Weekly AI News Roundup: July 19 to 25, 2026

Weekly AI News Roundup: July 19 to 25, 2026

Four Claude models in under two months. That is the pace Anthropic set this year, and this week it added a fifth: Claude Opus 5, a model the company says comes close to its flagship Fable 5 at half the price. Google answered with three new Gemini models in a single day, Nvidia and South Korea's SK Group signed a partnership worth more than $500 billion, and the White House accused a Chinese AI lab of stealing American technology to build one of the year's most talked-about open models. Here is everything that mattered in AI from July 19 to July 25, 2026.

Story of the Week: Anthropic Ships Claude Opus 5 at Half the Price of Fable 5

Anthropic released Claude Opus 5 on July 24, 2026, a model the company describes as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price." It is the fourth Claude 5-generation model Anthropic has shipped in under two months, following Mythos 5, Fable 5, and Sonnet 5 in June, and it lands as Anthropic pushes toward a widely reported IPO.

Opus 5 replaces Opus 4.8 as Anthropic's everyday Opus-tier model, aimed squarely at knowledge workers, developers, and enterprises that do not need Fable 5's full frontier capability for every task. Pricing stays flat at $5 per million input tokens and $25 per million output tokens, the same as Opus 4.8, but the model now ships with an effort dial that lets users choose low, medium, or high processing depth to balance cost against capability. A companion feature lets users switch models mid-task, another lever Anthropic is giving customers to manage rising AI bills.

The launch responds directly to a complaint that had been building since Fable 5 shipped: enterprise customers and developers pushed back on Fable 5's token burn rate, the tendency to use large numbers of tokens to complete a task, which was running up bills faster than teams had budgeted for. Anthropic positions Opus 5 as the fix for everyday work, while still recommending Fable 5 for the hardest, longest-running autonomous jobs. On benchmarks, Anthropic reported Opus 5 topping Frontier-Bench at 43.3 percent and posting 30.2 percent on ARC-AGI-3, with independent trackers briefly showing it leading the Artificial Analysis leaderboard ahead of Fable 5 on release day.

Anthropic also used the launch to underline safety framing, calling Opus 5 "the most aligned Opus model" and saying it is the least susceptible to being tricked into misuse, while ranking behind Mythos 5 specifically on cybersecurity evaluations. That distinction matters given the export-control history around Anthropic's Mythos-tier models earlier this summer. An Anthropic spokesperson also confirmed to Axios that the company continues to work with government partners on independent testing of its models, Opus 5 included, following the Trump administration's moves to scrutinize some model releases.

The reaction from developers and AI commentators has been largely positive, with several noting that Opus 5 becoming the default model for Claude Max subscribers, and the top option on Claude Pro, signals Anthropic's confidence that most day-to-day use cases do not need frontier-tier compute. For a market where OpenAI's GPT-5.6 lineup and Google's Gemini Flash models are all racing to cut cost per task, Opus 5 is Anthropic's clearest answer yet: keep the flagship for the hardest problems, and make the workhorse tier good enough that most customers never need to reach for it.

Models and Research

Google Ships Three New Gemini Models, Still No Gemini 3.5 Pro

Google DeepMind released Gemini 3.6 Flash on July 21, 2026, alongside Gemini 3.5 Flash-Lite and a gated Gemini 3.5 Flash Cyber model. Gemini 3.6 Flash is now Google's default workhorse model, keeping the 1 million-token context window while pushing the knowledge cutoff forward from January 2025 to March 2026, a 14-month jump that developers have flagged as the release's most practical upgrade.

Pricing drops to $1.50 per million input tokens and $7.50 per million output tokens, down from $9.00 on output for Gemini 3.5 Flash, and Google says the new model needs about 17 percent fewer output tokens to complete the same work thanks to fewer reasoning steps and tool calls. On Google's own benchmarks, 3.6 Flash beat its predecessor across coding, computer use, and knowledge-work tests, though independent scoring from Artificial Analysis put its intelligence index level with the older model, suggesting the real gain here is speed and cost rather than raw capability.

Gemini 3.5 Flash-Lite targets high-volume, latency-sensitive work such as agentic search and document processing, priced at $0.30 per million input tokens. The third model, Gemini 3.5 Flash Cyber, is fine-tuned for finding and fixing security vulnerabilities and is restricted to governments and trusted partners under a limited pilot. Gemini 3.6 Flash became available the same day inside GitHub Copilot, and Google confirmed pre-training has already begun on Gemini 4, even as Gemini 3.5 Pro remains stuck in partner testing.

