Enterprise AI is getting more affordable

Welcome to Runtime! Today: Anthropic's latest Sonnet model could make getting capable AI agents up and running much cheaper, Meta makes a big bet on Nvidia's (other) chips, and the latest funding rounds in enterprise tech.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime each week, and if you value independent enterprise tech journalism, click the button below and become a Runtime supporter today.

Like a summer's day

Late last year, after rolling their eyes during months of hype that enterprise AI agents were going to change everything, software developers started to realize something really had changed. The November release of Anthropic's Opus 4.5 model turned Claude Code into an incredibly powerful tool for creating software and managing agents, setting off a chain of events that convinced Wall Street over the last several weeks that the entire enterprise software industry was doomed.

Anthropic announced Tuesday that most of the head-turning capabilities introduced in Opus 4.5 are now available in a new release of its Sonnet model lineup, which is significantly cheaper to run than Opus. "Performance that would have previously required reaching for an Opus-class model — including on real-world, economically valuable office tasks — is now available with Sonnet 4.6," Anthropic said in a blog post.

• Sonnet 4.6 is better than its older sibling (which was released last September) at tasks such as coding, computer operation, and reasoning, and the company said some early testers liked it better than Opus 4.5.
• Anthropic thinks enterprises will want to employ Sonnet 4.6's computer-use skills to introduce agents to legacy but still-critical software applications that would have been otherwise hard to modernize.
• When it comes to coding, Anthropic said Sonnet 4.6 "more effectively read the context before modifying code and consolidated shared logic rather than duplicating it," which makes it less frustrating to work with.
• And it will become the new default model in Claude Cowork, which really set stock traders' hair on fire when it was released last month.

But Sonnet 4.6 is just the latest example of how powerful AI capabilities are becoming much more affordable for companies that don't need state-of-the-art performance. Opus 4.5 was itself a much cheaper version of the original Opus 4 series models, and now those capabilities cost even less.

• Sonnet 4.6 costs $3 for every million input tokens and $15 per every million output tokens, which compares to $5/$25 for Opus 4.5 and $15/$75 for Opus 4.1, which was released last August.
• If agents are the future of enterprise software, they are going to be running more or less constantly and in parallel as they code or analyze projects, and that is going to get real expensive real fast.
• But over the last year we've seen a steady decline in what Anthropic's almost-state-of-the-art AI model performance costs, which could allow a much wider variety of companies to jump on the agentic bandwagon.

Anthropic has been a little more circumspect than OpenAI when it comes to making big splashy AI infrastructure announcements, but it too needs to generate a ton of revenue over the next several years to offset the cost of training and serving these models. And while Anthropic enjoys a stronger presence inside the enterprise compared to its rival, OpenAI has made it very clear it intends to focus on business users during 2026.

• As businesses finally start to wrap their heads around AI-assisted coding and business-process agents, pricing is going to be a much more important part of the buying decision than capabilities alone.
• OpenAI and Google's models are generally cheaper than Anthropic's at the moment, but so far enterprises have been willing to pay for Claude's performance, especially when it comes to coding.
• That might not always remain the case, but if Anthropic can keep lowering the price of accessing that performance it will continue to have a strong argument inside the enterprise.

CPU to-do

Nvidia's GPUs need no introduction at this point in the AI boom, but its CPUs are far less common inside enterprise data centers and cloud providers, which is one reason why Nvidia and Intel cut a deal last year to add NVLink to Intel's server processors. Meta gave its CPU business a big vote of confidence Tuesday with a multiyear deal to add those CPUs to its chip arsenal.

Meta will build new data centers around the Grace CPU, making it "the first large data center operator to use Nvidia’s CPUs in standalone servers," Bloomberg reported. Grace is based on Arm's chip designs, and Meta also plans to use the next-generation Vera CPU once that comes out later this year.

CPUs remain indispensable to enterprise computing at scale, even after Nvidia CEO Jensen Huang tried to convince everyone in the early days of the boom that GPUs are all you need. Training and serving AI models creates a ton of other workloads that need processing, and CPUs are a far more cost-effective way to get those jobs done.

Enterprise funding

Temporal raised $300 million in Series D funding, valuing the development-platform company at $5 billion.

Render scored $100 million in an extension to its last Series C funding round as it builds out an AI development platform.

Braintrust landed $80 million in Series B funding for its observability software, which tracks AI workloads.

Mesh Optical Technologies launched with $50 million in new funding for its data-center networking chips.

GitGuardian raised $50 million in Series C funding to build out identity security technology designed to work with agents.

OPAQUE scored $24 million in Series B funding for its AI compliance software, which helps companies make sure they have the right guardrails in place while running AI agents.

The Runtime roundup

Palo Alto Networks beat expectations for revenue and earnings during the last quarter but disappointed investors with a lighter-than-expected profit projection, showing that even security companies aren't immune to SaaSpocalypse 2026.

ServiceNow CEO Bill McDermott will purchase $3 million of the company's stock and other executives will cancel pre-arranged trading plans in hopes of shoring up its stock price, which closed down on the news.

Thanks for reading — see you Thursday!