Claude Opus 5 lands at half the price, Google teases Gemini 4, and Kimi K3 opens its weights

Every Tuesday I round up what actually happened in AI the week before, filtered through one question: does it change what I do Monday morning. Not what sounds big in a headline.

This week the news was the price tag, not the benchmark. Here is what mattered.

Claude Opus 5 shipped, and the frontier tier stopped being the expensive tier

On July 24 Anthropic released Claude Opus 5 across every platform, API included. It posts record coding scores and near top of the pack intelligence. The number that matters is the price. Opus 5 costs the same as Opus 4.8 did, five dollars per million tokens in and twenty five out, which is half the input price of Fable 5, the model that sat above it. It ships with a one million token context window and a low, medium, high effort dial so you pay for deep thinking only when you ask for it.

What it means for builders: the good coding model used to be the one you rationed. Now the near best one costs what the mid tier cost last month. If you picked a cheaper model earlier this year to keep the bill down, that trade is worth re-running. The gap between good enough and best just got cheaper to cross.

Google shipped the workhorse and teased the flagship

A few weeks ago the Google story was a rumor about them tearing down their next model and starting over. This week they shipped. On July 21 Google released Gemini 3.6 Flash, plus a cheaper Flash-Lite tier and a gated security model, and teased Gemini 4 in the same breath. Flash is the volume model, a dollar fifty in and seven fifty out, faster than the version before it and burning about 17 percent fewer tokens to finish the same job. It landed in GitHub Copilot on day one.

What it means for builders: most real work is not frontier work. It is a thousand small calls where speed and cost decide the bill. A faster, cheaper workhorse that finishes in fewer steps is a bigger deal for a running system than a flagship you call twice a day. And the Gemini 4 tease tells you the top end race is not slowing down. Build so you can swap the engine without rewiring the car.

Kimi K3 and the fine print on open weights

Moonshot put Kimi K3 up as a hosted model on July 16 and promised the full open weights by July 27, which was yesterday. It is a 2.8 trillion parameter model that only lights up a sliver of itself per token, so it runs lighter than the headline number. It ranks in the top five on one popular intelligence index and took first place on a front end coding leaderboard, above Fable 5. The catch is the wallet. Even before the weights drop, the hosted price is real money, three dollars per million fresh tokens in and fifteen out.

What it means for builders: open weights and cheap to run are not the same sentence. A 2.8 trillion parameter model is free to download and expensive to host. Go and look at whether the weights actually landed, but price the hardware before you assume open means cheap. For most people, calling a hosted model is still the rational move.

What I did about any of it

Seven serious models shipped in the seven days I am writing about. I did not touch my stack for a single one.

My pipeline shipped two new calculators, a cluster of three practice exams for electricians, a security fix on the signup form, and an accessibility check that now fails the build if a page ships broken. None of it needed a frontier model. It needed a cheap one that does the job and a set of checks that catch it when it is wrong.

A cheaper Opus is a nice tailwind. It is not the thing that let me ship safely this week. The test that fails the build when the math is off did that. The robot that presses the button before anything goes live did that. Better and cheaper models raise the ceiling on what one person can build. They still do not look at what actually went out. That part stays the job.

More next Tuesday. If someone forwarded you this, subscribe at blog.jaredhebb.com/newsletter.

Jared