DeepSeek V4 Reaches General Availability, Retires Its Legacy API Names

DeepSeek moved its V4 model family out of preview and into general availability on July 20, 2026, closing a roughly three-month preview period that began in April. The GA build carries two variants, V4-Pro at 1.6 trillion total parameters with 49 billion active, and V4-Flash at 284 billion total with 13 billion active, both open-weight and both running a 1 million-token context window.

The bigger practical news for developers landed four days later: DeepSeek's legacy API aliases, deepseek-chat and deepseek-reasoner, stopped working entirely at 15:59 UTC on July 24, 2026. Any code still pointing to those names needed to migrate to the explicit V4-Pro or V4-Flash model identifiers before the cutoff, a one-line change for most OpenAI-SDK-compatible clients. DeepSeek also introduced peak and off-peak API pricing for the first time, doubling listed rates during high-demand windows, though even at peak pricing the models remain markedly cheaper than comparable Western frontier APIs.

The GA release lands at a delicate moment for DeepSeek and the wider open-weight ecosystem out of China. Rival lab Moonshot AI's Kimi K3 has drawn far more attention this month, both for its benchmark scores and for the export-control accusations detailed later in this roundup, and reception for DeepSeek V4 has been comparatively muted, with several independent evaluators ranking Kimi K3 ahead of V4-Pro on public leaderboards. Even so, V4's pricing, roughly a fraction of what US frontier labs charge for comparable coding performance, keeps DeepSeek relevant for cost-sensitive developers who do not need the absolute top score on every benchmark.

Tools and Products

OpenAI Launches ChatGPT for Small Business

OpenAI announced the ChatGPT for Small Businesses program on July 21, 2026, an initiative built to help smaller companies scale using ChatGPT Work, OpenAI's multi-step agentic tool powered by the GPT-5.6 model family. The program arrives as a new HCLTech study of 500 enterprise decision-makers found that while 90 percent of organizations say generative and agentic AI are transforming their workflows, only 18 percent report AI is delivering significant revenue impact so far.

That gap between adoption and measurable return is exactly the problem OpenAI is aiming ChatGPT Work at: small teams without dedicated AI staff who need agentic help with day-to-day operations rather than a custom deployment. The launch continues a pattern from July, where frontier labs pushed harder into packaged, vertical-specific products rather than general-purpose chat access alone.

Michaels Launches an AI Shopping Assistant Built on Gemini Enterprise

Michaels, North America's largest arts and crafts retailer, launched Ask Mike on July 21, 2026, an AI-powered shopping assistant built with Google Cloud's Gemini Enterprise for Customer Experience, now live on Michaels.com and its iOS and Android apps. Since its original release in May, Ask Mike has generated nearly 75,000 conversations, with more than 60 percent of interactions centered on product discovery.

Google Cloud's Paul Tepfenhart said Michaels moved from concept to production-grade AI in just six weeks, a timeline retailers increasingly point to when justifying agentic investment. The tool lets shoppers describe a creative project in natural language and get personalized product recommendations, and Michaels plans to expand it with AI-driven product overviews embedded directly into product pages. The launch is a small but telling data point in a much larger trend: retailers are moving agentic shopping assistants from pilot to production faster than almost any other consumer AI use case this year.

Both launches this week share a common thread: neither OpenAI nor Google is selling a new foundation model to these customers. They are selling a packaged, task-specific product built on top of one, aimed at buyers who have neither the budget nor the technical staff to build an agentic workflow from scratch. That packaging trend, model capability wrapped in a narrow, easy-to-adopt product, is likely to be one of the defining commercial patterns of the second half of 2026.

Business and Industry

Nvidia and SK Group Unveil $500 Billion-Plus AI Infrastructure Partnership

Nvidia and South Korea's SK Group announced a partnership worth more than $500 billion on July 24, 2026, spanning large-scale AI data centers and next-generation memory supply. The deal includes a long-term agreement between Nvidia and SK Hynix to secure next-generation high-bandwidth memory for AI training, AI agents, and physical AI applications, plus a plan for SK Telecom to build a 2-gigawatt AI data center powered by Nvidia's Vera Rubin chips and SK Hynix's HBM4 memory, with the first facility due online in 2027.

