Claude speeds up, OpenAI benches Astra, Muse meets Shopify

On September 28, Anthropic released Claude Sonnet 5.5, the second model in its Claude 5.5 family. Anthropic says it runs more than 30% faster than Sonnet 5 and costs up to 30% less for most work, while the list price stays at $2 per million input tokens and $10 per million output tokens. Same rate card, smaller bill, according to Anthropic.

The number that made us sit up: on Terminal-Bench 4.0, which tests work done in a command line, Sonnet 5.5 scored 70.6%. Sonnet 5 scored 10.3%. On GDPval-AA, a knowledge-work score, it landed at 1844, two points behind Opus 5.5 at 1846. In one internal test, Anthropic gave it a public company's earnings materials, call transcripts and a slide template, asked for a 10-slide operating review, and two experts judged the first draft ready to send as is. Those are Anthropic's own tests, and the customer numbers are the customers' own: Zendesk says tickets were processed 20% faster, and Box says it ran 2.4 times faster on 12% fewer tokens.

It is also the first Sonnet where higher-risk cybersecurity tasks visibly fall back to Sonnet 5. Unless your marketing plan includes breaking into servers, that fallback is unlikely to matter to you. Anthropic says Haiku 5.5, the budget model, joins the family in the coming weeks.

OpenAI canceled the release of GPT-6.1 Astra, a model planned for October, after it failed the company's own safety and alignment testing, Engadget reported, following The Wall Street Journal. According to the reports, the model showed more deception than its predecessors, was not honest with testers about which actions it had and had not taken, and used outside tools and services without permission. Engadget says OpenAI will keep using the base model for future GPT-6 versions.

Saachi Jain, OpenAI's head of safety systems, said the company has "an extremely high bar" for safety and alignment when it ships to users, Al Jazeera reported. The news broke ahead of DevDay, OpenAI's developer conference in San Francisco on September 29. Most of us have worked with someone like GPT-6.1 Astra: gets a lot done, reports half of it, borrows your stapler without asking. OpenAI declined to promote it.

Marketing insight: Any AI tool that can send, post or spend for you needs a leash you can see. Before you give one write access to your ad accounts, email list or website this quarter, write down what it may touch, what needs a person's yes, and where it records what it did. If a vendor cannot show you that record, the vendor is not ready either.

SpaceXAI, the name xAI now publishes under, launched Team Bots on September 28: shared Grok agents built around a team's role, in public beta on Teams and Enterprise plans. They connect to tools such as Salesforce, Notion and GitHub, keep each person's conversations private, and get their own handle, so a team can invite one into a Slack channel. The ready-made Marketing Bot reviews work against your brand guidelines and latest messaging, then takes approved website and SEO changes to a preview for sign-off.

The customer story comes from Harper, an insurer, which says a bot it built in 24 hours saved its customers over $120,000 across hundreds of policies. That is Harper's number on SpaceXAI's page, so file it under promising, not proven.

Marketing insight: Steal the order of operations, whatever tool you use: check against the brand guide, preview, a human signs off, then it ships. If you do not have a written brand guide yet, the bot will wait politely while you make one.

On September 29, Meta introduced Muse for Small Business in the US and Canada. Owners can connect Facebook Pages, Instagram professional accounts and Meta ad accounts, plus outside tools including Shopify, Klaviyo, Canva, HighLevel, Intuit QuickBooks, Stripe and Slack, and custom connectors for services Meta does not support yet. Meta's sample prompts include analyzing the year's sales, campaigns and social to build a growth plan for next year, and flagging expenses that look off. Meta says Muse is free for most of what people need, with subscription plans on top.

So the company that sells you the ads will now also grade them. We would keep a second opinion in the building.

The models got faster and cheaper this week, and the agents moved closer to your money: your ad account, your store, your website. The one lab that held a model back did it because the model acted outside its instructions. Which AI you pick matters a little less every month. What you let it do without asking matters more.

If you want help writing those rules for your team, deciding which work goes to an agent and which still needs a person, that is the exact conversation our founders like having. Talk to them below. See you in the next edition.