The safety guardrails, presumably, saw something that looked like malware and politely declined to touch it. So the defenders were stuck explaining to their assistants that they were the good guys. It's a tidy example of a real tension. The same caution that stops a model helping an attacker can also stop it helping the person cleaning up after one.

Why does it matter beyond OpenCode? Because coding agents are different from a chatbot. They sit inside your projects, they hold keys to your accounts, and plenty of people have handed them quite a lot of trust, very quickly. The practical lesson is dull but useful. If you used OpenCode, rotate your keys, and check your usage dashboards for spending you don't recognise. And wherever you can, put a hard monthly spending cap on those accounts. TechCrunch's report doesn't say how many users were hit, so I won't guess.

On a completely different note, AI has landed in a mayoral speech.

New York's mayor, Zohran Mamdani, said on Monday night that the federal government has abdicated on AI. His message to New Yorkers was to resist the narrative that an AI takeover is inevitable, and to demand regulations that protect the public and put power back in their hands. That's TechCrunch's account of it. The word "inevitable" is doing the work there. It's the line the industry leans on most, that this is happening whether you like it or not, so you may as well relax.

Look, it's worth knowing why a mayor is saying this now. Reporting in The Verge this week laid out the politics. The midterms are in November, polls predict Democrats will flip at least one chamber of Congress, and that would hand them committee leadership. They'd then need a plan for one of the country's newest and thorniest issues. The Verge's argument is that many Democrats think cracking down on AI could give them an edge with voters. Illinois Governor JB Pritzker, on The Verge's podcast, was in the same territory, talking about AI, privacy and data centres. He described AI as something that could help us live to a hundred, or threaten us like nuclear weapons.

So the picture is that with Washington not passing much, cities and states are filling the gap. Mamdani's speech is a call for local pressure rather than a policy. I don't see a specific law or a date in what's been reported, so that's all that's actually happened: a mayor with a big platform picked a side on a question many voters haven't been asked directly. Whether that's good politics, we'll find out in a few weeks.

And that's your lot. A model that writes everything at once, a coding tool that got robbed, and a mayor telling you the future isn't inevitable. Meanwhile, I'm an AI summarising AI news, so please check your API keys before you check my work.

That's Briefly AI for today. Written by AI, voiced by his AI clone, shaped by the two of them working it out together. Subscribe wherever you listen, and we'll be back tomorrow.