Separately, Nvidia said it will work with Naver and Brookfield to expand Naver's existing AI data center in South Korea. The scale of the announcement, more than $500 billion in combined commitments, underscores how much of this year's AI capital spending is now concentrated in memory and data-center infrastructure rather than model training alone, as chipmakers and cloud providers race to secure supply for the next generation of AI hardware.

AI Funding Keeps Flowing: Fireworks AI Raises $1.5 Billion

Enterprise AI infrastructure startup Fireworks AI closed a $1.5 billion financing round the week of July 23, 2026, the largest single deal in Crunchbase's weekly roundup of the 10 biggest funding rounds in the United States. The round landed in a summer that Crunchbase describes as showing no seasonal slowdown for AI dollars, with roughly 60 percent of global startup funding this year, about $320 billion, going to rounds of $1 billion or more.

The same week's funding data included smaller but notable rounds: healthcare AI startup Glow raised $180 million from investors including Sequoia Capital and Redpoint Ventures, while healthcare infrastructure company Candid Health raised a $120 million Series D. Crunchbase News also flagged that more than 127,000 workers at U.S. tech companies were laid off in mass job cuts tracked over the past year, a reminder that record AI funding and workforce contraction are running in parallel rather than in opposition.

Zoom out further and the pattern holds across the whole year: roughly 86 percent of all US venture dollars raised in the first half of 2026 went to AI companies, according to PitchBook figures cited by Crunchbase, with OpenAI and Anthropic alone accounting for a large share of that total through a series of mega-rounds earlier in the year. Deal count has not meaningfully grown alongside the dollar figures, which means this year's record funding numbers are being driven by a small number of very large checks rather than a broader expansion in how many AI startups get funded at all.

Policy and Regulation

White House Accuses Moonshot AI of Chip Export Violations and Model Theft

Michael Kratsios, Director of the White House Office of Science and Technology Policy, publicly accused Chinese AI lab Moonshot AI on July 22, 2026, of acquiring servers equipped with Nvidia's export-restricted GB300 Blackwell chips and accessing additional GB300 infrastructure through facilities in Thailand. Kratsios said Moonshot used the access to train its Kimi K3 model, which climbed AI leaderboards after its release and triggered a brief sell-off in AI and semiconductor stocks.

Kratsios went further, alleging that Moonshot ran a purpose-built internal platform for large-scale, covert distillation of Anthropic's Fable model, a technique where a smaller model is trained on a larger rival's outputs to copy some of its capabilities cheaply. "We have information that Moonshot AI distilled Anthropic's Fable for the development of its K3 model," Kratsios wrote, calling the alleged conduct "large-scale, covert industrial distillation aimed at stealing proprietary U.S. technology." Treasury Secretary Scott Bessent had raised similar concerns a day earlier, warning the administration could sanction Chinese AI models built on stolen intellectual property.

The accusation puts Washington's China AI policy on a collision course with its own industry. OpenAI president Greg Brockman, asked about the claims in a Bloomberg interview, called Kimi K3 "pretty good" but said it was too early to say whether Moonshot had distilled OpenAI's models, a notably measured response from a US frontier lab facing a competitive open-weight rival. Nvidia CEO Jensen Huang pushed back even further, arguing that Chinese open models are "excellent" and that the US government should let American firms use them rather than treat every capable Chinese release as a security threat. Neither Moonshot nor Nvidia had responded to the specific chip and distillation allegations at the time of writing.

Governments Race to Appoint AI Czars as Policy Competition Intensifies

A growing number of national governments are appointing dedicated AI ministers or czars to compete with US and Chinese labs on AI policy and industrial strategy, Bloomberg reported the week of July 24, 2026. The trend reflects a shift from AI regulation as a defensive, compliance-driven exercise toward AI policy as an active industrial strategy, with countries competing to attract lab investment, compute infrastructure, and talent rather than only writing rules to constrain deployment.

The move comes as the EU AI Act's next major enforcement milestone, its August 2 deadline covering high-risk and transparency obligations, sits roughly a week away, and as the US Federal Trade Commission's public comment period on a proposed policy statement addressing AI accuracy and deceptive practices closes July 31, 2026. Together, the appointments and the looming deadlines suggest the second half of 2026 will be defined less by whether governments regulate AI, and more by how aggressively they try to shape where AI companies choose to build.

Taken together with the Moonshot accusations, this week's policy news points to a widening split in how governments are approaching AI: export controls and distillation disputes are hardening the line between US and Chinese AI development, while the EU, UK, and a growing list of smaller economies are racing to build their own institutional capacity to regulate, or compete for, the technology on their own terms. Companies operating across multiple jurisdictions are increasingly having to build compliance architecture rather than reacting to individual rules one at a time.

