The Brief #26.34 · Five new models, GitHub's long Monday, and Firefox 154

Five models entered the weekly table at once, GitHub lost most of a Monday, and Firefox 154 finished two CSS features the other engines already had.

Five new rows in the model table

Artificial Analysis scored five new models this week, and every model already on the board kept last week’s score. 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, so this week’s scores compare cleanly with last week’s, though the cost column did move in three places. Gemini 3.7 Flash, Grok 4.6, GLM-5.3, and the new Qwen weights each get their own story below.

The fifth new model is DeepSeek V4-Pro, released to general availability on August 13. It enters the table at 53 on the Index for $0.25 a task. The company also delivered the API price increase that the last issue flagged, effective August 16.

Its rates now split into peak and off-peak hours, and off-peak costs half as much. Peak covers 01:00 to 04:00 and 06:00 to 10:00 UTC, and every other hour is off-peak. The leaderboard quotes the peak rate, which is why V4-Flash climbed from $0.03 to $0.11 a task.

Two other returning costs moved, GLM-5.2 up from $0.31 to $0.44 with no announcement I could find, and Gemini 3.6 Flash down from $0.56 to $0.34. Those three all sit outside the ten shown below. Qwen3.8-27B enters without a cost because Artificial Analysis hasn’t priced the row yet.

Table The Intelligence Index top ten, plus the lower-ranked models this issue covers.

#	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	$1.23
4	Grok 4.6 (high)New	61	$0.84
5	Kimi K3 (max)	60	$0.84
6	GLM-5.3 (max)New	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)New	56	$0.40
13	DeepSeek V4-Pro (max)New	53	$0.25
14	GPT-5.6 Luna (max)	52	$0.05
15	Qwen3.8-27BNew	52	not yet priced

Gemini 3.7 Flash arrives three weeks after 3.6

Google released Gemini 3.7 Flash on August 13, three weeks after 3.6 Flash, and calls it its most intelligent “workhorse” model for coding and agents. A workhorse AI model is the mid-priced one you run all day, as opposed to the flagship you save for hard problems. GPT-5.6 Terra and Claude Sonnet 5 fill that same role in their own families. All day means jobs like these:

• Reviewing every pull request as it lands, one after another.
• Turning a folder of page designs into first-draft components.
• Drafting a reply to each incoming support ticket for a human to approve.

For web developers, Google claims Gemini 3.7 Flash builds better web apps than 3.6 Flash did, producing working layouts and more complete features in fewer prompts. The evidence it offers is Arena.ai’s WebDev Arena, where 3.7 Flash scores 1588 against its predecessor’s 1538. On the arena, a visitor writes a prompt and two randomly paired models each build an app from it. The visitor compares the two side by side and votes for the better one, and the model names appear only after the vote, so nobody votes for a brand. Ratings build from votes across many visitors.

Independently, Artificial Analysis scores it 56 at high effort for $0.40 a task, four points above 3.6 Flash, and measures it as the fastest model in the table at 322 output tokens per second. The introductory price is $0.75 input / $3.75 output per million tokens through December 31, 2026, and the post’s own footnote already schedules the double, $1.50 / $7.50 from January 1, 2027. The model is available in the Gemini API through AI Studio, in Antigravity, and in Spark for Google AI Pro and Ultra subscribers.

Grok 4.6 ties GPT-5.6 Sol

SpaceXAI released Grok 4.6 on August 12, aimed at long-running agents and interactive visual work. Artificial Analysis scores it 61 at high effort, tied with GPT-5.6 Sol and behind only the two Claude rows. The posted price is $2 input / $6 output per million tokens, the same rate as Grok 4.5, yet a task costs $0.84 against its predecessor’s $0.36, more than double.

The new model emits far more billed tokens per finished task, and that volume is what the cost-per-task column prices. There’s a fast variant at twice the price, and the model is live in Cursor, Grok Build, the API, and gateways like OpenRouter.

GLM-5.3 reaches 60 before its weights do

Z.ai released GLM-5.3 on August 14. It’s built on the same base model as GLM-5.2, and all of the gain comes from post-training, the teaching phase that follows the big training run. Artificial Analysis scores it 60 at max effort for $0.68 a task, matching Kimi K3 at a lower cost.

The weights, the downloadable model files, aren’t out. The company promises them about two weeks after launch, so until they land Kimi K3 remains the strongest model you can actually download.

One breaking change hits anyone already on the API: GLM-5.3 supports three thinking effort levels and removes the option to switch thinking off, so a request that sends thinking.type: "disabled" fails on the new model id. The fix is to send enabled and set reasoning_effort to low. Pricing holds at $1.40 input / $4.40 output per million tokens, the same rate listed for GLM-5.2.

The Qwen weights landed, on two licenses

Alibaba’s Qwen team put the Qwen3.8-Max weights up on Hugging Face on August 12, when the model card and license were finished, and published a new, much smaller Qwen3.8-27B on August 14. The Max is the frontier-sized one, 2.4 trillion parameters with about 95 billion doing the work on any single request. It ships under its own Qwen3.8-Max License, which allows commercial use but requires a separate agreement with Qwen once a company running a model-serving or AI-assistant business passes 50 million dollars of revenue in a year.

The 27B is the one sized for hardware you might own. It reads images and video as well as text, and it ships under Apache 2.0, the permissive license that allows commercial use without asking anyone. It scores 52 on the Index, level with GPT-5.6 Luna.

