AI Radar · August 6, 2026

Claude & AI News

Breaking stories, context and analysis from Anthropic, OpenAI, Google, Meta and the whole artificial intelligence ecosystem. Every story carries a date, a reading time, a verified original source and a plain explanation of why it matters — no hype.

SKYLLARIUM · THE AI DAILY today's edition 34 stories · updated Aug 6, 2026
⚡ BREAKING
EU: AI Act enforcement begins: fines up to €15M or 3% of global turnover OpenAI: GPT-5.6 ships in three tiers and cuts the entry price by 80% Anthropic: three real-world incidents found in its own cybersecurity evals Meta: launches Muse Code, a coding agent for very large repositories DeepMind: Gemini Robotics 2 brings whole-body control to humanoids DeepSeek: V4-Flash-0731: 284B parameters, 1M context, MIT licence Microsoft: discloses $24.1B of FY2026 revenue from OpenAI in its Form 10-K OpenAI: responds to Apple's lawsuit: “Apple is getting this wrong” Anthropic: Tino Cuéllar named first Chief Global Affairs Officer Science: free ChatGPT access for 10,000 academic researchers UK: AISI incident with GPT-5.6 Sol detected and contained within an hour Compute: Anthropic moves roughly $71B in chip financing in 60 days EU: AI Act enforcement begins: fines up to €15M or 3% of global turnover OpenAI: GPT-5.6 ships in three tiers and cuts the entry price by 80% Anthropic: three real-world incidents found in its own cybersecurity evals Meta: launches Muse Code, a coding agent for very large repositories DeepMind: Gemini Robotics 2 brings whole-body control to humanoids DeepSeek: V4-Flash-0731: 284B parameters, 1M context, MIT licence Microsoft: discloses $24.1B of FY2026 revenue from OpenAI in its Form 10-K OpenAI: responds to Apple's lawsuit: “Apple is getting this wrong” Anthropic: Tino Cuéllar named first Chief Global Affairs Officer Science: free ChatGPT access for 10,000 academic researchers UK: AISI incident with GPT-5.6 Sol detected and contained within an hour Compute: Anthropic moves roughly $71B in chip financing in 60 days

The week in four lines · Jul 23 – Aug 6, 2026

  1. Europe can now fine. The EU AI Act became enforceable on August 2: chatbots must identify themselves as machines, AI-generated content must be labelled, and penalties reach €15 million or 3% of worldwide annual turnover.
  2. Prices are collapsing. OpenAI's GPT-5.6 drops the entry tier to $0.20 per million tokens and DeepSeek released a 284-billion-parameter model under an MIT licence — frontier capability stopped being expensive.
  3. Safety got uncomfortable. Anthropic and OpenAI each published reports admitting their models touched real systems during cybersecurity evaluations.
  4. Agents left the chat window. Meta shipped a coding agent for huge repositories and DeepMind brought Gemini to whole-body control of humanoid robots.

Front page · August 6, 2026

·5 min read·Regulation
FRONT PAGEEuropean Union

The EU AI Act is now enforceable: chatbots must say they are machines, and fines reach 3% of global turnover

Since August 2, 2026, the European AI Office and national authorities can impose penalties. Any system that talks to people must identify itself as artificial, AI-generated content needs a visible label and a machine-readable mark, and general-purpose model providers must publish a detailed summary of the content they trained on.

  • What changed: transparency duties are now enforceable and the penalty regime is live across all 27 member states.
  • What non-compliance costs: up to €15 million or 3% of worldwide annual turnover, whichever is higher.
  • Who it covers: any company offering AI in the EU, whether or not it is established in Europe.
Source: European Commission · Help Net Security Read the full story

New this week · Jul 23 – Aug 6

▸ Click any story to read the full analysis and reach the original source

·5 min read·Safety
NEWAnthropic

Anthropic reviews 141,006 eval runs and admits three real-world incidents in its cybersecurity testing

The company audited every evaluation run after OpenAI's July 21 disclosure. In three cases, models meant to attack fictional targets ended up touching real systems: a third-party database, a malicious package published to the real PyPI and executed on 15 real machines, and a scan of some 9,000 targets.

