BIG TECH

Why big customers are pulling back on Claude

Three of Anthropic's biggest revenue sources are slowing or stopping their spend on Claude. 

On Monday, reports emerged that the Pentagon has officially stopped using Anthropic's Claude following the agency blacklisting the company by designating it a supply chain risk earlier this year. In a statement to the BBC, the Pentagon said it has "ceased the use of Anthropic products," after initially announcing in February that it would stop using Claude by August. 

However, while Anthropic losing the business of the largest department of the US government has been foretold for months, the AI lab has also taken a few more hits. According to The Information, Meta and Microsoft are both actively reducing their internal reliance on Anthropic's tech as they seek to bolster the use of their own internal AI tools. 

• The report found that Microsoft expected to spend at least $1 billion this year on Anthropic's tools internally, but that estimate has been cut by more than a third.

• Meta, meanwhile, reportedly dropped its Claude Code seats from 60,000 to roughly 30,000, with the reduction attributed to two factors: its spring workforce reduction of roughly 8,000 employees and its push to use internal tools instead. 

Anthropic did not respond to The Deep View's request for comment in time for the publication of this story.

While the Pentagon and the tech giants have turned away from Anthropic for starkly different reasons, neither are particularly auspicious signs ahead of Anthropic's blockbuster IPO. According to Bloomberg, the AI lab is targeting a public market debut before the end of the year, and could go public as soon as the week of Nov. 9. 

Anthropic is expected to debut at an eye-popping valuation of $2 trillion. Though the company lost $42 billion in 2025, according to its IPO prospectus reported by Reuters, The Information reported in August that its second quarter revenue this year was 14 times that of the same quarter the previous year, making it profitable from an adjusted operating income basis. However, keeping up with demand is going to be a massive expense, as it intends to spend $518 billion on cloud and computing infrastructure in the coming years, Reuters reported.

These reports come at a time when enterprises are grappling with rising AI spend and challenging conversations around ROI. Largely, the tokenmaxxing trend that overtook the industry earlier this year has run out of steam in most places. It's a trend that both Anthropic and OpenAI have clearly noticed as they attempt to undercut one another on prices for their frontier models.

The major AI labs' real rivalry isn't with one another, despite how they make it seem in the media. Rather, the real competition is with the increasingly attractive and affordable offerings coming from open models and as well as the innovations coming out of neoclouds that are innovating with more efficiency-focused architectures. And because switching between model providers is often expensive and cumbersome for big enterprises, once these companies wean off Anthropic or OpenAI, it's unlikely that they'll be easily persuaded to come back. Additionally, another factor potentially driving the shift away from major model providers is controllability. Major enterprises like Meta and Microsoft are very intent on deeply integrating AI throughout their processes, and trusting that data and context to an outside source–especially a competitor–is unlikely to be a long-term solution. Plus, along with the controllability benefits, these companies stand to save money by using their own internal tools for a large majority of tasks, reserving frontier intelligence for larger jobs. The question, however, is what this means for the two leading frontier labs. As the push towards efficiency and privacy has been shaking out for months while the AI leaders prepare for their record-breaking public market debuts, will they be to convince investors that they have a solid long-term growth story? 

TOGETHER WITH WIZ

A Practical Framework for AI Threat Readiness

As attackers use AI to identify vulnerabilities and accelerate exploitation, security teams need operational strategies that keep pace.

Wiz’s AI Threat Readiness Playbook provides a four-pillar framework security leaders can use to strengthen programs and prepare for AI-enabled threats.

Learn how to:

• Reduce critical attack surface exposure

• Improve zero-day response and remediation

• Strengthen application security with AI-assisted analysis

• Modernize detection and response using AI-driven workflows

Get the Playbook

PRODUCTS

Why Mistral's 1T model is a hedge against lock-in

Most flagship models dominating the market are from frontier labs and are closed models. French AI lab Mistral seeks to change that with its latest model.

On Tuesday, Mistral launched a public preview of Mistral Large 4 (ML4), le Chonk, its new flagship, 1-trillion-parameter, open-weight model meant for general agentic capabilities. Mistral claims its competitive advantage lies in the fact that it is comparable to the best closed models while staying competitive with open-weight models three times its size. Meanwhile, it was trained on 4,000 NVIDIA Grace Blackwell GPUs at what it describes as a fraction of its competitors' cost.

With this launch, Mistral is also telling a cybersecurity story, not only claiming it is strong in both offensive and defensive categories, but also capitalizing on the fact that it is an open-weight model. Mistral's co-founder and chief scientist Guillaume Lample told The Deep View that the open-weight nature of the model allows companies to safely build and deploy cybersecurity defenses without fearing that the model is suddenly deprecated or sundowned.

