The Brief #26.39 · Claude Opus 5.5, Jev, and Chrome 154

We had a very busy week. Claude Opus 5.5 led Artificial Analysis’ tests. Everyone started talking about Jev. OpenAI released GPT-6 Sol and Luna, SpaceXAI introduced Grok 4.7, and Xiaomi’s MiMo-V2.6-Pro is now free. Chrome 154 allows embedded frames to resize to fit the page, Lighthouse 13.5 checks support for AI assistants, and Cloudflare added previews for each branch. Google, Anthropic, and OpenAI updated their programming tools, and nine mathematicians will help OpenAI with sharing math research.

Claude speeds up the AI race

This week was packed with Artificial Analysis rating six new models that came out between September 18 and September 22. Anthropic’s Claude Opus 5.5 scored 58, topping the leaderboard and beating their previous best model, Fable 5.1. This score comes from Artificial Analysis’ Intelligence Index, which averages results from ten tough tests for each model. A higher score means a smarter model.

At high effort, Opus 5.5 scored 53.58 for $1.82 per task. For comparison, Claude’s Fable 5.1 at max effort scored 53.35 for $7.63 per task. This means Opus 5.5 matches Fable 5.1’s top score at about a quarter of the price.

Anthropic says it gave Opus 5.5 the job of cutting load times on every page of a web app, and it succeeded 39 out of 40 times. In the same test, Opus 5 made smaller improvements that also changed how the app behaved.

OpenAI released two new models: GPT-6 Sol, which scored 48 for $1.06 per task, and GPT-6 Luna, which scored 37 for $0.07 per task. Luna is especially good for inexpensive, non-critical tasks.

SpaceXAI’s Grok 4.7 and Xiaomi’s MiMo-V2.6-Pro (from the phone maker) both scored 46. However, MiMo-V2.6-Pro costs just $0.13 per task, compared to $3.74 for Grok 4.7. The next most affordable model in the top ten, GPT-6 Sol, costs about eight times more than MiMo-V2.6-Pro.

StepFun, a Chinese AI startup, launched Step 5 Preview as the sixth new model. It scored 44 for $0.72 per task.

Table Intelligence Index

#	Model	Intelligence Index v4.3.2	Cost per Index task
1	Claude Opus 5.5 (max)New	58	$5.98
2	Claude Opus 5.5 (high)New	54	$1.82
3	Claude Fable 5.1 (max)	53	$7.63
4	GPT-6 Astra (max)	53	$3.26
5	Claude Fable 5.1 (high)	51	$3.91
6	Muse Spark 1.3 (max)	48	$1.60
7	GPT-6 Sol (max)New	48	$1.06
8	Grok 4.7 (xhigh)New	46	$3.74
9	MiMo-V2.6-ProNew	46	$0.13
10	Qwen3.8 Max (0902)	45	$5.41
11	GLM-5.3 (max)	45	$2.01
12	Grok 4.6 (high)	44	$1.86
13	Step 5 PreviewNew	44	$0.72
14	Kimi K3 (max)	44	$2.00
15	GPT-5.6 Terra (max)	42	$1.40
16	GLM-5.3-Flash	42	$0.25
17	Gemini 3.8 Flash (high)	41	$1.24
18	GPT-6 Luna (max)New	37	$0.07

What’s Jev?

Jev has been getting a lot of attention lately, so I wanted to explain what it is and when you might choose it instead of a regular AI model.

TypeSafe AI, a US AI lab founded in 2024, released Jev on September 15, and within three days it was also available through Vercel’s AI Gateway, Netlify’s AI Gateway, and OpenRouter. With these gateways, you can access models from different companies using a single account and get one bill.

Vercel says nearly 13% of paid teams on its gateway started using Jev within the first 24 hours. They also said that no other model on their gateway has been adopted as quickly.

Jev is an AI model that doesn’t generate text, so it cannot draft replies or write code for you. Instead, you send it some text (which TypeSafe calls the state) along with one or more questions. For each question, you provide the possible answers, similar to a multiple choice question. For each question, Jev returns one of your provided answers in a format your code can use, along with a probability.

TypeSafe offers three kinds of questions.

• A Choice question lets you list up to 255 options. Jev returns one of them, plus a probability for each option.
• With a Score question, Jev places the text on a scale you define, such as “low”, “medium”, or “high” urgency.
• A Noul question, which Vercel calls Boolean, is a simple “yes” or “no” question. Jev answers with the probability of “yes” from 0 to 1.

All the questions in a single request are processed at the same time, so adding more questions doesn’t significantly increase the wait time. You specify the options for each request, rather than training Jev with your own data. TypeSafe describes Jev as a “System One model”, inspired by the fast, instinctive thinking Daniel Kahneman called System 1 in his book Thinking, Fast and Slow.

Jev only accepts text as input, so images or voice messages need to be converted to text before you use them. TypeSafe also lists math and comparing dates among Jev’s weak spots.

TypeSafe charges $0.042 input / $0 output per million tokens for Jev. Yes, free output. You only pay for the text you send. The state and your longest question together can be up to about 32,000 tokens.

