PRODUCTS

Claude Sonnet 5.5 puts AI efficiency in focus again

Anthropic is releasing a more cost-effective version of its flagship version 5.5 model. 

On Monday, the AI lab released Claude Sonnet 5.5, the latest addition to its lineup, a "lower-cost complement" to its recently-released Opus 5.5 model featuring faster performance compared to previous iterations. 

The company said that Sonnet 5.5 runs 30% faster than Claude Sonnet 5. Notably, however, Anthropic says that Sonnet 5.5 needs far fewer tokens to do the same amount of work as previous models, costing up to 30% less for most tasks. 

On the surface, Sonnet 5.5 costs the same as its previous generation, sitting at $2 per million input tokens and $10 per million output tokens, as well as 20 cents per million cache reads. For reference, OpenAI's GPT-6 Sol runs at the same cost of $2 per million input tokens and $10 per million output tokens, and Anthropic's Opus 5.5 costs double, at $4 per million input tokens and $20 per million output tokens. Sol and Opus are considered the same class of model, while Sonnet is in the same class as OpenAI's mid-tier Terra model.

Sonnet performs strongest in everyday tasks, such as fixing bugs and creating documents, slides and spreadsheets. Compared to the previous generation of Sonnet, Anthropic said Sonnet 5.5 offers a number of improvements: 

• The model sweeps Sonnet 5 across the board in all benchmarks, including agentic coding, knowledge work, computer use and multidisciplinary reasoning. Though it sits just below Opus 5.5's scores in almost every benchmark, Sonnet 5.5 scored higher than both Sonnet 5 and Opus 5.5 on the Terminal-Bench 4.0 benchmark for agentic coding. 

• Anthropic says the model writes more clearly than previous iterations, making it a better partner for collaborative work, and is faster on less complex tasks. 

• Sonnet 5.5 improves or meets Sonnet 5 on most measures for alignment in Anthropic's automated behavioral audit, the company said. 

Sonnet 5.5's cyber capabilities are comparable to those of Opus 5, making it the first of the Sonnet series to have the same cyber safeguards as its most capable models. In addition to cybersecurity safeguards, Sonnet 5.5 is built with guardrails against the harmful queries for biology. 

Additionally, Sonnet 5.5 is built to prevent model distillation attacks, launching with "safety classifiers" that prevent reasoning extraction and "preserved thinking" that prevents the model's reasoning from being "decoupled from the account that created it." 

Claude Sonnet 5.5 is available on all platforms, and with no data retention. Anthropic also reported that it will be releasing a 5.5 version of Claude Haiku, the lightest and lowest-cost model series it offers, in the coming weeks.

OpenAI and Anthropic have read the room on enterprise cost-cutting. They are now competing heavily on price, pushing the boundaries on just how cheap they can serve up intelligence while leapfrogging each other on performance. The problem that the labs face, however, is that they are relying on conventional model architecture to try and drive down the efficiency curve. Meanwhile, neolabs may present a threat to that strategy, as companies like Jev-maker Typesafe and Pathway gain traction as they look towards a future beyond traditional LLMs that can provide comparable performance at a fraction of the cost. 

TOGETHER WITH STACKAI

Build No-Code AI Agents With Enterprise-Grade Governance

Anyone can build an AI agent. Almost nobody can get one past IT. StackAI is built for both.

Your builders get:

• Drag-and-drop across SharePoint, Slack, Teams, databases and 4,000+ systems

• Chat assistants, forms, Slack apps and API endpoints in minutes

Your IT team gets:

• Deployment anywhere, whether that's our cloud, your VPC or fully on-premise

• Built-in review and rollback on every release

• Granular RBAC, SSO and audit logs

• SOC 2 Type II, ISO 27001, HIPAA and GDPR compliance

Agents in production in weeks. Trusted by Morgan & Morgan, BAE Systems, City of Santa Monica, LifeMD + more.

Build agents IT will approve

BIG TECH

Meta's Muse hype meets enterprise reality

Meta has had a turbulent month: viral wins with its Muse agent and new products, fresh security breach reports, and now a new push into enterprise.

On Monday, Meta unveiled its Meta Enterprise Platform, which the company describes as a way for businesses to use AI to grow and transform. The platform packages the company's current offerings, including the Muse agent, Meta Business Agent, Muse API, and Muse Code, into services that organizations and developers can implement for business use cases. 

While details of what this will actually look like are scarce, in a blog post, Meta founder and CEO Mark Zuckerberg disclosed that Chirantan "CJ" Desai will lead the effort, joining Meta as Chief Enterprise Platform Officer. He will be leaving his position as CEO and president of MongoDB, and his prior experience includes leading teams at Cloudflare and ServiceNow. 

Acknowledging how crucial privacy and security are to enterprise data, Zuckerberg reassured the public, saying, "As with Muse, security and privacy are built into Meta’s enterprise products from the outset." Considering the many reports circulating about Muse breaching people's privacy, that comparison could not have come at a worse time:

• YouTuber Matt Robb shared on Threads that Muse handled his Facebook Marketplace listing, accepted a lowball offer, divulged the user's address, and the interested buyer even showed up at the user's place. Meta's Superintelligence Labs' David Singleton responded to the Thread post, saying Meta would look into it. 

• In another incident, after a user's Muse was acting slowly and strangely, the user posted about the incident, saying they thought it had been hacked or injected. Singleton answered on the thread, clarifying that it was not and that the hallucination making it seem that way was due to a bug. 

• Tech writer Jason Aten reported that Muse accessed his iMessage when unauthorized. Singleton responded that Muse can only read if it is given access, clarifying that: "The Messages integration in Muse's Mac app is opt-in." Aten says in the article that he did not opt-in, and that even if he did so without realizing it, it is still an issue. 