Weights like these are what you load into Ollama or LM Studio, the free apps that download a model and run it on your own machine. Matching that 52 score means running the full-size download, roughly 56 GB, on a machine with about 64 GB of RAM. Both apps offer compressed versions instead, near 20 GB for a 24 GB graphics card or a 32 GB laptop, and those give up some of the quality the score measured.

GitHub lost most of a Monday

GitHub.com served elevated errors and latency for 7 hours and 47 minutes on Monday, August 17, across Issues, pull requests, the API, Actions, and Copilot. At the worst, roughly one web or API request in five failed, and about half of archive downloads and raw file fetches did. Sign-in through SAML and OIDC, git operations, and Pages joined the incident as the day went on, so stuck merges, failed pushes, and Copilot timeouts all had the same cause.

GitHub’s report traces the outage to its Central US data centers, where a new traffic high left the machines that spread requests across servers short of network capacity. Every service there runs behind a small helper program that carries its network traffic, and that helper handles only so many requests at a time. It hit that limit, and the system that adds more copies under load never added any, because it was watching how busy the service was instead of how busy the helper was.

Hacker News, the developer forum where stories rise on reader upvotes, carried four threads about the outage in one day, and they collected about 2,200 upvotes between them. One of the four was readers asking for GitHub alternatives. A 20 percent error rate isn’t a blackout, but it’s plenty to stall merges, deploys, and sign-ins for most of a working day.

VS Code’s built-in browser refreshes itself

Microsoft released VS Code 1.133 on August 12. The integrated browser, the browser tab that lives inside the editor, now reloads a local HTML file automatically when the file changes on disk. The release notes frame the feature as watching agent edits land immediately, and it covers your own saves just the same. The wording covers a local file open in that tab, so a project served by a local development server keeps the reload setup it already has.

The model picker now separates models billed through your own API key from the ones in your subscription, and you can switch providers between turns. With a Claude key already configured, an experimental setting opens the Agents window without a GitHub sign-in. And in long chats, the prompt you scrolled past now stays pinned to the top.

Chrome patches 15 more security bugs, two of them Critical

Google shipped a stable Chrome update on August 18 carrying 15 security fixes, two rated Critical, both buffer overflows in graphics code, one in WebGL and one in Dawn, the layer underneath WebGPU. A week earlier a smaller update fixed five more, and Chrome 152 reached its early slice of users on August 12, with full rollout scheduled for August 25 and no feature notes published yet.

One of the High fixes, a use-after-free in WebGL, is credited to OpenAI Codex Security alongside a human handle, and the same team was credited in the August 6 update too. Neither the August 18 nor the August 11 post reports in-the-wild exploitation. Update Chrome today anyway, the patch only protects you once the browser restarts.

Cursor sold to SpaceX, then became a code host

Cursor said on August 14 that its acquisition by SpaceX had closed. TechCrunch puts the deal at $60 billion in stock and says SpaceX exercised an option it has held since April.

Cursor says it joined SpaceX for the computing power. The post promises the largest fleet of GPUs in the world, the chips that train and run AI models. It says nothing about the models from other AI companies that the editor offers today, or whether they stay.

Three days later, on the Monday of the GitHub outage, Cursor launched Origin, its own git hosting and code review, in early beta on every paid plan, with an opt-out for enterprise administrators. Origin is a forge, the whole product around a repository, the web view, pull requests, review, and permissions. That’s the job GitHub, GitLab, and Bitbucket hold now. GitHub repositories can sync alongside the Cursor-hosted ones, and a comment on a synced pull request posts to both sides.

Cursor also shipped Builds, a ready-made copy of your project that its cloud agents start from. Without it, asking a cloud agent for something meant waiting minutes while it booted a machine, downloaded the code, and installed everything the project needs, before it read a single line. Cursor refreshes that copy every hour and reports its agents starting to answer three times sooner. It bought Firetiger too, whose agents watch software while real people are using it and pass the problems they find back to the coding agent.

Figma wraps text the way CSS does

Figma shipped Balance and Pretty text wrapping on August 14 across Design, Slides, Buzz, and Sites, settable per text layer, text style, or paragraph. Balance evens a heading’s lines out instead of leaving one long line over a short one, and Pretty keeps a single word from stranding alone on the last line. The names come straight from text-wrap: balance and text-wrap: pretty, so a mockup can finally show the wrapping the CSS will produce, though the release notes don’t promise the line breaking matches any browser’s exactly. The day before, the company opened an agent skills library on the Community, more than 50 skills for design workflows, and the Figma agent can now write a new skill from your file’s context and publish it for anyone to remix.

Firefox 154 finishes text-box-trim and sibling-index()

Mozilla shipped Firefox 154 on August 18, desktop and Android, and with it the last missing engine support for the text-box properties and the sibling counting functions. text-box-trim and text-box-edge cut away the built-in empty space a font carries above its capitals and below its baseline. That invisible padding is why a heading never sits exactly where your spacing math says it should, and trimming it makes vertical rhythm come out as written, no compensating negative margins required.

sibling-index() and sibling-count() give an element its position among its siblings, and the total count, as numbers that work inside calc(). A staggered animation delay no longer needs a Sass loop or an --i custom property set from JavaScript, and the index starts at 1, not 0.

The other engines shipped both earlier, the text-box properties in Chrome 133 and Safari 18.2, the sibling functions in Chrome 138 and Safari 26.2. As of this release, each feature works in the current version of every major browser. The same update adds a native picker UI for <input type="time"> and support for the No-Vary-Search response header, where a server declares which query parameters don’t change the response so the cache can reuse one stored copy across them.