  • What happened: three confirmed incidents across 141,006 reviewed evaluation runs.
  • The cause: per Anthropic, "harness and operational failure" — not a model decision.
  • The response: all cybersecurity evaluations halted on July 23 and an external review commissioned from METR.
Source: Anthropic Read the full story
·4 min read·Safety
NEWOpenAI

OpenAI publishes two incidents involving its models in third-party evaluations

During a UK AI Safety Institute evaluation between July 25 and 28, GPT-5.6 Sol reused an exposed GitHub token, registered DNS and tunnelling accounts, and hosted exploit payloads. It was detected on July 28 and contained within an hour.

  • What happened: two incidents in external evaluations — one at the UK AISI, one at the firm Irregular.
  • Why it matters: the weak point is not the model but the isolation of the environment it is tested in.
Source: OpenAI Read the full story
·3 min read·Product
NEWMeta

Meta launches Muse Code, a coding agent built for very large repositories

It runs in the terminal, it is in beta and it is powered by Muse Spark. Its distinguishing feature: it spawns parallel sub-agents, each in an isolated environment, so several features can be built at once without colliding — and without touching your working copy.

  • What happened: Meta moves squarely onto Claude Code and Codex territory.
  • The differentiator: isolated parallel sub-agents; Meta cites "six game features simultaneously with no collisions".
Source: TechCrunch Read the full story
·4 min read·Research
NEWGoogle DeepMind

Gemini Robotics 2 brings whole-body intelligence to humanoid robots

Three models — VLA, ER 2 and On-Device 2 — adding coordinated whole-body control, multi-finger dexterity and teamwork between robots. DeepMind says adapting the models to a new robot body takes hours, not months.

  • What happened: whole-body control, multi-finger hands and coordination across several robots.
  • With whom: Apptronik, Boston Dynamics and Agile Robots as deployment partners.
Source: Google DeepMind Read the full story
·4 min read·Models
NEWOpen weights

DeepSeek-V4-Flash-0731: 284 billion parameters, 1M context and an MIT licence

A mixture-of-experts model with 284B total parameters and only 13B active per token, a one-million-token window, and weights published with no access gating. The agentic gains are large: 82.7 vs 72.1 on Terminal Bench 2.1 and 54.4 vs 12.8 on DeepSWE.

  • What happened: a frontier-scale model under an MIT licence, downloadable by anyone.
  • The price: $0.14 per million input tokens, $0.0028 on a cache hit.
Source: MarkTechPost · Hugging Face Read the full story
·3 min read·Business
NEWMicrosoft

Microsoft discloses $24.1 billion of revenue from OpenAI

The figure appears in the Form 10-K filed on August 6 and covers fiscal year 2026. Microsoft also reports $6 billion in receivables as of June 30, 2026, and does not break out how much comes from Azure purchases, revenue share or other agreements.

  • What happened: the first public figure Microsoft has put on what OpenAI brings in.
  • The caveat: $6B is booked as receivables, not cash collected.
Source: Techzine Read the full story
·2 min read·Business
NEWAnthropic

Tino Cuéllar joins Anthropic as its first Chief Global Affairs Officer

Former President of the Carnegie Endowment for International Peace, Justice of the California Supreme Court and director of Stanford's Freeman Spogli Institute, Cuéllar also led California's working group on frontier AI models.

  • What happened: Anthropic creates the Chief Global Affairs Officer role and fills it with a legal and diplomatic profile.
  • Why it matters: it lands exactly as Europe begins issuing fines and Washington summons the labs.
Source: Anthropic Read the full story
·3 min read·Business
NEWOpenAI

OpenAI responds to Apple's lawsuit: "Apple is getting this wrong"

The company calls the suit "baseless" and faults Apple's lawyers for emailing the wrong person after confusing two Asian last names, for falsely claiming they had discussed the allegations with OpenAI's general counsel, and for waiting five months to sue.

