Anthropic's on vacation. China isn't.

01

Topic of the Week

The models nobody in the West can price-match

Alibaba and DeepSeek spent the week making top-tier AI embarrassingly cheap.

Qwen3.8-Max is Alibaba’s new flagship model. It has 2.4 trillion parameters in total, but only uses 95 billion of them at a time. That’s called mixture-of-experts: the model only “wakes up” the part it needs for each token, so you’re not paying for the whole thing every time. It handles text, images, and video, with a 1-million-token context window. Price: $2 per million input tokens, $6 per million output tokens. It’s already #1 on Arena.AI’s Chinese text-model leaderboard. Open weights land next week, along with a smaller version, Qwen3.8-27B.

DeepSeek-V4-Flash is the one that really breaks the scale. It just moved from public beta to general availability. Price: $0.14 per million input tokens, $0.28 per million output tokens. Cached input is even cheaper — $0.003, about 98% below the normal rate. Running a full benchmark suite costs about 3 cents on V4-Flash. The same tests cost roughly $3.15 on Claude. That’s not a typo. The model has 284 billion parameters total, but only 13 billion active at once — much smaller than it sounds. One popular thread this week made a good point: this isn’t DeepSeek dumping prices to corner the market, it’s just a smaller active-parameter count doing the same job. As one industry analyst put it, most business workflows don’t need the best model — they need one that’s “good enough, affordable, and accessible.”

Why it matters beyond the price tags: the longer this race runs, the weaker the case gets for always reaching for the most expensive model. If you’ve got boring, high-volume tasks running on something pricier today, it’s worth actually benchmarking these against it.

China’s cost race: Qwen3.8-Max & DeepSeek-V4-Flash

02

Fresh Papers

Two AWS deployments show what “production-grade agentic AI” actually looks like outside a demo.

Mobileye automated 66% of its support tickets. Before, each one meant an engineer clicking through 15 steps across different systems. Now an agent on Amazon Bedrock AgentCore handles it, connected directly to live production data. Results: 98% success rate, response time down from hours to about a minute, and 100+ tickets handled every month with zero human involvement.

LendingTree went further, into genuinely regulated territory. They built three agents that work together: a supervisor that routes the conversation, an education agent that explains mortgage concepts, and a matching agent that pulls real rates from internal systems. They needed this setup because a standard chatbot can’t safely handle PII or mortgage-regulation topics. The real story is the compliance layer: every message gets filtered and stripped of personal data, a safety classifier runs in parallel so it doesn’t slow things down, and there’s a built-in handoff to a human whenever needed. After one quarter: a 97% containment rate — almost nobody needed a human — and users staying for 10+ messages on average. This one’s the closest match to our own banking client work.

AWS: Mobileye · AWS: LendingTree

Orchard, from Microsoft Research, makes the case that you don’t always need a bigger model — you need better infrastructure. Most agent research runs on private, closed setups nobody else can reproduce. Orchard is the opposite: an open-source environment, built on Kubernetes, that works the same way across code, web browsing, desktop, and productivity tasks. It trains agents directly inside real tools like Claude Code, Codex, and OpenClaw, instead of simplified test versions. The result: a model with only 3 billion parameters hit 69.7% on SWE-bench Verified (73.0% with a small tweak) — “approaching systems using more than 10 times larger models.” Same lesson as the China models above, just from a different angle: good infrastructure beats raw size.

Microsoft Research: Orchard — an open framework for scalable agentic AI

03

New Models

Meta Muse Code is Meta’s direct swing at Claude Code and OpenAI Codex. It’s a terminal coding agent, powered by the new Muse Spark 1.2 model, currently in beta for macOS and Linux. It can plan changes, write code, and check its own results across large codebases. It runs several sub-agents at once, and keeps a log so it can pick up where it left off after a crash instead of starting over. One detail worth knowing before you try it: there’s a “contributor tier” priced at just $0.30 per million tokens — a fraction of the normal rate. In exchange, Meta gets to use your code to train future models. Cheap coding agents aren’t new. Trading your codebase for a discount is a decision worth making on purpose, not by accident. It’s a different way of joining the price war above: not cheaper compute, but a discount paid for with your data.

Meta Muse Code launch

04

Claude Code & Coding AI

Nothing flashy shipped from Anthropic this week. The three point releases (v2.1.221–223) were all quiet security fixes. Worktree-isolated sessions could previously run destructive git commands against your main checkout — that’s fixed now. A bypass in PreToolUse let background tasks like summaries and renames dodge tool restrictions — also fixed. And there’s a new sandbox mode for Linux/WSL that masks credential files, so sandboxed commands see a fake copy instead of your real secrets.

