Last summer, mining startup KoBold made a splash when it said it had discovered in Zambia one of the world’s largest copper deposits in more than a decade.

Now, another startup, Earth AI, exclusively told TechCrunch about its own discovery: promising deposits of critical minerals in parts of Australia that other mining outfits had ignored for decades. While it’s still not known whether they are as large as KoBold’s, the news suggests that future supplies of critical minerals are likely to emerge from a combination of field data parsed by artificial intelligence.

“The actual, real frontier [in mining] is not so much geographical as it is technological,” Roman Teslyuk, founder and CEO of Earth AI, told TechCrunch. 

Earth AI has identified deposits of copper, cobalt, and gold in the Northern Territory and silver, molybdenum, and tin at another site in New South Wales, 310 miles (500 kilometers) northwest of Sydney.

Earth AI emerged from Teslyuk’s graduate studies. Teslyuk, a native of Ukraine, was working toward a doctorate at the University of Sydney, where he became familiar with the mining industry in Australia. There, the government owns the rights to mineral deposits, and it leases them in six-year terms. Since the 1970s, he said, exploration companies are required to submit their data to a national archive.

“For some reason, nobody’s using them,” he said. “If I could build an algorithm that can absorb all that knowledge and learn from the failures and successes of millions of geologists in the past, I can make much better predictions about where to find minerals in the future.”

Teslyuk started Earth AI as a software company focused on making predictions about potential deposits, then approaching customers who might be interested in exploring sites further. But the customers were hesitant to invest, in part because they didn’t want to bet millions on the predictions of an unproven technology.

“Mining is a very conservative industry,” Teslyuk said. “Everything outside of the approved dogma is considered heresy.”

So Earth AI decided to develop its own drilling equipment to prove that the sites it identified were as promising as its software suggested. The company was accepted to Y Combinator’s spring 2019 cohort, and it spent the next few years refining its hardware and software. In January, Earth AI raised a $20 million Series B.

Though the company uses AI to search for minerals like KoBold, Teslyuk says it takes a different tack. Earth AI’s algorithms, he said, are trained to scan wide areas quickly and efficiently to find deposits that might otherwise have been overlooked.

“The way we used to explore for metals in the past, the 20th century, it just takes very, very long. It takes decades to find something,” Teslyuk said. “With the modern pace of the world, you just can’t wait for that long.”

During a livestream on Tuesday, OpenAI CEO Sam Altman announced the first major upgrade to ChatGPT’s image-generation capabilities in over a year.

ChatGPT can now leverage the company’s GPT-4o model to natively create and modify images and photos. GPT-4o has long underpinned the AI-powered chatbot platform, but until now, the model has been able to generate and edit only text — not images.

Altman said GPT-4o native image generation is live today in ChatGPT and Sora, OpenAI’s AI video-generation product, for subscribers to the company’s $200-a-month Pro plan. OpenAI says the feature is rolling out soon to Plus and free users of ChatGPT, as well as developers using the company’s API service.

GPT-4o with image output “thinks” a bit longer than the image-generation model it effectively replaces, DALL-E 3, to make what OpenAI describes as more accurate and detailed images. GPT-4o can edit existing images, including images with people in them — transforming them or “inpainting” details like foreground and background objects.

To power the new image feature, OpenAI told the Wall Street Journal it trained GPT-4o on “publicly available data,” as well as proprietary data from its partnerships with companies like Shutterstock.

Many generative AI vendors see training data as a competitive advantage, so they keep it and any information related to it close to the chest. But training data details are also a potential source of IP-related lawsuits, another disincentive for companies to reveal much. 

“We’re respecting of the artists’ rights in terms of how we do the output, and we have policies in place that prevent us from generating images that directly mimic any living artists’ work,” said Brad Lightcap, OpenAI’s chief operating officer, in a statement to the Journal.

OpenAI offers an opt-out form that allows creators to request that their works be removed from its training datasets. The company also says that it respects requests to disallow its web-scraping bots from collecting training data, including images, from websites.

ChatGPT’s upgraded image-generation feature follows on the heels of Google’s experimental native image output for Gemini 2.0 Flash, one of the company’s flagship models. The powerful feature went viral on social media — but not necessarily for the best reasons. Gemini 2.0 Flash’s image component turned out to have few guardrails, allowing people to remove watermarks and create images depicting copyrighted characters.

This article was update at 12pm PT to include OpenAI’s statement to the Wall Street Journal around GPT-4o’s training data.

On Tuesday, Google unveiled Gemini 2.5, a new family of AI reasoning models that pauses to “think” before answering a question.

To kick off the new family of models, Google is launching Gemini 2.5 Pro Experimental, a multimodal, reasoning AI model that the company claims is its most intelligent model yet. This model will be available on Tuesday in the company’s developer platform, Google AI Studio, as well as in the Gemini app for subscribers to the company’s $20-a-month AI plan, Gemini Advanced.