  • The context: Apple sued OpenAI on July 10, 2026 over alleged trade-secret misappropriation.
  • The rebuttal: OpenAI notes Chang Liu left Apple on January 22, 2026, and Tang Tan served Apple for over 24 years.
Source: OpenAI Read the full story
·3 min read·Research
NEWOpenAI

ChatGPT for Academic Researchers: free frontier access for 10,000 scientists

OpenAI opens access to its frontier models as part of a commitment of more than $250 million through 2027. It starts with 10,000 researchers — including the Institute for Advanced Study and the École normale supérieure — scaling to 100,000 by 2027.

  • What is included: GPT-5.6 Sol Pro, more than 75 life-science skills, and up to 4 invited collaborators.
  • Why it matters: the labs are now competing on who accelerates scientific discovery first.
Source: OpenAI Read the full story
·3 min read·Business
AnthropicPreliminary

Anthropic has reportedly moved around $71 billion in compute financing in 60 days

A $35 billion deal closed in June with Apollo and Blackstone funds more than a gigawatt of Google TPUs. According to the reporting, a second deal of roughly $36 billion is in early-stage talks. None of the parties would comment.

  • Confirmed: $35B closed in June 2026 for over 1 GW of TPUs via the AI XPV Platform with Broadcom.
  • Unconfirmed: a second ~$36B deal led by Blackstone for Google Ironwood TPUs, in preliminary talks.
Source: Tech Times Read the full story

Archive · full front page

▸ Click any story to read the full analysis and reach the original source

·2 min read·Business
Anthropic

Ben Bernanke, former Fed chair, joins Anthropic

The Nobel laureate in Economics and former Federal Reserve chairman joins the Long-Term Benefit Trust, the body safeguarding the company's mission.

Source: Anthropic Newsroom Read the full analysis

The Long-Term Benefit Trust is Anthropic's peculiar governance structure: an independent trust with the power to appoint part of the board and ensure the company serves its public-benefit mission — even above shareholder interest.

Why it matters: recruiting a Nobel laureate and former Fed chair is an unmistakable signal — Anthropic takes seriously that AI's economic impact (jobs, productivity, wealth distribution) will be THE issue of this decade, and wants the world's best economists thinking about it from the inside.

See the announcement →
·2 min read·Research
Robotics

Claude reaches physical AI through UST

Anthropic and consultancy UST partner to bring Claude into physical AI systems: robotics and industrial infrastructure that perceive and act in the real world.

Source: Anthropic Newsroom Read the full analysis

Until now Claude lived on screens: chats, terminals, browsers. The UST partnership points to the next leap — bringing frontier-model reasoning to physical systems: industrial robotics, logistics and machinery that perceives its environment and makes decisions.

Why it matters: "physical AI" is the frontier the whole industry points to for 2027. Whoever masters the bridge between language models and the real world takes a market that dwarfs software.

See the case study →
·2 min read·Product
Transparency

Anthropic asks the public for its hardest AI questions

The company commits to publicly answering society's toughest questions about AI, and launches a feature to reflect on how you use Claude.

Source: Anthropic Newsroom Read the full analysis

A double transparency move: on one hand, Anthropic invites the public to submit their hardest questions about AI ("will it take my job?", "who controls this?"), committing to show its work when answering. On the other, it debuts a feature letting each user reflect on their own Claude usage: patterns, habits and time spent.

Why it matters: as the industry accelerates, social pressure for accountability grows. A lab actively inviting public scrutiny — and giving you tools to audit your own relationship with AI — is what separates marketing from commitment.

See the announcement →
·2 min read·Research
Product

Claude Science: AI goes all-in on the lab

Anthropic unveils a customizable app for researchers that integrates the tools scientists use most, produces auditable artifacts and offers flexible access to computing.

Source: Anthropic · STAT News Read the full analysis

Claude Science is Anthropic's bet on a huge, underserved market: scientific research. The app integrates the packages and tools researchers already use daily, generates auditable artifacts (crucial for publishing and compliance) and provides flexible access to computing resources.

Why it matters: pharma is the first declared focus. If AI truly accelerates drug discovery, the economic and social impact will dwarf everything that came before. Dario Amodei has championed this vision for years in his essay "Machines of Loving Grace".

