Is AI slowing down?

Today’s issue: Lessons in etiquette, DHH’s complicated relationship to the Pope, and a VM you can definitely trust.

Welcome to #524.

---

Is AI “slowing down”?

After all the big labs agreed it was time to slow down the pace of AI development, OpenAI, Anthropic, and SpaceXAI all released models this week. It feels like telling your girlfriend that you want to take things slow only to find yourself walking into a Tiffany’s against your will.

So are we slowing down? Well kind of. On the surface it appears like the labs are coordinating their releases because they still gotta ship something to keep their IPOs on track. But for the most part these updates focused on optimizations and efficiency rather than new capabilities. Let’s take a look at the highlights.

Opus got smarter

Opus 5.5 “fixed” many of the issues that plagued the previous generation like the incoherent walls of text it would generate for pretty much any query. Its input and output tokens are 20% cheaper than Opus 5, and cache reads are 60% cheaper than Opus 5, despite beating Fable 5.1 on many benchmarks. The main thing you need to know is that it already performs well on medium reasoning, so try that before burning through tokens on xhigh.

Sol & Luna got cheaper

OpenAI released GPT-6 Sol & Luna, cutting API prices by about half compared with the promotional prices for their 5.6 counterparts. They also decided not to release GPT-6 Terra, probably because Luna performs just as well as GPT-5.6 Terra despite being significantly cheaper. Similar to Opus 5.5, Sol doesn’t give you too much of an improvement at higher reasoning levels, so if you’re not subscription stacking, use GPT-6 Astra if you need high reasoning.

Grok got stamina

Early reviews for Grok 4.7 were not great… It’s bigger than 4.6, and word on the street is that it’s really bad (comparatively) at front-end design and 3D. But if there’s one thing our beefy boy improved in this release, it’s his stamina on long-running tasks.

Bottom line: The AI slowdown somehow resulted in a race to make existing models cheaper and more efficient, but sadly we’re stuck subscription stacking because you never know if your model of choice is going to have a bad cook.

---

Realtime verification inside your agent loop

The current generation of coding models relies on guess-and-check. They generate a bunch of code and you have to wait till it’s done in order to see if everything works.

Sonar Vortex collapses code verification and security checks directly into the agent’s loop. Here’s how it works:

• Inject context and constraints: Vortex adds the right project context and constraints.

• Generate code: The agent writes code with that guidance.

• Verify in real time: Each change is checked with algorithmic analysis.

Sonar Vortex connects to the AI tools your team already uses. Try it today and experience the difference.

---

1. Vadim Makeev wrote about the new CSS property: link-parameters that lets you pass CSS values into linked resources like SVGs. The only catch is that it’s “nowhere near production” 😭.

2. Speaking of CSS, Ahmad Shadeed shows how to detect if elements are overlapping using CSS. It’s getting weird out here in the land of the cascade.

3. Apify Store search now matches intent. You can filter by the data you need, the site you’re searching, or whether the tool connects to an MCP server. Use code BYTES to get $25 in credits. [sponsored]

4. The team at Meta wrote a very long article and made a nice video showing how they built a secure VM architecture for Muse. You can definitely trust them with your data.

5. Addy Osmani wrote about how to get the most out of Opus 5.5. It’s the lesson in etiquette you didn’t know you needed.

6. DHH broke Rails developers’ hearts at Rails World 2026 when he said he’s retired from professional programming and is having agents write Rust for HEY’s backend. Is this like if the Pope renounced Catholicism?

7. Convex’s CTO James Cowling gave an important talk about why abstractions still matter and how engineers are still responsible for what they ship even if they didn’t write the code. [sponsored]

8. Anant Goel wrote about the blind spots in harness evals.

9. Vue Router is working on experimental scroll restoration. Better late than never 🫡.

10. Hyperagent lets you build a team of specialist agents running their own browser, shell, and file system. They integrate with your tools and share context so they can hand off specific tasks without you babysitting them. Get free credits here. [sponsored]

11. Manu Martínez-Almeida wrote about how he built Figma-to-code for Builder.io before LLMs.

12. Google published a blog about how they automate “coherent” long-form video generation. Too bad they got everyone hooked on YT shorts.