OpenAI Drops o3 Model That Makes PhD-Level Researchers Nervous
If you thought AI peaked with ChatGPT writing your emails, this week just changed the game entirely. OpenAI released benchmarks for its upcoming o3 model that scored 87.7% on the ARC-AGI test—a benchmark specifically designed to measure AI’s ability to reason through novel problems it’s never seen before. For context, the previous record was 55%, and most humans score around 85%.
The o3 model isn’t publicly available yet, but OpenAI demonstrated it solving complex mathematical proofs and scientific reasoning tasks that typically require graduate-level training. This isn’t about pattern matching or regurgitating training data—it’s about genuine logical reasoning through unfamiliar problems.
What Makes o3 Different From GPT-4
Unlike the conversational models you’re used to, o3 belongs to OpenAI’s reasoning-focused lineup that started with o1. These models literally think before they respond, spending extra compute time on multi-step reasoning instead of instantly generating text. When you ask o3 a complex question, it might take 30 seconds to respond—but that response is far more likely to be correct.
The model comes in two flavors: o3 and o3-mini. The full o3 model can use what OpenAI calls “high compute” mode, where it dedicates significantly more processing power to particularly difficult problems. In this mode, it achieved those record-breaking scores, though at a cost that would make your AWS bill weep.
The catch? OpenAI hasn’t announced pricing or a public release date yet. They’re currently allowing safety researchers to test it before wider deployment, which brings us to the elephant in the room: these reasoning capabilities are powerful enough that even OpenAI seems cautious about rushing it out the door.
Google Quietly Restructures Its Entire Gemini Strategy
While everyone was distracted by OpenAI’s announcements, Google made several significant moves with its Gemini lineup this week. The company consolidated its previously fragmented AI offerings and introduced Gemini 2.0 Flash, which it’s positioning as the model for agentic AI applications.
Translation: Google wants Gemini to actually do things for you, not just chat. Gemini 2.0 Flash can control software interfaces, search the web in real-time, and execute multi-step tasks with less hand-holding than previous versions.
The Real Numbers Behind Gemini 2.0
According to TechCrunch, Gemini 2.0 Flash processes multimodal inputs—text, images, video, and audio—simultaneously rather than converting everything to text first. This architectural change makes it roughly 2x faster than Gemini 1.5 Pro at similar tasks while maintaining comparable accuracy.
Google is offering Gemini 2.0 Flash through its API at $0.10 per million input tokens and $0.40 per million output tokens. For comparison, that’s about half the cost of GPT-4 Turbo for similar capability levels. The free tier in Google AI Studio includes 50 requests per day with rate limits that reset every 60 seconds.
The limitation nobody’s talking about: Gemini 2.0 still struggles with sustained reasoning over very long contexts. While it can technically handle up to 1 million tokens of context, its performance degrades noticeably after about 100,000 tokens—roughly 75,000 words. If you need true long-document analysis, you’re still better off with Anthropic’s Claude.
Anthropic’s Claude Achieves RSP Level 2 Safety Classification
Speaking of Anthropic, the company announced this week that Claude 3.5 Sonnet has been evaluated under its Responsible Scaling Policy and classified as ASL-2 (AI Safety Level 2). This sounds bureaucratic until you realize what it means: Anthropic is publicly committing that this model doesn’t pose catastrophic risks in areas like bioweapons development or autonomous cyber attacks.
The announcement came with detailed technical documentation showing how Anthropic tested Claude’s capabilities in sensitive domains. According to their published results, Claude 3.5 Sonnet can’t meaningfully accelerate an expert’s ability to create dangerous biological agents, and its cybersecurity capabilities remain below the threshold where it could autonomously find and exploit zero-day vulnerabilities.
Why Safety Benchmarks Actually Matter Now
Here’s why this isn’t just corporate PR: Anthropic published the specific tests they used and the threshold scores that would trigger escalation to ASL-3. That means competitors and researchers can verify their methodology and hold them accountable. It’s the AI equivalent of showing your work in math class.
