PRODUCTS

Claude Opus 5 joins race for AI token savings

When it comes to its latest AI models, Anthropic is scaling confusion.  

On Friday, Claude officially got its fifth-generation Opus model. And while the best news is that it's half the cost of Fable, the bad news is that users are likely to be even more confused about which Claude models they should use. 

Anthropic says that Opus 5 is a "proactive" model that comes close to Fable 5's frontier intelligence.

• The model hit state-of-the-art performance on coding and knowledge work evaluations, but still lags behind Mythos 5 on cybersecurity (but then again, what model doesn't). 

• The model provides a significant jump in performance from its predecessor, Opus 4.8, which it released in late May. Opus 5 is also much stronger at verifying its work and iterating until success. 

• Opus 5, however, has the same price tag as its previous generation: $5 per million input tokens and $25 per million output tokens. It's still on the expensive side, since it costs virtually the same as OpenAI's top-tier model GPT-5.6 ($5/$30 per million) and over twice as expensive as similar models from Google Gemini and xAI's Grok. 

Early testing by companies like Cognition, Cursor, Lovable, Zapier, Box and more found that the model outperformed competitors on things like analytical work, agentic coding tasks and debugging. Sualeh Asif, co-founder of Cursor, said in a statement that Opus 5 delivered "near Fable 5 intelligence at Opus speed and cost," offering many of the same behaviors. 

Additionally, Anthropic said Fable 5 is its most aligned model to date, adhering to Claude's constitution better than Opus 4.8, Sonnet 5, or Fable 5, with the lowest rates of deception and the least susceptibility to being fooled into misuse. In other words, Anthropic is making the case for Opus 5 as its safest model.

It also makes sense that Anthropic is targeting cost and efficiency with Opus 5. AI costs have become a major 2026 pain point for enterprises. Offering ways to trim token budgets has come sharply into focus for tech giants like Microsoft, Uber, Nebius and Databricks as the tokenmaxxing fad has sputtered out in favor of efficiency.

However, a cheaper and more efficient model may be in Anthropic's favor, too, as the company has struggled to find the compute needed to run its ultrapowerful Fable 5, tapering access by only making it available to Max and Team Premium users, and for only 50% of standard weekly usage limits. 

This adds to the already confusing question of which Claude model to use. Opus used to be Anthropic's flagship tier for Claude models, representing the most powerful and most expensive performance for the hardest questions and tasks. Then came Sonnet, the workhorse with a mix of power and speed, and then Haiku for quickness and cost savings. Mythos emerged as an extra-high tier of performance, especially focused on cyber capabilities. But then it turned out to be a little too powerful and dangerous, and so Anthropic made Fable, which is basically Mythos with the cybersecurity and biological capabilities removed. All of this makes it harder to sort out which model to use, and all of the Mythos/Fable hype has made Opus feel like a mix of Sonnet and Haiku.

There are multiple benefits to figuring out how to serve up cheaper models. Along with the fact that everyone is cutting costs, Anthropic has been making deal after deal to try and secure the data center capacity it needs to run its compute-hungry Mythos and Fable models as it plays catch-up to rival OpenAI, which has been making big data center bets for the past two years. However, the bigger reason to undercut competitors on cost while keeping performance high is that eventually, cost may become the key differentiator as these models commoditize. Though the major AI labs are currently trying to position their models as distinctly better than one another, cost is likely to become a more important factor than performance, especially as these models leapfrog one another with incremental updates. Remember that there's still a huge capability overhang that separates what the models can do and what most people are using them for.

TOGETHER WITH CHECKSUM

Don't stop at CI/CD. Ship faster with continuous verification.

63% of engineering teams now ship code faster with AI. 72% have already had a production incident from AI-generated code. The bottleneck didn't disappear. It moved downstream.

Checksum is an AI-native continuous testing platform that auto-generates and self-heals your E2E test suite, runs inside your existing CI/CD pipeline, and keeps pace with the velocity you're already getting from AI coding agents.

Clearpoint Strategy saves $500k a year. Postilize ships 30% faster.

Setup takes under 30 minutes.

MARKETS

Why OpenAI's compute spree now looks prescient

OpenAI is not afraid of a compute spending spree. 

Last week, the Wall Street Journal reported that the company expects to spend $750 billion on AI infrastructure by 2030, a 25% increase from its estimates earlier this year. Also announced last week, OpenAI is planning a 3.2 gigawatt data center in Georgia with a price tag anywhere from $20 to $30 billion. 

While these numbers may give sticker shock, let's not forget that this isn't the first time OpenAI has made massive commitments to building out data centers, such as its partnership with the $500 billion Stargate project. With Stargate seemingly on ice and compute demand only ramping up further, OpenAI is finding new ways to fill the gap. 

But OpenAI is not the only one vying for raw compute power. Rival Anthropic has sought out a handful of pricey deals, largely to lease compute from other tech giants, including SpaceX, Amazon, Microsoft, Google, and most recently, Meta and AMD. 

