The Brief #26.35 · A cheaper Sol, Stripe to buy OpenRouter, and JPEG XL
Want more of these in your Google results?
OpenAI made its best model cheaper, Stripe is buying OpenRouter, an Amazon for AI models, and Mozilla cleared the way for JPEG XL.
Sol’s price cut reaches the model table
OpenAI cut GPT-5.6 Sol’s API price on August 21, 20% off input and 33% off output, to $4 input / $20 output per million tokens. Input covers what you send the model, and output covers what it writes back.
The announcement calls it promotional pricing, good at least through November 21, 2026. The same cut reaches every other way of buying the model, from Fast mode and long-context requests to Batch and Flex processing, credits for the Codex coding agent, and ChatGPT Work. July’s cuts covered Terra and Luna and skipped Sol, so the best model in the family was the last one to get cheaper.
On the weekly table, the cut shows up in the cost column, with Sol down from $1.23 to $0.96 per task. The Intelligence Index runs nine hard tests and reports them as one number, higher is smarter, and the cost column is the average price of one of those tasks.
The Index is still on version 4.1.1 and every returning row kept last week’s score, so this was a quiet week everywhere but that one column. The only other move is a price for Qwen3.8-27B, the row that entered the table last week unpriced, and it now runs $0.25 a task at xhigh effort, the setting that gives the model the most thinking time.
| # | Model | Intelligence Index | Cost per Index task |
|---|---|---|---|
| 1 | Claude Opus 5 (max) | 63 | $2.34 |
| 2 | Claude Fable 5 (max) | 62 | $3.14 |
| 3 | GPT-5.6 Sol (max) | 61 | $0.96 |
| 4 | Grok 4.6 (high) | 61 | $0.84 |
| 5 | Kimi K3 (max) | 60 | $0.84 |
| 6 | GLM-5.3 (max) | 60 | $0.68 |
| 7 | Qwen3.8-Max | 58 | $1.13 |
| 8 | Muse Spark 1.2 (xhigh) | 57 | $0.40 |
| 9 | GPT-5.6 Terra (max) | 57 | $0.51 |
| 10 | Gemini 3.7 Flash (high) | 56 | $0.40 |
| 16 | Qwen3.8-27B (xhigh) | 52 | $0.25 |
Your laptop is only about five months behind the cloud
The best laptop-class open model now scores 52 on the Intelligence Index, the number the table above shows in its last row. GPT-5.6 Luna scores that same 52. That score arrived on August 14 with the Qwen3.8-27B weights, the downloadable model files last week’s Brief covered.
The chart below tracks the best of those laptop models against the best hosted model, like Gemini, Claude, or ChatGPT, on the same index over time. Laptop-class means open (a model anyone can download and run on their own machine) and small enough for a strong laptop. The size limit is about 35 billion parameters, the internal numbers a model is made of.
Every score belongs to the full-quality download, which for Qwen3.8-27B is a roughly 56 GB file that needs about 64 GB of RAM, so that 52 assumes a seriously equipped laptop. The 16 to 21 GB compressed versions that fit a laptop with 32 GB of RAM give up some of the measured score.
The cloud line ends at 63 with Claude Opus 5, and the laptop line at 52. The cloud first passed 52 on March 5 with GPT-5.4, so the newest laptop step showed up only about five months behind.
The lag here is how long ago the cloud first passed the level the laptop just reached, and a year ago the same measurement came out at roughly eleven months. Epoch AI runs the same comparison for open models of any size on its own capability index and measures an average lag of four months since January 2026, so with no size limit the gap comes out even smaller.
Downloadable also doesn’t mean laptop-class, since Kimi K3’s open weights score 60 but need a supercomputer (at least 1.5 TB of RAM) to run.
Mozilla will ship JPEG XL
Mozilla posted its intent to ship JPEG XL on August 24, the public notice a browser team files when a feature is headed for release. Firefox Nightly, the browser’s testing build, already displays the format. JPEG XL is a newer image format designed to replace the JPEG most photos on the web still use.