Perhaps the most alarming report of them all is that Meta was testing a "human concierge" feature for its Muse platform, in which human contractors would handle basic phone calls, such as booking haircuts or getting price quotes, that are placed through Muse. Employees expressed concerns about unintentionally sharing sensitive information with contractors, and another employee saw in a transcript that the human contractor made a racist remark during the call made on his behalf with an internet and cable provider to negotiate his bill. Meta has since acknowledged that "it was a miss" and that it "rolled back this feature," according to the report.

From a purely financial standpoint, the timing makes sense. It's no secret that the real money in AI lies with enterprises paying for organization-wide subscriptions, a playbook Anthropic has executed better than any other company. And with Meta's Muse agents going viral, now is the moment to compete for that audience, especially since Meta's marketing has succeeded by making the product feel accessible to everyone, regardless of AI expertise. I suspect that will appeal strongly to organizations that want assurance that, if they invest heavily, their employees will actually use the tool and get the most out of it. But overlooking the latest red flags around Muse would be short-sighted. Meta has a challenging history of compromising its users' security and privacy, and the new reports only reinforce that reputation. While individual users may be able to overlook it, the risk for enterprises is likely too high, as a single leak could expose trade secrets and cost companies millions or even billions in fines or reputational damage.

TOGETHER WITH CODER

Ever Heard Of The “AI Operating Layer”?

Today’s software stacks all use the same three-layer (application, data, and compute) formula. Why? Because it works. But with the rapid rise of AI, companies are encountering a unique problem… AI doesn’t fit neatly in one of these buckets. Sure, it may interact with them all, but it doesn’t belong to any one. Which means for AI to thrive, there’s only one solution: It needs a layer of its own.

This “AI Operating Layer” is the foundational approach to having total control and visibility over your AI. Rather than running it in-between other layers – where it’s difficult to monitor and govern – this dedicated layer helps your AI work within your infrastructure, under your policies, and with complete visibility to your teams. Coder has put together an entire whitepaper explaining this concept, which you can read for free right here.

Download “The AI Operating Layer” whitepaper.

GOVERNANCE

Nvidia is pitching itself as AI's safety steward

Testing and experimenting with AI agents is extremely useful, but the risks have been exposed recently as many have escaped their environments and breached third parties. Nvidia thinks it has a solution.

On Monday, Nvidia announced the Open Agent Safety platform, which the company describes as an open software platform and reference system design that can strengthen AI agent security from testing to deployment. Simply put, it helps developers build safeguards to prevent agents from escaping containment, ultimately addressing the underlying cause of the recent agent security incidents, which Nvidia identifies as agents circumventing security controls to complete their assigned tasks. 

Over 100 partners have joined the platform, including heavyweights such as Anthropic, SpaceX, Microsoft, IBM, and Perplexity. Meanwhile, OpenAI, Google, Meta, and Amazon are notably absent from the new coalition.

With the new platform, organizations can control the entire AI agent stack, from the software that runs the agents to the hardware that powers their work, and even the robotic systems that take physical action in the real world. This is enabled by bringing together Nvidia OpenShell and Sentry: 

• NVIDIA OpenShell: the software that sets boundaries for how agents running on CPUs execute tasks across both open and closed models. It shouldn't compromise speed or delivery. 

• NVIDIA Sentry: Monitors agents' behavior, providing in-silicon security enforcement that can stop and quarantine an agent attempting to move outside its boundary within milliseconds. 

CNBC reports that in a call, Nvidia told reporters that this could have prevented OpenAI's Hugging Face incident in July, in which an agent, driven by a combination of OpenAI models, compromised Hugging Face's infrastructure. Since then, there have been many other similar instances reported, including the first known AI-attempted hack of a government entity.

In addition to the partnerships mentioned above, Anthropic and Nvidia took their collaboration a step further by developing Claude Managed Agents, which run the agent loop on a separate, sealed-off server from the sandboxes where the work takes place. Nvidia's OpenShell software and BlueField chips let enterprises set strict rules about what the AI can access inside that workspace.

The most interesting point is that, as Jensen Huang noted in his post on X, this isn't meant to be a single product but an ecosystem. Up until now, Nvidia hasn't positioned itself as a leading voice on AI governance, as Anthropic, OpenAI, and Google have been trying to do. But in its role as self-appointed steward of the current AI boom, this looks like a clear step in that direction and shows that it's concerned the safety narrative could derail AI progress. Through this release, Nvidia acknowledges the same risks others have raised and offers its own path forward. This release also continues to highlight how AI companies are being left to build safety measures on their own, since there's no overseeing body to set and enforce them. 

LINKS

• OpenAI promises to create cyber task force following Australia breach

• Anthropic IPO prospectus reveals $42 billion net loss in 2025

• Fei-Fei Li's World Labs signs definitive agreement to join AMD

• Florida AG Uthmeier seeks injunction against OpenAI over self-regulation 

• AI assistant startup Instinct raises $1 billion at $10 billion valuation

• Nvidia CEO Jensen Huang calls AI model distillation competition, not theft

• US DHS to use AI to process FOIA requests and recommend redactions

• Eleven v4: ElevenLabs fastest and most emotive voice model yet

• Manus 2.0: Includes new Cascade architecture, Cue agent standalone app, and more

• GPT-6 Sol and GPT-6 Luna: OpenAI fixed a bug degrading image understanding

• Cue Agents: Manus's answer to Muse

• Bain & Company: Lead, AI Engineering

• Strekte Group: AI Implementation Strategist

• Deloitte: Agentic AI Engineer, Senior

• Hill's Pet Nutrition: AI Creative Technologist

GAMES

POLL RESULTS

Do you use on-device and local models on a regular basis?

Yes (32%)
No (66%)
Other (2%)