Moving forward, Google says all of its new AI models will have reasoning capabilities baked in.

Since OpenAI launched the first AI reasoning model in September 2024, o1, the tech industry has raced to match or exceed that model’s capabilities with their own. Today, Anthropic, DeepSeek, Google, and xAI all have AI reasoning models, which use extra computing power and time to fact-check and reason through problems before delivering an answer.

Reasoning techniques have helped AI models achieve new heights in math and coding tasks. Many in the tech world believe reasoning models will be a key component of AI agents, autonomous systems that can perform tasks largely sans human intervention. However, these models are also more expensive.

Google has experimented with AI reasoning models before, previously releasing a “thinking” version of Gemini in December. But Gemini 2.5 represents the company’s most serious attempt yet at besting OpenAI’s “o” series of models.

Google claims that Gemini 2.5 Pro outperforms its previous frontier AI models, and some of the leading competing AI models, on several benchmarks. Specifically, Google says it designed Gemini 2.5 to excel at creating visually compelling web apps and agentic coding applications.

On an evaluation measuring code editing, called Aider Polyglot, Google says Gemini 2.5 Pro scores 68.6%, outperforming top AI models from OpenAI, Anthropic, and Chinese AI lab DeepSeek.

However, on another test measuring software dev abilities, SWE-bench Verified, Gemini 2.5 Pro scores 63.8%, outperforming OpenAI’s o3-mini and DeepSeek’s R1, but underperforming Anthropic’s Claude 3.7 Sonnet, which scored 70.3%.

On Humanity’s Last Exam, a multimodal test consisting of thousands of crowdsourced questions relating to mathematics, humanities, and the natural sciences, Google says Gemini 2.5 Pro scores 18.8%, performing better than most rival flagship models.

To start, Google says Gemini 2.5 Pro is shipping with a 1 million token context window, which means the AI model can take in roughly 750,000 words in a single go. That’s longer than the entire “Lord of The Rings” book series. And soon, Gemini 2.5 Pro will support double the input length (2 million tokens).

Google didn’t publish API pricing for Gemini 2.5 Pro. The company says it’ll share more in the coming weeks.

Poe, Quora’s chatbot app, launched one of its most affordable subscription options on Tuesday, priced at just $5 per month. 

In addition, the company introduced its highest-priced plan at $250 per month, designed for users who need to send a large volume of messages on Poe.

Poe allows users to utilize several AI-powered bots — including DeepSeek-R1, GPT-4o, Claude 3.7 Sonnet, o3-mini, ElevenLabs, and more — in one place. It operates on a point system, enabling users to spend points across different models, with each bot having its own point cost per message.

Under the new $5/month plan, users can spend up to 10,000 points per day. In contrast, the $250/month tier offers 12.5 million points, which the company says is better for more “expensive” models, such as GPT-4.5, OpenAI’s o1-pro, and Google DeepMind’s Veo 2.

By popular demand, we are introducing two new subscription options today, at $5/month and $250/month. These align Poe with two simultaneous trends in AI: normal models are getting cheaper and the most advanced models are getting more expensive. (1/5) pic.twitter.com/1sUqOMdWfs

— Poe (@poe_platform) March 25, 2025

With the introduction of these two new subscription plans, users now have a wider variety of options, which the company says was highly requested.

Previously, the least expensive tier was $20 per month, which provided 1 million points. A free plan is also available, but users can only ask a limited number of questions each day.

Poe is available on iOS, Android, Mac, and Windows.

As the year of the AI agent takes shape, a new trend is emerging: startups offering the picks and shovels that help employers build a workforce of bots.

Manny Medina, best known as the founder and former CEO of the $4.4 billion valued sales automation company Outreach, just launched one such startup called Paid, he told TechCrunch exclusively.

Paid doesn’t make AI agents. It offers a platform that makes sure they get paid, profitably. Paid announced Monday that it raised €10 million (about $11 million) in a pre-seed investment from European powerhouse EQT Ventures, Sequoia, and GTMFund.

Medina came up with the idea for Paid after spending months talking to dozens of agentic platform startups. In these conversations, a common complaint emerged. “They didn’t really know what to charge,” Medina told TechCrunch.

The premise of Paid is that the old ways of charging for software won’t work with AI agents. Agentic companies can’t charge per user or per seat, meaning based on how many people are using the software (like old-school Microsoft Office). The whole point is that one employee could run lots of agents. Or agents will run by themselves with no human overseer at all.

Companies developing AI agents also can’t charge like the last big generational change in software, SaaS, charging by usage because, if agents work properly, they “are taking over a whole role,” Medina says. 