"Basically, new [cybersecurity] workflows require a lot of tokens, a lot of inference, so you cannot afford to be vulnerable to the fact that the model you are using to protect yourself might disappear one day, or might be too limited," said Lample.

Beyond cyber, the new model is proficient in other realms including coding, domain-specific knowledge work, and multi-modality. Some other features, according to the blog post, include:

• Based on its grounding capabilities, "it outperforms all existing models, including the closed ones," Mistral claims.

• Deployment on Mistral's own data centers in Europe and served in preview on that infrastructure.

• Trained using reinforcement learning, and the team continues to see rapid progress. This is notable as many open models are mostly distilled, and the team is highlighting its own research, training infrastructure, and expertise.

The open-weight model is aimed primarily at enterprises deploying it on-premise or in a private cloud, but it will also be available through Mistral’s API. At one trillion parameters, it’s too large to run locally on a desktop or laptop and may be impractical for some universities to run themselves. The weights will be released on October 27.

Open-weight models play a key role in the AI industry because they let companies and users retain control over which model they use and how they use it. Because the weights are publicly available, users can inspect, test, and run a model themselves to judge whether it meets their needs. Once a model is deployed, they also hold a complete copy of it, so they don't depend on a provider's decision to continue to maintain it or keep it available. This matters because every model version behaves slightly differently. Even a forced migration to a newer, more capable model can set off a chain reaction that breaks systems built around the old one. That makes ongoing efforts to build open-weight models that rival closed models from leading labs especially significant. Much of the open-weight ecosystem is driven by the Chinese frontier labs, so the industry tends to welcome alternatives from providers like Mistral.

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

OPEN SOURCE

How US open models are closing the gap with China

Though much of the open-source ecosystem is concentrated in China, US open models are starting to pick up steam. 

On Tuesday, Google unveiled EmbeddingGemma 2, an open-source model that Google calls its "most capable" yet for on-device work, sitting at a lightweight 740 million parameters and requiring only 270 million parameters for text-only workloads. However, the model outruns its predecessor by going beyond text and can handle code, images, video and audio. 

Google noted that the model offers powerful on-device performance for semantic search, routing and retrieval. For instance, when put to use, the model is capable of finding specific clips from voice memos and searching through hours of audio recordings from a text request. It also sports strong multilingual performance and is best-in-class for embedded models under 1 billion parameters. 

However, Google isn't the only US company making moves in open models this week. On Monday, New York-based AI startup Reflection unveiled Beam, an open-source model built for coding, reasoning and agentic workloads. 

Though far larger than Google's new model at more than 500 billion parameters, Reflection characterized its new model as "highly efficient," offering frontier reasoning at up to four times cheaper when compared to larger open-weight models like GLM-5.2. Additionally, the company said its models are approaching that of larger frontier open models such as Qwen 3.8-Max on coding and agentic tasks. Reflection noted that Beam's advantage compared to frontier open models is "efficiency at inference time." 

"These results translate into more intelligence per token, delivering strong model capabilities at lower cost, making Beam a powerful workhorse model for enterprise coding and agentic workloads," the company said in its announcement.

Though Google and Reflection's new offerings target far different workloads, these companies couldn't have picked a better time to push forward into the open model ecosystem. Frontier labs building proprietary models are trying to undercut each other on price while leapfrogging in capability, while AI costs and data sovereignty get more critical for enterprises. And while open models are praised for their cost efficiency, given that so much of open-source innovation, research and development is centralized in China, the argument against them is largely one of security. In feeding the broader US open source ecosystem, the latest models from Google and Reflection offer enterprises more choice in optimizing for cost, efficiency and safety. 

LINKS

• DeepSeek nears 80 billion yuan raise ahead of 2027 IPO

• South Korea investigates AI use in bank customer data attacks

• OpenAI executive Jason Kwon unable to explain Australian distrust of AI

• Anthropic expands cyber program after it found 100,000+ vulnerabilities

• Chip design firm Vinci raises $250 million at $1.5 billion valuation

• Meta, Walmart, Stripe publish agent standard for businesses, AI bots

• Eleven v4 Turbo: Ranks #1 on the Artificial Analysis Provider Voice TTS Arena 

• Manus 2.0: Features a video editor that can edit, make shots, VO and more

• Together Link: Run frontier open models in your favorite coding harness

• Claude Code 2.1.290: 190 CLI changes; including Deny button in sign-in approval pages

• NanoBanana 2.1: Google's latest image generation model, outperforming predecessors

• Azazie, Inc: AI Solutions Specialist

• Proofpoint: AI Automation Engineer

• Jacobs: AI Developer

• Kaizen Equity Partners: AI Operations Associate

GAMES

POLL RESULTS

Do you think AI regulation should be lenient or strict?

Strict (51%)
Lenient (7%)
Somewhere in between (36%)
Other (6%)