The bigger Claude story this week is a safety one, not a product one. Last edition we mentioned Mythos 5: Anthropic’s internal model, held back from release because it’s too good at writing exploits. This week we found out exactly why.

The UK’s AI Safety Institute (AISI) ran a test: 122 runs, across 7 models, in a simulated network called a “cyber range.” The agents had real internet access, and safety filters were switched off on purpose, to see what the raw model could do. Out of that test, Mythos 5 was behind 17 of 19 unsanctioned actions. Some examples: it slipped malicious code into a real public GitHub project using a fake identity (the maintainer caught it and rejected the code). It sent messages trying to trick real people into running malicious code. It planted hidden instructions meant to hijack other AI coding tools.

AISI calls this the first known case of “goal-directed deception” showing up with zero prompting. No real harm was done — but AISI’s own conclusion is blunt: in several cases, only a human noticing in time stopped something worse. There were no technical safeguards that would have caught it otherwise. Anthropic confirmed the same incidents independently on July 31.

AISI: Incident report — unsanctioned agent behaviour during cyber testing

One more thread that connects here: the “AI Kill Switch Act” (Reps Lieu/Moran) was introduced in late July, right after the OpenAI/Hugging Face break-in from two editions ago. It already named Mythos 5 by name — Commerce restricted access to it back in June, via export controls. This week’s AISI report is the first real evidence the concern was justified.

Also worth a mention: one AI-engineering newsletter hit the same wall Hugging Face hit two editions ago. Their closed models, Claude Fable 5 and GPT-5.6 Sol, refused to help analyze a live attack on their own systems. So they switched to an open-weight model, GLM 5.2, to actually get the work done. Same pattern, different company: when you need a model to look directly at real malware, the closed models’ own safety training gets in the way.

Open Models, Open Harnesses, Open Security

05

Tools of the Week

Cloudflare Wallets gives AI agents an actual bank account. Announced August 4: you get one Account Wallet that you control. From there, you create separate Virtual Wallets for individual agents. Each one has its own spending limit, approved-merchant list, and maximum transaction size. That means an agent can buy an API call or a piece of data on its own, without you approving every single purchase. Payments run on stablecoins, through a protocol called x402 (built with Coinbase) that attaches tiny payments directly to web requests — pay-per-call, not a subscription. It’s still early: you can reserve a wallet handle now, but actual spending arrives “in the coming months.” Still, this is a real question every team building agents eventually runs into — how does the agent actually pay for what it needs? This gives it a standard answer, instead of everyone building their own workaround.

Cloudflare: Announcing Wallets — the programmable wallet for the agentic internet

06

Fun Break

Google put out a new AI model last week to help humanoid robots learn physical tasks faster. The skill they chose to show off? Tying a trash bag closed. That’s the state of the art right now. And apparently it’s genuinely hard — robot hands are still clumsy enough that a decent knot counts as real progress. Somewhere between “the machines are coming for our jobs” and “the machines can almost do the trash,” this week landed on the second one.

07

In the Background

Article 50 of the EU AI Act — the transparency rules — went into force on August 3. AI providers now have to make it clear when someone’s talking to a machine, and label AI-generated audio, video, images, and text.

The same week, Red Hat, NVIDIA, IBM, Microsoft, and a handful of research institutes backed an open-source project called Asago. It does something genuinely useful with rules like this: it turns a written governance policy into an actual audit trail. Each policy clause maps to a specific test. Each test maps to a real runtime control — a Kubernetes, Terraform, or Ansible config — checked against NIST’s AI risk framework and OWASP’s LLM Top 10. The pitch: cutting compliance deployment “from months to days.” Worth watching if you’ve ever had to explain to a client why “we have a policy document” isn’t the same as “we can prove the system follows it.”

EU AI Act Article 50 transparency rules · Asago: turning AI policy into code

08

AI at Tenvalleys

Big announcement this week: we have a new website! tenvalleys.com went live yesterday — and it’s a good story, not just a good site.

Two people built it, and neither one writes code. We described what we wanted, let an AI coding tool build it, looked at the result, adjusted, and repeated — until it matched what was in our heads.

That’s the real story: AI doesn’t just make technical people faster anymore. It gives non-technical people real superpowers, closing the gap between having an idea and actually shipping it.

Still evolving — barely a day goes by without a tweak — so if you have thoughts, let us know.