Take a look at the examples below. Choose one, press “Send the request”, and you’ll see answers from Jev 1.13 and Claude Opus 5 side by side, each showing the typical response time from OpenRouter’s tests.

OpenRouter tested bank customer messages and found Opus 5 was more accurate, getting 84.4% right compared to Jev’s 81.0%. However, Jev answered much faster, taking 175 ms per reply, while Opus 5 took 2,266 ms, which is about 13 times longer. Jev was also way cheaper at $0.11 per 1,000 messages, while Opus 5 cost $2.42.

I picked five uses from TypeSafe’s list to help you get ideas for what you can build with Jev.

• Send each contact form or support message to the right team using a Choice question, and include a yes-or-no question to check if it’s urgent in the same request.
• Before sending each chat message to your main AI model, you can use a yes-or-no check to see if the message tries to make the model ignore your instructions, which is called a prompt injection.
• You can filter out spam and abusive comments or reviews so a person can review them.
• Only send the difficult requests to your expensive model. You can find these by asking a yes-or-no question about whether each request needs that model.
• By asking one Choice question per record, Jev can sort all your product descriptions or years of old support tickets into categories you choose.

TypeSafe paused new sign-ups for Jev on September 22, saying it had seen “such an immense swell of demand” that it had to stop for now. The company says current accounts will keep working while it works to give everyone access as soon as possible. OpenRouter’s Jev guide says you don’t need to join a waitlist or have a TypeSafe account to use Jev there. You only need an OpenRouter API key.

Chrome 154 lets an iframe grow to fit the page inside it

Google moved Chrome 154 to the stable channel, the version regular users get, on September 22. This update also includes 108 security fixes.

In Chrome 154, an iframe can now expand to fit the entire embedded page, so you see the whole page without its own scrollbar.

This sizing feature only works if it’s enabled on both the parent and embedded pages. Usually, for privacy and security, the outer page can’t see how big the content inside an iframe is. To use this, the embedded page needs to have <meta name="responsive-embedded-sizing"> in its <head>, and you need to set frame-sizing: content-height on the iframe in your CSS.

Keep in mind that if the embedded content changes size after loading, like when an accordion opens, the iframe will not resize on its own. You will have to adjust it manually.

Chrome154

Edge154

FirefoxNo

SafariNo

Safari iOSNo

Chrome Android154

Browser data updated September 24, 2026

Here are two more things you should know about Chrome 154:

• The new text-decoration-inset feature lets you control where an underline begins and ends, either inside the text edges or extending beyond them. Google suggests using this for underline reveal effects, which used to need a background gradient or an extra element (Chrome 154 beta post).
• Now, Chrome only closes a popover or dialog when a full click event happens outside it. This means it will stay open if you scroll on a touchscreen or right-click outside the dialog.

OpenAI released GPT-6 Sol and GPT-6 Luna, with changes to prompt caching

OpenAI released GPT-6 Sol and GPT-6 Luna on September 22. These models run faster and cost less than earlier versions in the GPT-6 family, which started with GPT-6 Astra on September 3. OpenAI says Astra is still its best-performing model overall.

For prompts up to 272,000 tokens, GPT-6 Sol costs $2 input / $10 output per million tokens, and GPT-6 Luna costs $0.10 input / $0.50 output.

If the start of your request repeats earlier instructions or tool definitions, OpenAI reuses its previous work for that part. This is called prompt caching. Repeated tokens are charged at just 10% of the usual input price.

In a September 22 post, OpenAI explained how prompt caching works on GPT-6. You get the discount if you send the same beginning of a prompt again within 30 minutes of its previous use. On both models, the first time you write data into the cache, it costs 1.25 times the normal input price, making the first request a bit more expensive. After that, requests for that data are billed at just 10% of the regular price.

Grok 4.7 from SpaceXAI and MiMo-V2.6-Pro from Xiaomi

SpaceXAI released Grok 4.7 on September 21 at $2 input / $6 output per million tokens for prompts under 200,000 tokens, the same price as Grok 4.6. The company says Grok 4.7 uses a new, larger model than Grok 4.6 and was trained longer on harder tasks. Grok 4.7 uses more tokens for each task. This means that, even though the token price is the same, it ends up costing more per task than Grok 4.6.

Xiaomi put MiMo-V2.6-Pro on Hugging Face, a marketplace for AI models, on September 21. The model is under the MIT license, which lets anyone download its files (the weights) and use them commercially.

However, MiMo-V2.6-Pro has 1.02 trillion parameters, which means you can’t run it on a regular computer unless you have access to a supercomputer. The model is split into specialist sections, so it uses only 42 billion parameters per token.

There is an interesting detail about the cost of training AIs. According to Xiaomi’s technical report, the reinforcement learning stage alone (the practice-and-feedback training that comes after a model’s main training) cost $2.6 million.

What changed in AI coding tools this week

Antigravity 2.0, Google’s desktop app for running and supervising AI coding agents, got three updates between September 18 and 22, versions 2.15.0, 2.15.1, and 2.16.0. On Windows, you can now turn on Enable Sandbox Mode to run agent commands in a sandbox. This limits which files and network connections the agent can access. Antigravity now supports connecting to a Linux installation running inside Windows.

