o3: Faster, Cheaper, Smarter

OpenAI unveiled its new reasoning model, o3-mini, on Friday. The model is designed for technical challenges in STEM, programming, math, and science. It is both powerful and affordable.

O3-mini is an upgrade over the o1 family. Internal tests show that external reviewers preferred o3-mini’s responses over those from o1-mini more than half the time. Its answers come 24% faster and have 39% fewer major mistakes. It fact-checks itself thoroughly—a process that enhances reliability despite slightly longer processing times.

The model is now available in ChatGPT. Paid subscribers, including ChatGPT Plus and Team users, enjoy a higher rate limit of 150 queries per day. ChatGPT Pro subscribers have unlimited access. Soon, Enterprise and Edu customers will also benefit from o3-mini. Developers can select the model via OpenAI’s API. They can adjust the reasoning effort—low, medium, or high—depending on latency needs. Pricing is competitive: $0.55 per million cached input tokens and $4.40 per million output tokens, a saving of 63% compared to o1-mini.

OpenAI emphasizes that o3-mini strikes a balanced trade-off between speed, accuracy, and cost. The medium reasoning mode, which is the default in ChatGPT, is designed to meet most users’ needs. Premium users can opt for a high reasoning mode that offers sharper responses at the expense of speed. While o3-mini does not outclass every benchmark, it delivers improvements where it counts.

This release comes as OpenAI faces stiff competition from companies like DeepSeek. Allegations of IP theft and challenges in the AI race add pressure to the launch. Still, OpenAI’s latest model marks another step toward making advanced, reliable AI accessible to all.

Google Unveils Gemini 2.0 Pro Experimental

Google quietly launched Gemini 2.0 Pro Experimental.

The company revealed it in a changelog for its Gemini chatbot app. No big fanfare. Just a low-key update.

Gemini 2.0 Pro Experimental succeeds the Gemini 1.5 Pro model, which debuted last February. It now leads the Gemini AI family.

Google promises “better factuality” and “stronger performance” for coding and math tasks. The new model is tailored for complex coding challenges and advanced statistical or quantum algorithm development. It helps users tackle technical tasks with improved ease and accuracy.

The launch comes at a time when attention is focused on Chinese AI startup DeepSeek. DeepSeek’s models are widely available and match or surpass many American counterparts. This rivalry has stirred both Silicon Valley and U.S. government circles.

Gemini 2.0 Pro Experimental is available exclusively to Gemini Advanced users starting Thursday. Gemini Advanced is part of Google One AI Premium, and it also integrates with Google Workspace add-ons.

Google cautions that Gemini 2.0 Pro Experimental is in “early preview.” Unexpected behaviors are possible. Mistakes can happen. Unlike other Gemini models, it lacks real-time information access and isn’t compatible with some app features.

The changelog states, “We believe in rapid iteration and bringing the best of Gemini to the world.” Google is eager for feedback from Gemini Advanced subscribers. Their insights will help refine the model and shape future wider releases.

Alongside this release, Google introduced the Gemini 2.0 Flash model for all Gemini app users. For now, Gemini Flash remains the default model.

In summary, Google’s rollout of Gemini 2.0 Pro Experimental is a subtle yet significant step. It aims to enhance performance for technical tasks and keep pace with emerging global competition in AI.

Meta Doubles Down on AI Amid GPU Fears

U.S. markets panicked on Monday.

Speculation swirled that DeepSeek’s AI models would crush GPU demand. Nvidia’s stock fell almost 20%.

Meta, however, is not stepping back. During its first-quarter earnings call on Wednesday, CEO Mark Zuckerberg pledged that the company would invest “very heavily” in AI—even “hundreds of billions of dollars” over the long term.

Last week, Zuckerberg announced Meta’s plan to spend more than $60 billion in 2025 alone on capital expenditures. Most of that will go toward data centers.

The company sees robust AI infrastructure spending as a major strategic advantage.

When an analyst asked about DeepSeek’s impact on Meta’s chip demand, Zuckerberg remained confident. He pointed out that chips remain crucial for inference purposes, especially with billions of users relying on the service.

“Building out that kind of infrastructure is going to be a major advantage for both quality and scale,” he emphasized.

Meta also has plans for its next AI model, Llama 4.

The goal is clear: to become the world’s most competitive AI model—even against closed systems like ChatGPT. Zuckerberg expects Llama 4 to feature agentic and multimodal capabilities, aligning with moves by both OpenAI and Anthropic.

He reflected on the progress from Llama 3. “Our goal with Llama 3 was to make open source competitive. With Llama 4, we aim to lead,” he stated.

Even as market jitters continue, Meta’s vision remains unwavering.

The company is committed to a long-term, heavy investment in AI infrastructure—a bet that could redefine its future.

DeepSeek Faces Global Access Blockades

Companies and government agencies worldwide are cutting off access to DeepSeek’s tools.

Cybersecurity firms report that hundreds of organizations—especially those linked to governments—are blocking DeepSeek. Concerns center on potential data leaks and weak privacy safeguards.

Nearly 70% of Armis Inc. customers have requested blocks, and over half of Netskope Inc. clients have done the same.

The fear? That sensitive information might end up in the wrong hands, possibly reaching the Chinese government.

Nadir Izrael of Armis warns, “You don’t know where your information goes.”

DeepSeek’s own privacy terms reveal that data is stored on servers in China.
This means any dispute would be governed by Chinese law. It raises alarm bells for companies and regulators alike.

The backlash began after DeepSeek’s AI chatbot soared to the top of Apple Store downloads. High-profile tech executives even praised the tool. But the praise quickly turned to concern.

Cyber researchers uncovered a publicly accessible database belonging to DeepSeek.
This database contained chat histories and technical log data.
Wiz Inc. found the data before DeepSeek secured it.

Governments are also acting. Italy’s privacy regulator ordered an immediate block on DeepSeek. Ireland’s Data Protection Commission has requested more information to assess the risk.

In the UK, the Information Commissioner stressed transparency from AI developers regarding personal data use.

Meanwhile, U.S. officials and think tanks warn that Chinese national security laws can force companies to hand over encryption keys. These laws once fueled efforts to ban ByteDance’s TikTok.

DeepSeek’s low prices still threaten to lure developers away from competitors like OpenAI, despite security concerns.

Cybercrime experts caution that DeepSeek’s AI services lack robust safeguards.
Hackers might exploit these tools to craft phishing emails or analyze stolen data.

As one expert put it, minimal guardrails could mean big problems.