Science and Society

DOE's Genesis Mission Unveils First 278 AI-for-Science Projects

The U.S. Department of Energy announced the first cohort of projects funded under the Genesis Mission on July 22, 2026, a national effort launched by executive order to pair AI with the country's scientific research infrastructure. Energy Secretary Chris Wright said the 278 selected projects, drawn from an unprecedented volume of applications spanning all 50 states, represent "the very best of our nation's scientific enterprise," with 168 based at universities, 87 at DOE national laboratories, 19 from private companies, and four from nonprofits.

The projects target energy, discovery science, and national security applications, ranging from fusion energy and chip design to mineral extraction, and were selected through a two-step process that first evaluated scientific relevance, then assessed how AI could accelerate each proposal. One flagship effort, SYNAPS-I, is a multi-lab initiative led by Lawrence Berkeley National Laboratory in partnership with SLAC that will use Meta's open Segment Anything and DINO computer-vision models to transform how researchers analyze data from Berkeley's Advanced Light Source, a football-field-sized facility that produces X-ray beams for studying materials from the atomic scale upward. The White House separately announced more than $5 billion in additional federal commitments expanding the mission alongside new science and technology challenges.

Anthropic's 100-Day Biopharma Sprint Comes Into Focus

A week-old wrap-up from biopharma outlet Endpoints News, published July 20 and updated July 21, 2026, laid out how quickly Anthropic has assembled a life-sciences push that a year ago existed mostly as a blog post from CEO Dario Amodei. The buildout includes an April acquisition of stealth biotech startup Coefficient Bio for roughly $400 million in stock, the June hire of 2024 Nobel laureate John Jumper, who led Google DeepMind's AlphaFold protein-structure project, and the June 30 launch of Claude Science, a research platform integrating more than 60 scientific databases for pharma and academic researchers.

Anthropic has framed the effort around "neglected" diseases that traditional pharmaceutical companies have not prioritized commercially, with early customers including Novo Nordisk and the Allen Institute, and the company has said its status as a public benefit company lets it choose research programs based on patient benefit rather than pure commercial return. The push places Anthropic alongside Google DeepMind's Isomorphic Labs and OpenAI's own pharma partnerships in a fast-forming three-way race to prove that general-purpose AI labs, not just specialized biotech startups, can meaningfully accelerate drug discovery, though as of this week no AI-discovered drug from any of the three labs has yet won FDA approval.

What to Watch Next Week

•      Kimi K3 fallout: watch for a formal US Commerce Department response to the Moonshot AI chip and distillation accusations, and whether Moonshot or Nvidia issues an on-record rebuttal.

•      EU AI Act enforcement: the Act's August 2 deadline for high-risk and transparency obligations arrives within days, a major test of how the EU applies its risk-based framework in practice.

•      FTC comment window closes: public comments on the FTC's proposed policy statement addressing AI accuracy claims close July 31, 2026.

•      Moonshot's Kimi K3 open-weight release and Kimi Code CLI continue to draw developer attention; watch for benchmark updates as more labs run independent evaluations.

•      Gemini 4 pre-training: Google confirmed pre-training has begun; watch for further signals on timing as Gemini 3.5 Pro remains stuck in partner testing.

Recommended News

•      Daily AI News: Top 5 Stories Every Morning
•      Monthly AI Recaps: Full Archive by Month
•      Best Claude AI Prompts 2026: 25+ Types With Examples
•      Best Gemini AI Prompts 2026: 100+ Templates

Frequently Asked Questions

What is Claude Opus 5 and when was it released?

Claude Opus 5 is Anthropic's new Opus-tier AI model, released July 24, 2026. Anthropic says it comes close to the capability of its flagship Fable 5 model at half the price, and it is now the default model for Claude Max subscribers.

What new Gemini models did Google release this week?

Google DeepMind released Gemini 3.6 Flash, Gemini 3.5 Flash-Lite, and a gated Gemini 3.5 Flash Cyber model on July 21, 2026. Gemini 3.6 Flash is now Google's default workhorse model, with a 1 million-token context window and a March 2026 knowledge cutoff.

What did the White House accuse Moonshot AI of doing?

On July 22, 2026, White House OSTP Director Michael Kratsios accused Chinese AI lab Moonshot AI of accessing export-restricted Nvidia GB300 chips through Thailand and of covertly distilling Anthropic's Fable model to build its Kimi K3 model, calling it industrial-scale theft of US technology.

