PRODUCTS

Sonnet 5 is Anthropic's answer to AI sticker shock

Despite its latest kerfuffle with the US government over the release of its Mythos and Fable models, Anthropic isn't slowing down.

On Tuesday, the company unveiled Sonnet 5, the latest addition to its Claude model family and the "most agentic" version yet. The company said the model significantly narrows the performance gap between efficient and more powerful models, delivering capabilities on par with larger models at a lower cost. 

Sonnet 5 can make plans, use tools like browsers and terminals, and run autonomously in ways that would have previously required more expensive systems, Anthropic said. 

• The company noted Sonnet 5 provides "substantial improvement" over its predecessor, Sonnet 4.6, showing an overall lower rate of "undesirable behaviors" and is safer to use in agentic contexts. 

• Though it performs close to Opus 4.8, Anthropic's most powerful publicly available model, the company noted that it has a "much lower ability" to perform cybersecurity tasks than its Opus lineup. 

• Starting today, Sonnet 5 is available across all plans and is the default model for Free and Pro users. On Claude Code and Claude Platform, the introductory pricing is $2 per million input tokens and $10 per million output tokens, and will be bumped to $3 per million input tokens and $15 per million output tokens starting in August. 

Early testers include Cursor, Rakuten, Lovable and Factory, with testers broadly impressed with the model's ability to finish complex tasks, check its own outputs and complete agentic work at a good price. 

"With Claude Sonnet 5, agents stay on plan, follow our conventions and ship clean multi-step changes, all at an efficient cost," Sualeh Asif, co-founder of Cursor, said in the company's blog post. 

Anthropic's cost-friendly model comes at a time when the tokenmaxxing fad is catching up with enterprises. More than ever, companies are considering the untenable costs of AI, with Gartner predicting that AI spending will reach nearly $2.6 trillion this year, up 47% from the prior year. As a result, enterprises are experiencing AI sticker shock. Meanwhile, AI vendors have started to shift their focus to cost efficiency.

Model providers like OpenAI and Anthropic stood to benefit the most from the tokenmaxxing trend. However, with the industry largely recoiling from this mindset after getting hit with massive AI usage bills, model providers may be taking notice that the days of unlimited token gluttony are over. The efficiency-focused design of Sonnet 5 may be a sign that Anthropic wants to keep up with the changing tide. However, efficient and cost-effective AI might not be a bad thing for Anthropic, either. Customers burning through fewer tokens means Anthropic can save on compute, which is vital as demand rises and the shortage worsens. Plus, with Mythos and Fable still caught up in government red tape, the company likely wants to keep its audience hooked, and Sonnet 5 is more economically attractive to large customers. 

TOGETHER WITH FUNCTION

When it comes to health, your AI is better with Function.

ChatGPT, Claude, and other AI platforms are incredible tools. But when it comes to health, AI is running off generic population data instead of your unique biology. Not your hormone levels. Not your cortisol.

Function changes that. When you securely connect Function to your AI tools, you can bring your lab results and actual biology into the conversation.

Smarter inputs. Smarter responses.

Connect Function to your AI tools

CULTURE

Google races to define creative AI

Following the rising popularity of its Nano Banana image model, Google wants to keep the momentum rolling. 

On Tuesday, the company announced two new additions to its generative AI creative suite: the general availability of Nano Banana 2 Lite and the public preview of Gemini Omni Flash. 

The models, now available in the Gemini Enterprise Agent Platform, aim to give users more speed, control and cost efficiency in their creative work, the company said in its announcement.

Here's the breakdown:

• Nano Banana 2 Lite, or Gemini 3.1 Flash-Lite Image, is the fastest and most cost-efficient image generation and editing model in the Nano Banana family. The model can generate an image in as little as four seconds, and rapidly iterate design concepts or image edits. The model also features significant improvements in visual quality, including better world knowledge, character consistency, and localized text by region. 

• Gemini Omni Flash offers high-quality video generation and conversational editing, embedded directly into media tools and apps. The model also includes multimodal outputs, improved simulation and physics understanding, and improved text generation. The company said that support for audio references, video references, last frame, scene extension and higher resolutions will be available soon.

Google said that several companies are already testing both Nano Banana 2 Lite and Gemini Omni Flash, including Adobe, Klarna, Picsart, Genspark, Figma, Manus and others. Itay Schiff, Figma Weave co-founder and creative director, said in a statement that Nano Banana 2 Lite is "ideal for rapid iteration while staying in the creative flow." 

Nishant Tahilramani, creative director at Invideo, an AI video platform that tested the model, said in a statement that the "sheer range" of Gemini Omni Flash caught his attention. "The VFX capabilities surprised me, and looking at it as a producer, that brings in some very interesting possibilities … You take the crews you have always worked with in the live-action world, and you bring the breadth of what AI can do now onto the same set.”

The models are Google's latest foray into the creative fields. Last week, the company's AI unit, Google DeepMind, announced a partnership and a $75 million investment in A24, the film studio responsible for Backrooms, Hereditary and Midsommar, a decision that stoked the ire of fans and Hollywood professionals alike. 

"This partnership exists because we want to dictate what tools get built for artists, and so they have a voice in shaping them rather than having tools handed to them,"  Sophia Shin, an A24 spokesperson, told Wired.

