A Chinese Model Just Beat Claude at Coding. For a Tenth of the Price.

Morning, folks!. A packed one, a Chinese model just beat Claude at coding for a tenth of the price, a CEO profited off his own company's acquisition, an Amsterdam startup doubled its valuation in six months, Claude's about to help you shop, and Meta's new model finally learned to ask before it breaks something. Let's get into it.

Qwen3.8-Max-0902 just took #1 on Code Arena's WebDev leaderboard, edging past Claude Opus 5 by 3 points. No new model version either, just a targeted update focused on coding and long-running agent work. A month ago it was sitting in 4th place, 37 points behind, and now it's beating the model it was chasing. Same pricing the whole time too, $2 input and $6 output per million tokens, versus roughly $20 for Opus 5 to score marginally lower.

Here's the catch, though, and it's a real one, not a footnote. Every API call still runs through Chinese legal jurisdiction, benchmark win or not. Being cheaper and technically ahead on a leaderboard doesn't make that consideration disappear, and it's exactly the kind of thing that gets buried under a flashy ranking headline.

Palo Alto Networks paid $500 million in cash and stock for Console, an AI agent startup automating IT help-desk work, according to sources, more than 3x its $157 million valuation before the deal. The console is barely two years old, having raised just $29M total. The detail that stands out, CEO Nikesh Arora was personally an angel investor in Console before his own company acquired it.

This is Palo Alto's fifth security acquisition this year, on top of a $25B CyberArk deal and $3.35B for Chronosphere, all pointed at one number they keep repeating to Wall Street, $20 billion in next-gen security revenue by 2030.

Wonderful just raised $550 million at a $5 billion valuation, more than doubling its value from $2 billion in March. The company, founded in early 2025, has gone from roughly $1 million to $70 million in annualized revenue in about a year and now has 650 employees across more than 35 markets.

And it's not slowing down. Wonderful plans to reach more than 1,000 employees by the end of the year, with much of that growth going toward engineers working directly with customers to deploy its AI systems. It has now raised more than $800 million in total. That's a pretty wild amount of capital for a company that's barely two years old.

Anthropic just launched new commerce agent blueprints for Claude, aimed at retailers, travel companies, marketplaces, telecoms, and ticketing platforms. Instead of building an AI shopping system from scratch, companies can use Anthropic's templates and plug them into their own products. There are two versions: a shopping agent for customers and a merchant agent for businesses.

Anthropic says companies already using Claude-powered shopping agents have seen carts grow by up to 35% and customers become 60% more likely to complete a purchase. The blueprints cover things like product discovery, recommendations, and purchase decisions, while businesses still control how the agents behave and what they can access. Claude is moving from answering shopping questions to actually helping close the sale.

Meta released Muse Spark 1.3, and the real upgrade isn't the benchmark scores, it's judgment. It now asks clarifying questions when a prompt's unclear, checks in before doing anything it considers irreversible, and can juggle a few tasks in one messy thread without losing the plot. On coding work, Meta's own engineers measured about 20% fewer tool calls and 25% fewer tokens than the last version, actually more efficient, not just marginally better.

There's real safety work in here too, stronger against prompt injection, better at telling what's actually risky versus routine. And Muse Spark 1.2 is getting a full open-weight release soon, once Meta wraps up final safety testing, same open push that started with Muse Glimmer.