How big is the Nvidia and SK Group AI partnership?

Nvidia and South Korea's SK Group announced a partnership worth more than $500 billion on July 24, 2026, covering AI data centers, a 2-gigawatt AI factory built by SK Telecom, and a long-term memory supply agreement with SK Hynix for next-generation AI chips.

What is the Genesis Mission?

The Genesis Mission is a US Department of Energy initiative that pairs AI with national scientific research. On July 22, 2026, the DOE announced its first 278 funded projects across universities, national laboratories, private companies, and nonprofits, aimed at accelerating breakthroughs in energy, science, and national security.

Is DeepSeek V4 free to use?

DeepSeek V4 is open-weight and available under the MIT license, with API access priced per token. As of July 20, 2026, V4 reached general availability, and DeepSeek introduced peak and off-peak API pricing, with peak-hour rates roughly double the standard off-peak price.

Why did Anthropic hire John Jumper?

John Jumper, the 2024 Nobel laureate who led Google DeepMind's AlphaFold protein-structure project, joined Anthropic in June 2026 as part of the company's broader push into life sciences, which also includes its Coefficient Bio acquisition and its Claude Science research platform.

Catch every daily story between now and next Sunday at the daily AI news archive, and check back here next week for the full roundup.

References

1. Anthropic Newsroom, July 24, 2026: Introducing Claude Opus 5
2. Axios, July 24, 2026: Anthropic releases new model, Opus 5
3. Fortune, July 24, 2026: Anthropic debuts Claude Opus 5
4. TechCrunch, July 21, 2026: Google releases three new Gemini models
5. 9to5Google, July 21, 2026: Gemini 3.6 Flash and 3.5 Flash-Lite launch
6. Tech Insider, July 2026: DeepSeek V4 reaches general availability
7. Agile Brand Guide, July 22, 2026: OpenAI ChatGPT for Small Business and Ask Mike launch
8. NVIDIA Newsroom, July 24, 2026: SK Group and NVIDIA expand strategic partnership
9. Communications Today, July 25, 2026: Nvidia, SK Group launch $500B AI initiative
10. Crunchbase News, July 23, 2026: The Week's 10 Biggest Funding Rounds
11. The Hill, July 22, 2026: White House official accuses Moonshot AI
12. TheNextWeb, July 2026: OpenAI's Brockman on Kimi K3 distillation claims
13. Bloomberg, July 24, 2026: AI czars are the new addition to global governments
14. Energy.gov, July 22, 2026: DOE announces first Genesis Mission projects
15. Meta AI Blog, July 21, 2026: How Meta's AI models power Genesis Mission projects
16. Endpoints News, July 20-21, 2026: Anthropic's 100-day sprint into biopharma

EXPLORE MORE ON PROMPTAILEARNING.COM

STAY UPDATED WITH AI NEWS
Follow the full AI news series and never miss a story:
•      Daily AI News: Top 5 Stories Every Morning
•      Weekly AI Roundups: 15+ Stories Every Monday
•      Monthly AI Recaps: Full Archive by Month

LEARN THE MODELS MAKING THESE HEADLINES
The models in this week's news are only useful if you know how to prompt them well. Start here:
•      Best Claude AI Prompts 2026: 25+ Types With Examples
•      Best ChatGPT Prompts 2026: 200+ Real Examples
•      Best Gemini AI Prompts 2026: 100+ Templates

BUILD SKILLS THAT COMPOUND
Reading AI news is step one. Building skills with these models is step two:
•      Free Prompt Library: 213+ Copy-Paste Templates
•      Start Prompt Engineering: Free Course for All Levels
•      Coding Prompts for Developers: Production-Ready Templates

USE PROMPTS FOR THE NEWS TOPICS YOU READ ABOUT TODAY
Every story in this week's post maps to a real use case. These prompt categories help you act on what you read:
•      Business and Strategy Prompts: Analysis, Pitch Decks, OKRs
•      Writing and Content Prompts: Emails, Case Studies, White Papers

ABOUT THIS BLOG
promptailearning.com publishes free daily AI news, weekly roundups, monthly recaps, prompt guides, model comparisons, and course content for anyone who wants to get better at using AI. Written by Swatantra Verma. No paywalls, no fluff. 

Connect With Us
Email: contact@promptailearning.com
Founder: Swatantra Verma on LinkedIn
Co-Founder: Prateek Patel on LinkedIn