Listen

Description

🔥 This Week’s Highlights

It was a heavyweight week. Google fired the biggest shot with Gemini 2.5 Pro Deep Think, which swept the science, math, and reasoning leaderboards on June 22, then promptly undercut its own moment by delaying Gemini 3.5 Pro to July. OpenAI answered with a whole new model family — GPT-5.6 Sol, Terra, and Luna — though access is locked to ‘trusted partners’ for now at the US government’s request. And the talent war went fully public: Anthropic poached three senior Google DeepMind researchers (including AlphaFold’s John Jumper) while accusing Alibaba of a 28.8-million-exchange scheme to copy Claude. On the hardware side, Meta made AI glasses genuinely affordable at $299.

🤖 AI Models & Features

GPT-5.6 Sol, Terra & Luna: OpenAI’s New Model Family

🌞 News & Announcements · High Priority · 15 min to try

OpenAI split its next generation into three models tuned for different jobs: Sol for the hardest coding and research, Terraas the balanced everyday workhorse (roughly GPT-5.5 quality at half the cost), and Luna as the cheap, fast option for drafting and routine automation. There are real new capabilities too — a max reasoning-effort dial and an ‘ultra’ multi-agent mode that spins up subagents. The catch: at launch it’s limited to a small group of partners via API and Codex, with ChatGPT access to follow. Keep an eye on your model picker over the next few weeks.

Read more →

Gemini 2.5 Pro with Deep Think Tops the Leaderboards

🧠 News & Announcements · High Priority · 15 min to try

This is the model to beat right now. Deep Think adds a separate ‘extended, parallel reasoning’ mode that kicks in for hard problems, and it paid off: 89.8% on MMLU-Pro (highest of any public model) and 82.4% on GPQA Diamond, ahead of Fable 5 and GPT-5.5. The 2-million-token context window is the other headline — you can drop in an entire codebase or a couple of full-length books and ask questions across all of it. If you use Gemini, switch to Deep Think for your genuinely thorny questions.

Read more →

Claude Is Now a Model Option in Microsoft 365 Copilot

🤝 Products & Platforms · High Priority · 5 min to try

If you spend your day in Word, Excel, and Teams, you can now pick Claude inside Copilot Chat for complex analysis, long-document understanding, and structured writing. It’s a quiet but meaningful change: you no longer have to leave Office (or copy-paste into a separate chatbot) to get Claude’s take on a messy spreadsheet or a 40-page contract. Worth checking your Copilot model selector to see if it’s live on your plan.

Read more →

Gemini 3.5 Live Translate: Speech-to-Speech in 70+ Languages

🗣️ Products & Platforms · High Priority · 5 min to try

Google’s new Live Translate does near real-time voice translation across 70+ languages — including during phone calls, at the system level. That’s broader language coverage than anything comparable from OpenAI or Anthropic at launch. The one limit: it’s starting on Pixel 10 and Galaxy S26 with Android 17, and will reach more chipsets later. If you travel or have family across a language barrier, this is the kind of feature that genuinely changes a phone call.

Read more →

Google Finance Gets a Dedicated App with AI ‘Key Moments’

📈 Products & Platforms · Medium Priority · 5 min to try

Google Finance is now a standalone app (Android first), bundling watchlists, real-time data, and live news with an AI ‘Key Moments’ feature that explains why a stock is moving in plain language. For casual investors, that context is the useful part — less staring at a red number, more understanding what just happened. iOS is coming in the next few months. (Informational, not financial advice.)

Read more →

DeepSeek V4-Pro Makes Its 75% Price Cut Permanent

💸 News & Announcements · Medium Priority · 15 min to try

DeepSeek turned its promo discount into the new normal: V4-Pro now costs about $0.44 per million input tokens and $0.87 output — more than 5x cheaper than GPT-5.5 on input and 17x cheaper on output. If you build or automate anything on top of an API, this resets what ‘cheap’ means. The open-weight Chinese models are now in a different price universe than Western frontier models, which is worth factoring into your defaults.