Support for it had sat behind a flag in Firefox since 2021, a setting you have to turn on yourself, blocked by a security objection. The worry was the decoder, the program that turns an image file back into pixels on your screen. A browser runs one on every image every page shows, including images from strangers, and a specially crafted image can use a decoder bug to run an attacker’s code on your machine.
The objection was about what one bug could do rather than any hole someone had already found. The standard JPEG XL decoder is about 100,000 lines of C++, in that language one memory mistake is enough, and Mozilla didn’t want that much new risk in the browser.
So Mozilla told the JPEG XL team at Google Research it would ship the format if they built a safe, fast decoder in Rust, a language designed so that kind of memory mistake can’t happen. The result is jxl-rs, and it’s the decoder shipping now.
Chrome filed its own intent to ship the same day, also on jxl-rs. Safari has shipped the format since 2023, minus the progressive loading part, and Mozilla expects support across browsers before the end of the year.
Browser data updated September 24, 2026
Mozilla’s post compares it with AVIF, another modern image format browsers already support, and puts real numbers on the comparison. At web quality AVIF tends to come out smaller. At lossless, compression that keeps every pixel exactly, JPEG XL wins, 1.45 MB against 1.76 MB on a test photo and 92 kB against 164 kB on a screenshot.
JPEG XL can also recompress the JPEGs a site already serves into smaller files with no quality loss. And it renders progressively, meaning the image draws while it downloads, so a reader on a slow connection can tell what a photo shows after just a few kilobytes.
Nothing needs changing this week.
New formats arrive behind <picture>, the HTML tag that offers the same image in several formats and lets each browser take the first one it understands.
So a JPEG XL version becomes worth adding the moment stable support lands.
You can follow that on Chrome’s status entry, and I’ll let you know in a future Brief too.
Claude’s browser tool reads your HTML
Anthropic launched a browser use tool on August 19 and took several features out of beta the same day. The tool is part of Claude’s API, so this is for developers building AI products rather than something you turn on in the Claude app. A product built on it runs a normal browser on its own computers, and Claude tells that browser what to open, click, and type, with Anthropic hosting nothing.
Teams already glued this together with outside tools that let a model steer a browser. Now that ability is built into the API itself, and it stays inside one browser window rather than reaching the whole desktop.
The older computer use tool only sees pixels, screenshots in and coordinates out. The browser tool also reads the page’s structure. It pulls the accessibility tree (the element-by-element outline a screen reader gets) and tags every entry with a reference.
From there the browser tool acts on those references, sets form values directly, manages tabs, and reports downloads, with screenshot-and-click still available underneath. File upload exists too, off by default.
The same day’s release notes move three more developer features out of beta, the computer use tool, the Files API, and Agent Skills. The browser and computer use toolsets run on the Claude 5 models and Opus 4.8.
The accessibility tree is built from your HTML, so an agent that targets elements by reference is navigating by the names, roles, and structure you wrote.
A real <button>, a labeled input, and a proper form give an agent something to drive reliably, and a pile of styled divs, generic boxes with no stated role, forces it back to screenshots and guessed coordinates.
The HTML work that serves screen-reader users is the same work that now makes a page automatable.
ChatGPT can read your Apple Messages
OpenAI added an Apple Messages plugin to the ChatGPT desktop app for macOS in the week of August 17, on all plans. It finds chats, summarizes conversations, drafts replies, and sends them through Messages, covering iMessage, SMS, and RCS.
The plugin’s documentation says it runs on your Mac and reads your Messages history there once you grant the macOS permissions. It works inside Codex and ChatGPT Work sessions rather than regular chats, and this release ships only in the Apple silicon build.
By default a message goes out only after you approve the text and the recipients. A per-chat setting can make that approval permanent, and OpenAI’s own docs warn against using it where a chat might contain untrusted instructions, since nobody reviews a message before it sends under your name.
End-to-end encryption protects a message while it travels between devices, and it does nothing for the readable history sitting on your Mac. The plugin asks for permission to read that stored history, the same permission any tool of this kind has to ask for.
Before you turn a tool like this on, check who reads that history, where the processing happens, and what leaves your Mac.
