I publish a free AI tool review every few days on AIToolScout. This is my first weekly roundup — 7 stories that actually matter, summarized so you can catch up in 5 minutes.
Published: July 5, 2026 | Reading time: 5 min | Stories: 7
TL;DR — This Week in AI
- Anthropic released Claude Science, an AI workbench that runs locally and integrates 60+ scientific databases
- Google’s Nano Banana 2 Lite generates 1K images in 4 seconds at $0.034 per thousand — the cheapest AI image gen yet
- Claude Sonnet 5 delivers near-Opus agentic performance at $2/$10 per million tokens
- Kimi K2.7 Code became the first open-weight model available in GitHub Copilot’s model picker
- OpenAI and Broadcom unveiled Jalapeno, a custom inference chip built in 9 months
- Reflection AI signed a $6.3 billion compute lease at SpaceX’s Colossus 2 facility
- Meta’s Brain2Qwerty v2 hit 61% word accuracy in non-invasive brain-to-text decoding
Anthropic Launches Claude Science — AI Officially Enters the Research Lab
It is Claude Science. No question. On June 30, Anthropic released a dedicated AI research workbench that integrates over 60 scientific databases — genomics, proteomics, structural biology, cheminformatics — and includes a reviewer agent that checks citations and computations in real time.
The feature that makes universities sit up: Claude Science runs locally on macOS or Linux. Your research data never leaves the lab. That is the difference between “ChatGPT for science” and “AI that meets IRB requirements.” Anthropic is preparing for an IPO, and this product positions them as the go-to AI provider for regulated industries. Beta access for Pro, Max, Team, and Enterprise plans.
Read the full announcement on Anthropic →
Google Nano Banana 2 Lite — 4-Second Images at $0.034 per Thousand
Google dropped a number on June 30 that changes the math for content creators. Nano Banana 2 Lite generates a 1K-resolution image in 4 seconds at $0.034 per thousand images. Three and a half cents. For a thousand images.
I tested it myself through Google AI Studio. The quality is not Midjourney-tier — but at this price, you do not need Midjourney for blog headers, social graphics, and thumbnails. Google also launched Gemini Omni Flash for conversational video editing at $0.10 per second of video. Both tools include SynthID watermarking and C2PA credentials by default.
Read the full announcement on Google AI Blog →
Claude Sonnet 5 Debuts — Near-Opus Agentic Performance at Mid-Tier Pricing
Claude Sonnet 5. Anthropic released it alongside Claude Science, and it is a substantial leap over Sonnet 4.6. Long-form coding, tool calling, debugging — performance approaching Opus 4.8 at $2/million input tokens and $10/million output tokens (introductory pricing through August 31).
Early adopters in insurance and legal services are running it in production already. The model’s agentic capabilities — sustained planning, self-verification, multi-step task completion — are what developers care about most. If you are building anything that requires an AI to actually finish a task rather than stop halfway, Sonnet 5 is the model to test against.
Read the full announcement on Anthropic →
Kimi K2.7 Code — the First Open-Weight Model in GitHub Copilot
Moonshot AI’s Kimi K2.7 Code made history: it is the first open-weight model available directly in GitHub Copilot’s model picker. Developers can now select it alongside GPT-4 and Claude within VS Code, JetBrains, Xcode, and github.com.
The symbolic weight here is bigger than the technical specs. GitHub is owned by Microsoft, which has an exclusive partnership with OpenAI. The fact that an open-weight model is now a first-class citizen in Copilot signals a real shift in the developer tools landscape. Enterprise admins need to enable it first, and billing is usage-based.
Read the full announcement on GitHub Changelog →
OpenAI and Broadcom Unveil Jalapeno — a Custom Inference Chip in 9 Months
Yes — and they shipped faster than anyone expected. On June 24, OpenAI and Broadcom unveiled Jalapeno, a custom ASIC inference chip that went from design to tape-out in nine months. That is exceptionally fast for silicon. Most chip designs take 18 to 24 months minimum.
Jalapeno is purpose-built for LLM inference, claiming significantly better performance-per-watt than existing GPUs. OpenAI’s own models (including GPT-5.3-Codex-Spark) are already running on engineering samples at production frequency. The goal is clear: escape the Nvidia bottleneck that keeps inference costs high for everyone.
Read the full announcement on OpenAI →
Reflection AI Signs $6.3B Compute Lease — the Largest AI Infrastructure Deal Yet
$6.3 billion. That is what Reflection AI committed to SpaceX for access to Nvidia GB300 chips at the Colossus 2 facility near Memphis, Tennessee — $150 million per month from July 2026 through 2029. It is the largest single compute infrastructure deal in AI history.
Reflection AI was founded in 2024 by two former Google DeepMind researchers. Their bet: train an American-made, open-weight frontier model. The open-weight commitment is notable — governments and enterprises are reassessing dependence on closed AI systems after the U.S. banned Anthropic’s Fable 5 and Mythos 5 models. Either party can exit with 90 days’ notice after the first three months.
Read the full story on The AI Insider →
Meta Brain2Qwerty v2 — Non-Invasive Brain-to-Text Hits 61% Accuracy
Getting closer. Meta’s neuroscience team published Brain2Qwerty v2 in Nature Neuroscience on June 29. The system converts non-invasive MEG brain signals into text with 61% average word accuracy — up from 8% in the previous version. The best participant hit 78%.
This is not a consumer product yet. You still need a helmet-sized MEG scanner. But 61% marks a threshold where the technology crosses from “interesting research paper” into “potentially useful for accessibility applications.” No surgery, no implants — just a scanner recording magnetic fields from your brain while you think about typing.
Read the full paper on Nature Neuroscience →
The Big Picture: 3 Trends Shaping AI Right Now
Three patterns are becoming hard to ignore:
- AI is entering regulated industries for real. Claude Science for research labs. Sonnet 5 in insurance and legal. Jalapeno for efficient inference. This is not the chatbot era anymore — it is the infrastructure era.
- Content creation costs are collapsing. Google’s $0.034 per 1,000 images changes who can afford AI visuals. If you are a content creator, small business, or indie developer, this is your moment.
- Hardware independence is the new competitive moat. OpenAI building chips. Reflection locking $6.3B in compute. The era of “just use Nvidia” is ending.
FAQ: AI Weekly Roundup
How often is the AI Weekly Roundup published?
Every weekend. I publish a new free AI tool review or tutorial every few days on AIToolScout, and this roundup drops on Sunday to wrap up the week’s biggest AI news.
How do you choose which news to include?
I filter for stories that matter to content creators, developers, and anyone building with AI tools. No stock market noise. No executive reshuffles (unless it changes the product landscape). Just launches, breakthroughs, and deals that affect what you can actually do with AI.
Where do you source the news from?
Official company blogs (Anthropic, Google, OpenAI, Meta), developer changelogs (GitHub), academic journals (Nature), and trusted tech media (TechCrunch, The AI Insider). Every story links to its original source so you can read the full details.
Can I suggest a story for next week?
Yes — drop me a message here with a link to the news or announcement. I read every submission and give credit when I include a reader tip.
What other content does AIToolScout publish?
Free AI tool reviews with hands-on testing, comparison tables, and real free-tier limits — no “it might be free” guesses. Check out AI Reviews for tool reviews or Blog for tutorials and weekly roundups like this one.
That is it for this week. Bookmark AIToolScout or subscribe to the Blog — I publish every few days and the next roundup drops next Sunday.