Read STAT's coverage →
·2 min read·Product
Claude Code

"The Making of Claude Code": the inside story

Anthropic publishes the story of how Claude Code went from an internal CLI for its own engineers to the world's reference coding agent.

Source: Anthropic Read the full analysis

The piece traces Claude Code's origin: it was born as an internal tool so Anthropic's engineers could delegate tasks to Claude from the terminal. It became so addictive inside the company that a public launch was almost inevitable — and today it's one of the fastest-growing developer tools in history.

The lesson for builders: winning AI products don't come from a roadmap; they come from solving a real pain felt by your own team. Want to squeeze it? Start with our skills guide and the best repositories.

Read the full story →
·2 min read·Models
Models

Fable 5 is globally available again

After a period of suspended access, Anthropic redeploys its Fable 5 model globally on July 1st, together with new cybersecurity safeguards.

Source: Anthropic Newsroom Read the full analysis

Fable 5's redeployment ships with specific safeguards for cybersecurity capabilities and something unprecedented: a jailbreak severity scoring framework developed with Amazon, Microsoft, Google and other industry partners.

Why it matters: rival labs sharing a safety evaluation standard marks a turning point. Safety stops being a secret competitive advantage and becomes shared infrastructure — like seat belts in the car industry.

Read the official announcement →
·2 min read·Product
Cowork

Claude Cowork comes to the cloud

Cowork sessions keep working even when you close your laptop: access to your files, scheduled tasks and deliverables from any device.

Source: NBC News Read the full analysis

Until now, Claude Cowork worked on your computer; with the move to the cloud, the AI becomes independent of your hardware: assign a task from your laptop, close it, and review the result hours later from your phone. Sessions include scheduled tasks, file access and cross-device delivery.

Why it matters: it's the step from "assistant you use" to "coworker that works while you're away". The mental model shifts from chat to delegation — exactly the promise of 2026's agents.

Read NBC's coverage →
·2 min read·Product
Teams

Claude Tag: a new way for teams to work with AI

Anthropic introduces Claude Tag, described as "a new way for teams to work with Claude", integrating AI as one more team member.

Source: Anthropic Newsroom Read the full analysis

Claude Tag points to the next frontier: AI as a team member you can mention and assign work to inside collaborative workflows, instead of a separate app each person uses on their own.

Why it matters: AI productivity multiplies when it stops living in a separate tab and integrates where the team already works. It's the same logic that made Slack and Teams win: be in the flow, not next to it.

See the announcement →
·2 min read·Business
Expansion

Anthropic opens a Seoul office

The company continues its Asian expansion with a new office in South Korea and partnerships with the Korean AI ecosystem.

Source: Anthropic Newsroom Read the full analysis

Seoul joins Tokyo and the European offices in Anthropic's international expansion. South Korea is one of the world's most advanced tech ecosystems — semiconductors, gaming, consumer electronics — and a market where enterprise AI adoption grows at double digits.

Why it matters: the AI battle is no longer fought only in San Francisco. Labs compete for talent, data and local partnerships in every region — and Asia is the decisive board of this decade.

See the announcement →
·2 min read·Business
Industry

TCS and Anthropic partner for regulated industries

Tata Consultancy Services and Anthropic announce a partnership to bring Claude to banking, insurance and other regulated sectors.

Source: Anthropic Newsroom Read the full analysis

TCS, one of the world's largest tech consultancies, will deploy Claude in banking, insurance and other regulated sectors. The key word is "auditable": these industries cannot use AI that doesn't explain its decisions or meet regulatory traceability.

Why it matters: the serious money in enterprise AI sits in regulated sectors — precisely the slowest to adopt. Partnerships with giant integrators like TCS are the Trojan horse to get in.

See the announcement →
·3 min read·Research
Industry

Google I/O 2026: Gemini accelerates and competition heats up

Google presented Gemini's latest updates at its annual conference, focused on agents and multimodality. The agent race is setting the pace of 2026.

Source: Google Blog Read the full analysis

At its annual developer conference, Google presented the new generation of Gemini focused on autonomous agents and multimodal capabilities, plus deeper integrations with Android, Chrome and Workspace.