Read more →

VibeThinker-3B: Frontier-Level Reasoning at Just 3B Parameters

🐤 Concepts & Research · Medium Priority · 30 min to try

A 3-billion-parameter model that claims to match frontier reasoners on math and code? That’s the pitch for WeiboAI’s MIT-licensed VibeThinker-3B. It’s small enough to run on a decent laptop, which makes it a fun, low-stakes way to see how far tiny models have come. Don’t expect a great generalist chatbot — but for math and coding experiments you can run entirely offline, it’s a neat download.

Read more →

Nvidia Nemotron 3 Ultra: A New Open-Weights Heavyweight

⚡ News & Announcements · Medium Priority · 30 min to try

Nvidia’s latest open model lands with praise for its capability-to-efficiency ratio — strong results without demanding the biggest hardware. If you like running or fine-tuning your own models rather than renting them, Nemotron 3 Ultra is worth a spot on your shortlist alongside the latest Qwen and DeepSeek releases.

Read more →

Perplexity Adds Multi-Model ‘Council’ and Learn Step-by-Step

🧩 Products & Platforms · Medium Priority · 15 min to try

Perplexity’s big update leans into a smart idea: its new Council feature runs your question past several models and blends the answers, so you’re not betting everything on one model’s quirks. There’s also a Learn step by step mode for studying a topic incrementally, plus better memory and a cleaner UI. If you do research-heavy work, Council is the standout to try first (some features are Pro-only).

Read more →

Grok Composer 2.5 Lands in Grok Build

🛠️ Coding Tools · Medium Priority · 30 min to try

xAI’s new Composer 2.5 is a fast model tuned for long-running, multi-step tasks, now available in Grok Build for SuperGrok and X Premium+ subscribers. It’s part of Grok’s broader June glow-up (1M context, image-to-video, and more). If you already tinker with agentic coding inside Grok, make Composer 2.5 your new default and see how it handles longer jobs.

Read more →

FrontierCode: A Benchmark for Production-Quality AI Code

📊 Concepts & Research · Medium Priority · 15 min to try

Most coding benchmarks ask ‘does the code run?’ Cognition’s new FrontierCode asks the better question: ‘would a senior engineer actually merge this?’ It scores agents on scope control, regression safety, and test quality across 150 tasks. Early results put Fable 5 on top at 46.3%, ahead of Claude Opus 4.8 (34.3%) and GPT-5.5 (25.5%). A more honest signal than SWE-bench if you’re deciding which model to trust with hands-off coding.

Read more →

Anthropic Accuses Alibaba of a 28.8M-Exchange Distillation Attack

🛡️ News & Announcements · Low Priority · 5 min to try

Anthropic told the US Senate that operators tied to Alibaba’s Qwen lab used ~25,000 fake accounts and 28.8 million exchanges to ‘distill’ Claude — copying its behavior without its safety guardrails. It’s the largest such campaign Anthropic has disclosed. Mostly an industry-and-policy story, but it explains why every lab is tightening high-volume API monitoring, which can affect how the rest of us use these tools.

Read more →

Google’s AI Talent Exodus: Four DeepMind Departures in Six Days

🏃 News & Announcements · Low Priority · 5 min to try

In six days, Google DeepMind lost four senior people: Noam Shazeer to OpenAI, and John Jumper (AlphaFold), Jonas Adler, and Alexander Pritzel to Anthropic. Reporting points to compute-allocation frustrations and pre-IPO equity as the pull. Nothing to ‘try’ here, but it’s a telling read on where momentum is concentrating — and useful background for why Gemini’s roadmap wobbled this week.

Read more →

The End of Cheap AI: Frontier Model Prices Are Reversing

📉 Concepts & Research · Medium Priority · 5 min to try

The years-long slide in AI prices flipped in 2026: GPT-5.5, Gemini 3.5 Flash, and Claude Fable 5 all launched…



This is a public episode. If you would like to discuss this with other subscribers or get access to bonus episodes, visit nicholascalcutti.substack.com