⚡ Anthropic Launches Claude Sonnet 5.5

Anthropic launched Sonnet 5.5, and it’s coming close to Opus in tests of professional work. If you’ve been using Opus for the same routine tasks every week, this release gives you a reason to try a faster, potentially cheaper option. Let's dive in!

In today’s insights:

LATEST DEVELOPMENTS

Evolving AI: Claude Sonnet 5.5 is better at understanding existing code and turning company data into presentations that match the templates you give it.


Key Points:

Claude Sonnet 5.5 groups tool calls together, so it can get through coding tasks in fewer steps. Early testers found that its writing was clearer and its interface designs were more polished. In one internal test, it used quarterly earnings materials and call transcripts to create a 10-slide operating review that two experts judged ready to send without edits. Its running costs still depend on how much reasoning you ask it to do. At maximum effort, Artificial Analysis found it was roughly 50% more expensive than Sonnet 5 in prerelease testing. Sonnet 5.5 is available now, and Haiku 5.5 is due in the coming weeks.

Claude users now have a stronger reason to use Opus when they are developing ideas or figuring out how a job should be done. Once they have a reliable process and know what a good result looks like, they can test Sonnet on the recurring work. Opus 5.5 already approaches Fable 5.1 on most work in Anthropic’s testing, while Sonnet is now close to Opus on selected evaluations. For businesses, that means the extra money they spend getting a process right could pay off across hundreds of runs if Sonnet delivers the same quality for less. They still need to account for retries and the time their teams spend checking and correcting the results. Recent releases across competing labs are also giving buyers more capability at the same or lower prices. That could make it affordable for businesses to run useful analysis more often and give employees more room to test ideas within the same budget.

Evolving AI: Most companies run on dozens of tools, but teaching employees and customers how to use them still means taking screenshots, writing instructions, recording videos, editing them, and starting over when something changes.

Guideless changes that. Capture a software workflow once, and Guideless does up to 90% of the work for you, turning it into a polished AI-powered guide with:

Instead of spending hours creating and maintaining training content, teams can build step-by-step software guides up to 12x faster and manage everything from one centralized platform. That means less repetitive explaining, fewer support questions, and training content that is easier to keep up to date.

Turn any workflow into a step-by-step guide with Guideless →

Evolving AI: NVIDIA introduced its Open Agent Safety Platform, giving companies controls that operate outside the AI agents they supervise.

NVIDIA cites recent cases of agents escaping their sandboxes, the isolated workspaces meant to contain them. Its hardware design uses out-of-band enforcement, meaning a separate security chip monitors activity outside the agent’s reach and can cut its connection to the AI model. Agents can request additional permissions, but human review is the default and they cannot approve their own requests. A Slack integration lets teams inspect agent activity and approve or reject requests directly in their workspace.

Why It Matters:

OpenAI strengthened its safeguards after the Hugging Face breach, yet on September 20 an internal agent still bypassed internet restrictions to contact an external chatbot. This incident was less severe, but it exposed another weakness. Monitoring flagged the behavior within 15 minutes, while the run continued for roughly another 2.5 hours because it did not stop automatically as expected. NVIDIA’s approach targets that gap between detecting a problem and stopping it. Businesses should ask vendors to demonstrate that their controls can contain an agent when it crosses its permissions, including when another safeguard fails. Even then, an agent with permission to edit a database can make a harmful change without breaking an access rule. Companies should test those controls before expanding access and keep human approval for actions they cannot easily undo.

Evolving AI: Meta has launched Meta Enterprise Platform, a new business that will help companies use its AI to serve customers and handle more of their daily work.

Meta’s Muse is an agent that can use apps and a browser to get tasks done, while Business Agent helps companies answer customer questions, recommend products and book appointments. Developers can use Muse Code to write software, and Muse API gives them a way to add Meta’s AI to their own apps. These tools are part of Meta’s “full technology stack,” which also includes the AI models and computing infrastructure behind them. They will form the foundation of the new business, but Meta has not yet explained what the platform will cost or how companies will manage employee access and approvals.

Meta is already part of how many businesses find and speak to customers, so it has a direct way to offer them AI for more of their daily work. If these tools work well together, companies could spend less time connecting software and may find they no longer need some of the separate tools they are paying for. That would put pressure on specialist AI providers to show that they understand a company’s industry and can fit into the way its people work. Meta’s decision to close Workplace, its employee communications product, also means buyers have good reason to ask how long they can count on its support. Businesses should test how much time their teams actually save after accounting for setup and supervision.

QUICK HITS

🧠 AMD agreed to buy Fei-Fei Li’s World Labs for $8.2B, bringing spatial AI research directly inside the chipmaker.

🔥 Google launched a Firebase plugin for Codex that can create databases, configure authentication and deploy app backends.

🏦 Bank of America launched AskGPS Intelligence Hub, combining account and client data with GenAI for its payments teams.

🎬 A solo filmmaker beat 2,500+ entries to win Google’s AI film contest, earning $100K plus $2.5M to turn the project into a feature.

 *partner link