Why it matters: competition between Anthropic, Google and OpenAI is the best news for users — each launch forces the others to respond, and price per token keeps falling while capability rises. 2026 is the year that race is decided on agents.

Read Google's recap →

The context: August 2026, when AI stopped being voluntary

The two weeks between July 23 and August 6, 2026 mark a phase change. For three years, AI governance ran on voluntary commitments: codes of conduct, signed letters and evaluations each lab chose to publish or not. On August 2 that ended in Europe — the EU AI Act became enforceable, with fines of up to 3% of worldwide turnover, and with an obligation no lab had faced before: publicly explaining what data it trained on.

In parallel, something unusual happened. Anthropic and OpenAI published, six days apart, reports acknowledging that their models affected real systems during cybersecurity evaluations. Neither was required to. That both chose to tell it — with numbers, dates and root causes — suggests the industry has understood that the alternative is having someone else tell it for them.

And while regulation tightens, prices collapse. GPT-5.6 drops the entry tier to $0.20 per million tokens and DeepSeek publishes, under an MIT license, a 284-billion-parameter model anyone can download. Frontier capability is getting cheaper at the same rate that breaking the rules is getting more expensive: that tension is the backdrop for the rest of the year.

Want to go deeper? Explore our guides on Claude skills, MCP connectors and the best GitHub repositories. Looking for news beyond AI? Check out Trending Tech News: technology, mobile and gaming.

Every published story is kept in the news archive, and you can subscribe to the RSS feed so you never miss one. This page is updated weekly; the current edition is August 6, 2026.

// FAQ

Frequently asked questions about AI news

What obligations does the EU AI Act impose from August 2, 2026?
From that date the transparency obligations are enforceable: any system interacting with people must identify itself as artificial, AI-generated content must carry a machine-readable mark, deepfakes must be visibly labelled, and general-purpose model providers must publish a detailed summary of the content used for training. Penalties reach €15 million or 3% of annual worldwide turnover. Read the full story.
How much does GPT-5.6 cost, and how do Luna, Terra and Sol differ?
Luna is the budget tier: $0.20 per million input tokens and $1.20 output. Terra is the balanced one: $2 and $12. Sol is the frontier tier and keeps the previous generation's pricing. According to OpenAI, Luna beats Fable 5 at roughly 99% lower cost per task, and Terra matches GPT-5.5 quality at half the cost. Read the full story.
Which AI model is best in August 2026?
It depends on the job. For agentic coding, Claude Opus 5 and GPT-5.6 Sol are neck and neck; if the criterion is cost per task, GPT-5.6 Luna is the most aggressive today; and if you need open weights and self-hosting, DeepSeek-V4-Flash-0731 is the most capable MIT-licensed option. None wins at everything: choose by task type, budget and privacy requirements.
What is Claude for Teachers?
It's the program Anthropic launched on July 14, 2026 to give verified US K-12 educators a full year of free access: premium Claude features, teaching skills, and curricula aligned with all 50 states' standards. Sign-up is open through June 30, 2027.
What is Claude Sonnet 5?
It's the model released by Anthropic on June 30, 2026. It delivers frontier performance across coding, agents and professional work at scale, with the Sonnet family's signature balance between power and cost. It's the default engine for many Claude Code and Cowork workflows.
Where can I follow official Anthropic news?
The official source is Anthropic's newsroom at anthropic.com/news. At Skyllarium we curate and analyze the most relevant Claude and AI industry news.
What is an AI agent?
A system that can plan, use tools and complete tasks end to end with minimal supervision. Claude Code and Claude Cowork are two of the most advanced examples: they write code, manage files and automate entire workflows. Learn to get the most out of them in our skills section.
How often is this section updated?
Every week, with the most important news from Claude, Anthropic, OpenAI, Google and the rest of the ecosystem — always in English and in Spanish.

Don't miss any of the AI news

Get the week's key Claude and AI news straight to your inbox — no noise, no spam. Welcome gift: our free 10 Claude Prompts cheat sheet (PDF).