Anthropic does, however, have some plans to build out its own data centers, including a $50 billion investment with Fluidstack to build facilities in Texas and New York. Still, comparatively, Anthropic has committed far less in constructing and owning AI infrastructure than OpenAI.

However, as both companies continue their heated race to develop and release highly-capable (and power hungry) models, coupled with the meteoric rise of token-burning agents, compute is at a premium. As Robert Bittencourt, head of thematic investing at Apollo Global, said in a recent blog post: "In a compute-constrained world, access itself becomes a competitive moat … What looked like an overreach only months ago increasingly looks like strategic foresight." 

And with its recent string of deals, Anthropic may be playing catch-up. This is evidenced by the fact that Claude users often report outages and hitting their rate limits faster than they do with OpenAI. One of the results of being compute-constrained is that Anthropic has had to taper access to its new flagship model, Fable 5. It's only available to Max and Team Premium users, and for only 50% of standard weekly usage limits, which means users have to consciously limit how they use the model. 

Still, even despite contrasting spending habits, a lot has to go right for either of these companies to be winning bets for investors, including finding substantial revenue growth that outpaces spending, Jeremy Roberts, senior director of research and content at Info-Tech Research Group, told The Deep View. 

"Ultimately, we will have to wait to review the companies’ financials as they proceed to their IPOs," said Roberts. "My speculative read is that both companies are struggling with profitability and taking on deeper and deeper commitments in an effort to win the compute race."

Initially, the massive, trillion-dollar price tag that OpenAI committed to compute at one point raised all kinds of doubts and red flags about its future. However, now the company's valuation is scraping a trillion ahead of its IPO, and the industry is eating up every bit of compute that it can get its hands on. And with its primary rival making deal after deal to lease compute power, OpenAI's spending spree may have been the right call, while Anthropic may be kicking itself for taking the conservative approach to the AI infrastructure buildout. Additionally, owning the entire stack is an advantage in and of itself. OpenAI is embedding itself within every layer of a technology that it claims is on course to transform society. In this case, owning the literal power means owning figurative power, too. 

TOGETHER WITH CDATA

Spending too much on Claude API calls?

It’s not necessarily a given, it might be how your data layer is set up.

We ran 56 live benchmark tests on a real enterprise query spanning Salesforce, Snowflake, and ServiceNow, and found that the right architecture cuts token usage by up to 97.6%, with per query cost dropping from $0.596 to $0.027 in some cases. 

CData Connect AI makes that possible. Curious how?

Check out the full breakdown → 

RESEARCH

Why AI must learn how to say 'I don’t know'

AI systems almost always have an answer, even when they should say, “I don’t know.”

In this episode of The Deep View Conversations, host Sabrina Ortiz speaks with Ruchir Puri, chief scientist at IBM Research, about why uncertainty modeling may be AI’s most urgent technical challenge.

Puri explains why today’s models struggle to recognize the limits of their own knowledge, how that failure contributes to hallucinations, and what researchers must solve before AI can become more reliable. He also explores the need for self-improving models, the enormous energy gap between artificial and human intelligence, and why the future of AI depends on doing more with less compute.

The conversation also covers:

• Why Puri predicted in 2020 that AI would transform software development

• How big data, GPUs, and transformer architectures created the current AI boom

• Why intelligence involves more than IQ

• The roles of emotional and relationship intelligence

• Why language models cannot capture the full complexity of the physical world

• How AI could help redesign software, quantum computing, and chip development

• Why Puri prefers "artificial useful intelligence" over AGI

Rather than chasing abstract definitions of general intelligence, Puri argues that the industry should focus on building AI that is useful, efficient, adaptable, and honest about what it does not know.

📺 Watch on YouTube
🎧 Listen in your favorite podcast player

Subscribe to Deep View Conversations for interviews with the leaders shaping the future of AI, business, and technology: tdv.transitor.fm 

LINKS

• Z.AI's GLM 5.2 helped stop OpenAI's Hugging Face cyberattack

• Lawsuit claims ChatGPT told man not to seek help for pulmonary embolism

• Google research finds that AI is not replacing workers, but assisting them

• Intel sees substantial growth as AI data centers fuel sales

• Jensen Huang donates $75 million to Vanderbilt art school 

• Midjourney acquires horoscope app Co-Star

• Runway Agent: The video platform's agent now offers a workflow tool for building, running or editing node-based workflows through natural language.

• Canva Code 2.0: Users can create and edit interactive websites through Canva, as well as import HTML, match their brand, and swap in visuals from their library.

• Meta Seller: Meta launched a free standalone app version of Facebook Marketplace. 

• ChatGPT Work agent: can now use websites that require you to sign in

• Google DeepMind: Research Engineer, Human Understanding

• Deloitte: Agentic Capability AI and Data Science Engineer

• Accenture: Google Agentic AI Delivery Senior Engineer

• EY: Forward Deployed AI Engineer - Associate Director

GAMES

A QUICK POLL BEFORE YOU GO