An agent’s customer doesn’t want to pay for all the discrete tasks an agent does — if it even knows them all, he says. They want to pay for its results, like an employee. So if an agent is hired in insurance and the role’s success is measured in completed policy renewals, a company doesn’t want to pay for each email the agent sent.

At the same time, the costs associated with providing agents are variable, depending on how many LLM tokens it needs to execute its training and its tasks.

“So how do you help them price for the job that they’re delivering?” Medina said of the startups offering agents. “They needed the ability to try new things with different customers. They needed the ability to measure their margins.”

Billing meets HR management

Agents are so new that startups haven’t had to deal with processes that provide profitable billing, let alone renewals. Paid allows agentic startups to create pricing — fixed or variable — with an eye to profitable margins. 

In doing so, it also tracks agents’ output, which also lets startups validate the return on investment.

It’s the AI agent era version of Zuora (SaaS renewal billing software) meets SuccessFactors (SaaS HR management software).

The Paid platform is being marketed to startups, rather than enterprises like Salesforce and Microsoft, which are also offering agentic platforms. Paid has three such companies as beta customers, it says: Logic.app, 11x, VidLab7, Artisan, and HappyRobot.

“Agents are replacing roles, human roles, not the entire job, but entire roles,” Medina says.

He’s also practicing what he preaches, using AI to build this new startup. Paid engineers vibe coded the initial product demos with tools like v0, Replit, and Lovable.

“This is what is so much fun about building a company right now. We have two engineers, and we have built the entirety of the building platform in a month. Why? Because we build everything on AI,” he said. 

Medina has experience building companies from nothing. The former Microsoftie, who has been a well known part of the Seattle tech scene for decades, took Outreach from $0 when he founded it in 2011 to 800 employees and $250 million in annual recurring revenue by the time he left the CEO role in September.

Medina left the executive chairman role in March, though he remains on the board. He, and Paid, are now based in London.

As the year of the AI agent takes shape, a new trend is emerging: startups offering the picks and shovels that help employers build a workforce of bots.

Manny Medina, best known as the founder and former CEO of the $4.4 billion valued sales automation company Outreach, just launched one such startup called Paid, he told TechCrunch exclusively.

Paid doesn’t make AI agents. It offers a platform that makes sure they get paid, profitably. Paid announced Monday that it raised €10 million (about $11 million) in a pre-seed investment from European powerhouse EQT Ventures, Sequoia, and GTMFund.

Medina came up with the idea for Paid after spending months talking to dozens of agentic platform startups. In these conversations, a common complaint emerged. “They didn’t really know what to charge,” Medina told TechCrunch.

The premise of Paid is that the old ways of charging for software won’t work with AI agents. Agentic companies can’t charge per user or per seat, meaning based on how many people are using the software (like old-school Microsoft Office). The whole point is that one employee could run lots of agents. Or agents will run by themselves with no human overseer at all.

Companies developing AI agents also can’t charge like the last big generational change in software, SaaS, charging by usage because, if agents work properly, they “are taking over a whole role,” Medina says. 

An agent’s customer doesn’t want to pay for all the discrete tasks an agent does — if it even knows them all, he says. They want to pay for its results, like an employee. So if an agent is hired in insurance and the role’s success is measured in completed policy renewals, a company doesn’t want to pay for each email the agent sent.

At the same time, the costs associated with providing agents are variable, depending on how many LLM tokens it needs to execute its training and its tasks.

“So how do you help them price for the job that they’re delivering?” Medina said of the startups offering agents. “They needed the ability to try new things with different customers. They needed the ability to measure their margins.”

Billing meets HR management

Agents are so new that startups haven’t had to deal with processes that provide profitable billing, let alone renewals. Paid allows agentic startups to create pricing — fixed or variable — with an eye to profitable margins. 

In doing so, it also tracks agents’ output, which also lets startups validate the return on investment.

It’s the AI agent era version of Zuora (SaaS renewal billing software) meets SuccessFactors (SaaS HR management software).

The Paid platform is being marketed to startups, rather than enterprises like Salesforce and Microsoft, which are also offering agentic platforms. Paid has three such companies as beta customers, it says: Logic.app, 11x, VidLab7, Artisan, and HappyRobot.

“Agents are replacing roles, human roles, not the entire job, but entire roles,” Medina says.

He’s also practicing what he preaches, using AI to build this new startup. Paid engineers vibe coded the initial product demos with tools like v0, Replit, and Lovable.

“This is what is so much fun about building a company right now. We have two engineers, and we have built the entirety of the building platform in a month. Why? Because we build everything on AI,” he said. 

Medina has experience building companies from nothing. The former Microsoftie, who has been a well known part of the Seattle tech scene for decades, took Outreach from $0 when he founded it in 2011 to 800 employees and $250 million in annual recurring revenue by the time he left the CEO role in September.

Medina left the executive chairman role in March, though he remains on the board. He, and Paid, are now based in London.