Subagents now show up as live cards, so you can monitor or stop them as needed. You can also drag and drop Word, Excel, and PowerPoint files into a prompt.

Anthropic’s Claude Code 2.1.277, out since September 18, loads AGENTS.md in a project that has no CLAUDE.md. Both files are Markdown documents at the root of your repository. Use them to write your project’s rules and commands for coding agents. If you already use an AGENTS.md file for other agents, you can remove your CLAUDE.md file.

OpenAI’s Codex now works in VS Code if you have either a GitHub Copilot or a ChatGPT subscription. You can also switch a Codex session between the ChatGPT app and VS Code without having to start over.

Lighthouse 13.5 adds checks for sites that work with AI agents

Google released Lighthouse 13.5 on September 18 with a new audit for a catalog file you can use to list your site’s AI tools. This catalog uses a format called Agentic Resource Discovery (ARD). Lighthouse is Google’s tool for checking websites. You can run it from the Lighthouse panel in Chrome’s developer tools, or on Google’s servers through PageSpeed Insights. Google expects version 13.5 to be available in PageSpeed Insights by October 2, and in the panel with Chrome 156, which comes out on October 20.

An ARD catalog is a JSON file that lists the agents and tools available on your site. It makes it easier for AI agents and directories to discover and verify them. The format is still a proposal, written by people at Google and Hugging Face together with R.V. Guha. If your site doesn’t have a catalog and your server answers the catalog’s address with an error like a 404, you will see a “not applicable” result.

Lighthouse 13.5 looks for the catalog at /.well-known/ai-catalog.json. However, the latest draft of the format, dated August 26, uses /.well-known/ard.json instead. Lighthouse also needs a "specVersion": "1.0" field in the catalog, but the latest draft doesn’t mention this requirement.

If you choose one of the three options below and click “Request the catalog,” you will see the response from /.well-known/ai-catalog.json and the report from Lighthouse 13.5.

In Lighthouse 13.5, the new audit appears alongside the older llms.txt check, both found under Agent Discoverability in the Agentic Browsing section of the report. llms.txt is a Markdown file that gives a short summary of your site and links to its key pages. It’s made for AI agents.

WebMCP is a proposed standard that allows you to give AI agents in the browser a set of tools on your page, such as “search events” or “add to order.” This way, the agents can use these tools by name instead of having to click through the page themselves. Lighthouse now shows a warning if a page has more than 40 WebMCP tools.

Cloudflare Worker Previews

Cloudflare launched Worker Previews on September 22, letting you run each app branch on Workers as a separate copy from production. Workers is Cloudflare’s platform for running sites and server code. Each preview comes with its own code, settings, URL, and logs. You get 100 previews per Worker on the free plan and 500 on paid plans.

To create a preview, use Wrangler, Cloudflare’s command-line tool. Run npx wrangler preview with version 4.135.0 or later. You can also link your repository to Workers Builds. It will create the preview for you and post the URL on your pull request.

Each preview has a single URL that updates every time you push changes. Every deployment also gets its own fixed URL, so you can return to any exact version.

Preview URLs are public by default, but you can make them require sign-in with Cloudflare Access.

A group of mathematicians will advise OpenAI on releasing its math results

OpenAI announced on September 21 that it’s working with the Advisory Group on Mathematics and Artificial Intelligence. This group includes nine mathematicians who will help the company check and publish the math results produced by its models. In the same post, OpenAI said that the internal model it began training on August 28, which produced its Navier-Stokes proof, has now solved over 100 long-standing open problems. The company hasn’t published those proofs or revealed which problems they solved.

On September 8, OpenAI published its Navier-Stokes proof. That same day, two mathematicians released their own proofs of a similar result for three related fluid equations, as The Brief covered. Three days later, winners of the Fields Medal, which is the best-known prize in mathematics, published a declaration opposing the use of famous problems as benchmarks by AI companies. A benchmark is a public test that shows how capable a model is. They wrote that announcing results too quickly doesn’t leave enough time for a proper write-up or to credit earlier work.

By September 22, almost 7,900 people had signed the declaration. This is up from about 7,000 when it was mentioned in last week’s Brief. OpenAI mentions the declaration in its post and introduces the group as its response to the criticism.

According to the group’s site, OpenAI first asked some of the members to join an advisory board, and with OpenAI’s agreement they formed an independent group instead. The group will publish its recommendations on its site.

Martin Hairer, a Fields medalist who signed the declaration and later joined the group, wrote on September 22 that only the nine members are involved in their discussions. He said that he and the other signers had asked AI companies to address the issues mentioned in the declaration. Because of this, he felt it would be hypocritical to refuse the first real chance to speak with one of them.

OpenAI has said that the internal model behind its Navier-Stokes proof is much more capable than GPT-6 Astra. However, it hasn’t named the model or announced when it will be available to people outside the company. The company says it is still figuring out how to make its math abilities more widely available.