CSS gets a class prefix selector, on paper
The CSS Working Group, the body that writes the standard for the web’s styling language, resolved on August 6 to add a class prefix selector, and the spec text landed in the Selectors Level 5 draft on August 18.
The new syntax is .btn-*, and it matches every element carrying a class that starts with btn-, so .btn-primary and .btn-danger both match while a bare .btn doesn’t.
The prefix must end at a hyphen, and there must be at least one character after it that isn’t another hyphen, so double-hyphen modifier classes like .btn--modifier stay out too.
Utility CSS and icon fonts, styling systems built on small prefixed labels, define thousands of these classes. Reaching them today takes attribute selectors, blunter patterns that search the raw class text and are easy to get subtly wrong.
The new selector reaches those classes directly. It’s a resolution and a draft, no browser implements it, and there’s nothing to try yet, but it says clearly where selectors are headed.
Drive Antigravity from any browser
Google shipped Remote Control for Antigravity, its code editor built around AI coding agents, on August 21. From a web browser, you connect to Antigravity sessions running on your other machines (a work desktop, a laptop, a server) and drive them remotely. A push notification arrives when an agent has finished its turn and needs input.
The sessions keep their local context, the files, workspaces, credentials, and settings already on each machine, so an agent carries on where it was instead of starting from scratch somewhere else. In practice that means starting a long agent run before you leave one machine, then answering its questions from wherever you are.
Codex passes 20 million users
Thibault Sottiaux, who runs Codex at OpenAI, said on August 21 that the coding agent passed 20 million active users some time that week. The number is his own, with no outside audit behind it.
That same weekend OpenAI also reset usage limits twice. The first reset, for all paid ChatGPT Work and Codex users, landed August 22.
Limits kept draining too fast after that, Sottiaux tied part of it to images in long sessions being handled inefficiently, and a second reset landed August 24 alongside fixes. If your limits looked refreshed twice in one weekend, that’s what happened.
Stripe is buying OpenRouter
Stripe, the company that moves money for a huge slice of the internet, agreed on August 19 to acquire OpenRouter, with no price disclosed. OpenRouter works like an Amazon for AI models. Instead of a separate signup with every model company, you make one account there and pay one bill for access to more than 400 models from over 80 of them.
Plenty of frontend developers pay for their models exactly that way, and its customers include NVIDIA, Zoom, and Lovable. Stripe already handles the money coming into a company, what its customers pay. OpenRouter sits on the money going out, what the company spends on AI models, and Stripe says it wants to run both sides.
“Tokens are the central currency for companies building with AI,” Stripe’s CEO Patrick Collison said. The pitch to an AI company is that OpenRouter can send each request to whichever model does the job for less, which cuts that outgoing bill.
The deal is an agreement rather than a closed sale, so nothing changes for customers today. Still, a payments company is buying the layer that sees which model every request went to and what it cost, and ledgers like that are its line of work.
A perfect ARC-AGI-3 score, without touching the model
NVIDIA reported on August 21 that a research system called Avo scored a perfect 100.00 on ARC-AGI-3, a public benchmark for AI agents, clearing all 183 levels across its 25 games. ARC-AGI-3 drops an agent into small games with no instructions, no stated rules, and no stated goal, so it has to work out how each game behaves, and what winning even means, by playing.
Avo, for Agentic Variation Operators, is a system built around an existing model rather than a new one. The model inside is Claude Opus 5, and ARC Prize, the group that runs the benchmark, puts the bare model at about 30%.
The rest of the score, by NVIDIA’s account, comes from the system built around the model. That system keeps a persistent memory of earlier attempts and their results, runs a supervision layer that watches for the agent going in circles, and loops through implementing a change, evaluating it, and going again.
There’s a paper and nothing to download or run. The result shows how much of an agent’s score lives outside the model, since memory, supervision, and an evaluate-and-retry loop moved the number from about 30 to 100 with the model unchanged.
The framing cuts both ways, since a perfect score shows the system handles long unfamiliar games while the post stops short of promising the same lift on your own work.
Want more of these in your Google results?