Benchmarks Misbehave, Prices Move, and Databricks Buys Again
Grok 4.6 Has an Effort Problem
Oh, this is deliciously inconvenient. The headline floating around the DeepSWE board is that Grok 4.6 /medium outperforms /high effort. That’s not just a leaderboard quirk; that’s a tiny grenade rolled under the whole “more thinking equals better answers” sales pitch.
One camp will say: relax, benchmarks are weird, single results can mislead, and DeepSWE is only one arena. Fair. But the other camp has the sharper TV line: if “high effort” costs more time, more compute, or more patience, shouldn’t it actually win? If the middle setting beats the fancy setting, users are going to ask why they’re being nudged toward the expensive button with the serious-sounding label.
The stakes are simple: model makers want effort knobs to feel like premium control panels. Skeptics see them as vibes with a price tag. Today, the skeptics get the better monologue.
DeepSeek Turns the Pricing Screw — and Opens the Toolbox
DeepSeek has two fresh sparks on the table: a GitHub project called DeepSeek Harness and an API Pricing Update. That combination is catnip for the AI crowd because it hits both sides of the developer brain: “Can I test this properly?” and “What’s this going to cost me?”
The debate practically writes itself. Builders want predictable pricing and clean tooling. Rival model shops don’t want DeepSeek to become the default cheap-and-serious option. And finance teams? They’re standing behind every engineer whispering, “Please don’t surprise me with another mystery bill.”
No, we don’t have to pretend a pricing update is glamorous. It’s not. It’s spreadsheets in a trench coat. But pricing is where AI hype gets mugged by reality. If DeepSeek can make the cost story clearer while giving developers a harness to poke, prod, and compare, that’s not confetti — that’s pressure.
Databricks Buys Electric, Because Apparently the Shopping Cart Still Has Wheels
Then there’s Databricks buying Electric, per Blocks & Files. Cue the industry groan: “Oh no, not another one!” The headline says the quiet part loudly enough.
The argument here isn’t whether acquisitions happen. Of course they do. The fight is whether the AI-data stack is consolidating into fewer, heavier platforms while everyone else gets squeezed into feature status. Buyers may like one throat to choke. Startups may see an exit ramp. Customers may wonder whether every useful tool is destined to become somebody else’s menu item.
Winner today? The giants with checkbooks. Loser? Anyone still pretending the AI infrastructure market is a quaint little neighborhood.