As it stands, AI is yanking creative industries in two directions. There's a large contingent fighting against it for reasons of copyright and creative integrity, and there's also a group that believes the tech can democratize the arts and give more power to hobbyists. With these new models and its industry partnerships, Google is walking straight into the firestorm. The thing that could give it more legitimacy in the creative industries, however, is its SynthID, Google DeepMind's watermarking system for synthetic content. With AI-generated content growing more lifelike and realistic by the day, people will increasingly be unable to tell the difference between authentic and synthetic content. However, if Google's tools, whether it be the models themselves or its synthetic content watermarking platform, become an entertainment industry standard, the industry and the population at large may stand a better chance at distinguishing real from fake. 

TOGETHER WITH KALSHI

The First Regulated Perps Market in the US Is Live

Perpetual futures (perps) are now live on Kalshi. Offshore, perps do ~$90 trillion in volume annually — until now, none of it was legal in the US.

Trade on whether an asset's price goes up or down. No expirations, no rolling costs, no complex payoffs.

Try Kalshi Perps today.

RESEARCH

Anthropic eyes biology after winning at code

If Anthropic co-founder Dario Amodei is right, AI will eventually be known more for the scientific breakthroughs it enables than the negative societal risks it introduces. 

On Tuesday, Anthropic announced what it's calling its "AI workbench for scientists," now dubbed Claude Science. The app launches in beta today and runs locally on Mac or Linux, or on a remote machine via SSH or an HPC login. It runs on Pro, Max, Team, and Enterprise plans. Anthropic also launched a new team plan for research labs.

It's important to note that Claude Science uses all of the existing Anthropic models and does not introduce any new science-specific models like the one OpenAI has created with GPT-Rosalind. Instead, Claude Science is focused on a science-specific interface that brings together tools and capabilities that can accelerate the work of scientists, researchers, students, and labs. It also runs science-specific agents that are designed to automate some of the most tedious and time-consuming parts of research. 

Anthropic said that Claude Science can run locally on a laboratory's own servers and doesn't need to run through the standard cloud-based Anthropic models. This is a key compromise since labs often run highly proprietary, valuable, and sensitive data that demand strong security, privacy, and data sovereignty. 

Another one of the key issues Anthropic is trying to accomplish with Claude Science is the problem of scientific work being spread across a disconnected set of tools, databases, and computing environments, which creates friction and obstacles to achieving important breakthroughs. 

In the livestream, the Anthropic team made the case that AI has completely transformed the work of software development, and that it can have a similar impact on life sciences. And that's where Claude Science will start, with the goal of "compressing timelines in biology."

Amodei has long championed the potential of accelerating scientific progress as one of AI's greatest opportunities to make a positive impact on civilization. Tuesday's announcement was introduced in a rare major public event for Anthropic and livestreamed on YouTube. 

To stimulate use of its new science tools, Anthropic also launched a new AI for Science Program that will support up to 50 projects and provide $30,000 in Claude Science credits. This is aimed at nonprofits and academic institutions, which can apply online by July 15, with winners notified by July 31 and projects running from September 1 to December 1.

The Deep View has covered the launch of Google's Gemini for Science and OpenAI's GPT-Rosalind model recently and the level of enthusiasm for the impact that AI could have on scientific discovery has bordered on hyperbolic. Anthropic's Claude Science launch matched that exuberance. Clearly, there's a ton of enthusiasm from both AI labs and scientific teams for how AI can streamline some of the most entrenched problems that bog down scientific research, especially in biology. It's reasonable that AI could assist with streamlining research and documentation. It also makes sense that it could allow scientific teams to model out more hypotheses and experiments simultaneously, as the head of Google Research Yossi Matias recently told me in an interview for The Deep View. But we should also keep in mind that biology and medical trials take a long time, sometimes up to a decade, to test on human subjects. And while AI might be able to simulate it, it's unlikely that it will accelerate those trials. And so we should be wary about any claims of AI dramatically accelerating the elimination of human disease or solving the problems of aging in the next few years. 

LINKS

• OpenAI reportedly has discovered a new way to cut inference costs in half

• Study finds heavy AI spenders are hiring staff faster than peers

• AWS invests $1 billion in forward deployed engineering organization

• US political campaigners are using AI to analyze voter data 

• Vibe coding has fueled a surge in mobile game development

• AI inference startup Etched raises $800 million at $5 billion

• Baz Planner: By startup Baz, this product sits between developers and codebases to automatically catch vulnerabilities.

• Cursor for iOS: Cursors mobile app lets you build from anywhere by launching always-on cloud agents.

• Short Video Overviews in NotebookLM: Turn complex sources into 60-second, vertical videos that deep dive into any concept.

• Seed Audio 1.0: ByteDance's new text-to-speech model that generates voice, music, and sound effects in a single pass.

• xAI: Member of Technical Staff - Post-Training and RL

• Nvidia: Research Scientist, Security and Privacy

• Reflection: Member of Technical Staff - Safety

• Glean: Machine Learning Engineer, LLM Evals & Observability

GAMES

POLL RESULTS

Do you think AI will replace or augment blue collar and manufacturing work?

Yes (24%)
Maybe, but not yet (57%)
No (14%)
Other (5%)