Wired reported that several AI labs are now adopting similar frameworks, creating an informal industry standard for safety evaluation. When models reach ASL-3, they’ll require significantly more restricted access and security protocols—think government contractor levels of operational security.
The honest limitation: these safety evaluations only measure current capabilities, not potential capabilities that might emerge from clever prompting or jailbreaking. Anthropic acknowledges this and plans quarterly re-evaluations as red-teaming techniques evolve.
Microsoft Integrates DALL-E 3 Directly Into Office Apps
Microsoft announced this week that DALL-E 3 image generation is now natively integrated into Word, PowerPoint, and Outlook for Microsoft 365 commercial customers. You can now generate custom images directly in your documents without leaving the app or using external tools.
The feature works through natural language prompts in a sidebar, and generated images automatically resize to fit your document layout. Microsoft is using a content credentials system that watermarks AI-generated images with cryptographic signatures—partly for transparency, partly to comply with emerging EU regulations around synthetic media.
The Pricing Breakdown You Actually Need
This feature requires a Microsoft 365 Copilot license, which costs $30 per user per month on top of your existing Microsoft 365 subscription. For most business plans, that means you’re paying around $50-60 per user per month total. You get 100 DALL-E image generations per user per month included in that price.
The catch that Microsoft buried in the fine print: images generated in commercial Office applications can’t be used for marketing or public-facing materials without an additional commercial use license. They’re technically licensed only for internal business communications. If you want unrestricted commercial rights, you still need to generate images through OpenAI’s DALL-E directly at $0.04-0.08 per image depending on resolution.
Perplexity Launches Publisher Program After Plagiarism Accusations
Perplexity, the AI search engine that’s been growing rapidly as a Google alternative, launched a revenue-sharing program with publishers this week. This comes after months of criticism from news organizations claiming Perplexity was reproducing their content without permission or compensation.
The program pays publishers based on how often Perplexity’s AI cites and links to their content in search results. Initial partners include Fortune, Time, and several smaller digital publishers. Perplexity hasn’t disclosed the exact payment formula, but The Verge reported that early participants are receiving payments in the “low five figures” monthly range.
What This Means for AI Search Economics
Perplexity is essentially betting that sharing revenue with content creators will give it a sustainable advantage over competitors who don’t. If users get better, more authoritative answers because premium publishers participate willingly, that could differentiate it from generic AI chatbots that scrape everything indiscriminately.
The limitation: Perplexity still uses content from publishers who aren’t in the program, and there’s no opt-out mechanism beyond robots.txt (which Perplexity has been accused of ignoring in the past). The program feels more like a peace offering to major publishers than a comprehensive solution to AI’s content attribution problem.
What Actually Matters From This Week’s News
Strip away the hype, and three trends are becoming clear. First, AI capabilities are advancing faster in reasoning and agentic behavior than in raw knowledge or creativity—o3’s performance proves that frontier models are getting genuinely smarter, not just bigger.
Second, the major labs are now competing on safety and responsibility as differentiators. Anthropic’s public safety commitments and OpenAI’s delayed o3 release suggest that moving cautiously is becoming a competitive advantage rather than a liability.
Third, the business models are still broken. Microsoft’s restrictive licensing, Perplexity’s awkward publisher payments, and the lack of clear pricing for breakthrough models like o3 show that nobody has figured out how to monetize these capabilities sustainably yet. That uncertainty might matter more than the technical achievements in determining which of these tools you’ll actually be using a year from now.
Disclaimer: Tool pricing and features change frequently. Always verify current information on official websites. Results vary based on individual use case.
Want More?
Ready to find your next favorite AI tool? Browse our full reviews at UntappedAI AI Tools.
Sources
openai.com • google.com • techcrunch.com • gemini.google.com • anthropic.com • wired.com • microsoft.com • theverge.com
