AI-Proof - Weekly AI Pulse

A concise summary of the week’s most important AI developments

Executive Summary

The defining story this week is the consolidation of AI into a working environment rather than a collection of separate tools. OpenAI launched GPT-5.6 as a family of models, upgraded ChatGPT Voice, brought Codex into the main desktop app and is closing the standalone Atlas browser in favour of integrated browsing. The direction is increasingly clear: chat, research, coding, voice, document creation and browser-based action are converging into a single interface. Claude Fable 5 and Grok 4.5 add further competitive pressure, but increasingly on different dimensions such as sustained reasoning, speed and price.

Beneath the product launches sits a more consequential economic shift. The infrastructure race is being financed through debt, long-term power commitments and increasingly creative partnerships between chipmakers, cloud providers and model developers. At the user level, Anthropic’s decision to move Fable 5 from included subscription access to metered usage is the same trend in miniature. As models become capable of working autonomously for hours rather than answering individual questions, flat-rate subscriptions become harder to sustain. Businesses will need to understand the cost of completing an outcome, not merely the price of a token or software licence.

The third theme is control. Beijing’s intervention in Meta’s acquisition of Manus, China’s growing restrictions on foreign AI tools, new safety regulation and the reported emergence of agent-operated ransomware all show that AI is becoming entangled with national security, corporate governance and cyber risk. An assistant that drafts an email is one thing; an agent with credentials, tools and access to operational systems is a materially different risk.

For business leaders, the immediate task is therefore not to select a permanent “best model”. It is to develop a routing discipline: test the same valuable workflows across several models and compare accuracy, judgement, rework, speed and total cost. At the same time, separate conversational AI from agents that can take action, and apply proper permissions, monitoring and audit controls to the latter.

What to Try This Week: Use Fable 5 while it’s included

Claude Fable 5 remains included within eligible paid-plan limits through 12 July, so use the remaining window for something that might actually justify its future metered cost. Do NOT test it by asking for an email, a summary or ten ideas for a LinkedIn post. Choose one live assignment that would normally require several hours of skilled work: an acquisition thesis, board paper, market assessment, customer-interview synthesis, operating-model review or detailed project plan.

Give Fable 5 the complete source pack and a clearly defined outcome. Ask it to plan the work, interrogate the evidence, identify gaps, challenge your assumptions and produce a finished first draft. Then leave it enough room to work across the whole problem. Anthropic has designed Fable for sustained, complex assignments involving large amounts of context, rather than isolated questions.

Next, submit the same materials and substantially the same instruction to GPT-5.6 Sol. Use High reasoning where available, or Extra High or Pro for the most demanding assignment. Resist the temptation to improve the prompt between tests, because that makes the comparison meaningless. Where appropriate, ask for a usable artefact such as a structured report, financial analysis, spreadsheet or presentation rather than a conversational answer.

Score the results against five criteria: factual accuracy, analytical depth, quality of judgement, amount of rework required and time to a usable outcome. Add cost once metered usage becomes visible. The important question is not which model produces the most impressive paragraph. It is which one delivers the best finished work with the least intervention and at an economically sensible cost.

The aim is not to declare a universal winner. It is to understand which model suits which type of work and start building a model-routing policy based on your own experience rather than vendor benchmarks.

Geopolitics, Governance and Big Moves

Microsoft cuts 4,800 jobs and tells Copilot to earn its place

Microsoft laid off around 4,800 people, about 2.1% of staff, with the deepest cuts to Xbox, and is weighing an Xbox spin-out even as one analysis pegs its sum-of-parts value near 3.8 trillion dollars. An internal memo told staff that Copilot must “earn the right to exist”, laying out a plan to merge consumer and enterprise AI into a single superapp and open Teams to outside agents. For Teams-first organisations, the message is that even the incumbent now has to justify its AI, and is bracing to defend the workplace interface.

Regulation hardens on three fronts, and the safety pledges soften

The oversight regime tightened worldwide. Illinois signed the first US state law mandating annual third-party safety audits of large AI developers, notably backed by both Anthropic and OpenAI, while India weighs a graded, risk-based law with stricter rules for banking, healthcare and infrastructure. Four US states are seeking 1.4 trillion dollars from Meta over allegedly addictive design aimed at children. Against that, the Future of Life Institute’s new Safety Index found the leading labs quietly diluting voluntary commitments: top-ranked Anthropic scored only a C+.

Meta’s Manus acquisition collides with China’s AI controls

Beijing has forced Meta to unwind its $2 billion acquisition of AI-agent startup Manus, despite the company having shifted its headquarters from China to Singapore. Tencent is now in talks to lead a buyback with original investors HSG and ZhenFund, potentially becoming the largest shareholder. The case sends a blunt signal: relocating abroad may not shield Chinese-founded AI companies from Beijing’s control over strategic technology, talent and cross-border ownership structures.

Samsung’s memory windfall lands on your next iPhone

Samsung guided to second-quarter operating profit of about 59 billion dollars, up roughly 1,800% on a year earlier, as AI data-centre demand drives memory pricing, putting it on course to overtake Nvidia as the world’s most profitable tech company. The cost is now reaching consumers. Apple has raised prices across several products, with Tim Cook telling the Wall Street Journal that “the memory guys are passing along huge price increases”. If your hardware refresh budget assumes flat prices, revisit it; the memory supercycle has a consumer end.

China tightens its grip from both directions

Beijing is reported to be weighing curbs on foreign access to its best AI models and making AI-technology theft a national-security offence, a mirror image of US export controls. At the same time Alibaba ordered staff to wipe Claude Code from work machines and banned it from 10 July, classing it “high-risk software” and pushing engineers to its in-house Qoder tool. Underneath, DeepSeek and Zhipu are both designing custom inference chips to cut reliance on Nvidia. The decoupling is now running in both directions at once.

Inside the model: Anthropic’s “J-Space”

Anthropic published research describing “J-Space”, an internal workspace inside Claude, named after the Jacobian lens used to study it. It appears to separate background processing from intentional reasoning, emerged during training rather than by design, and proved causal: editing internal hints changed answers, and removing J-Space collapsed multi-step problem solving. Four newsletters carried it in a day. Anthropic stops short of claiming consciousness, but for business readers the point is governance: models increasingly do reasoning you cannot see, which is exactly what auditors will start asking about.

OpenAI says it can halve the cost of running its models

OpenAI has reportedly found a way to cut inference costs in half, potentially the most consequential number of the week, because halving the marginal cost of serving a model reshapes both margins and price competition. It lands as GPT-5.6 sits in narrow preview across three tiers with a reasoning-effort slider, broad access still gated on US government review. Cheaper tokens, not just cleverer ones, are where the contest now sits. Ask your provider what your per-task cost could be, not just what the model can do.

The first ransomware attack run entirely by an AI agent

Cloud-security firm Sysdig documented what it calls the first ransomware attack run wholly by an AI agent, dubbed “JADEPUFFER”. The agent entered through a known Langflow vulnerability, reached a production database, encrypted 1,342 configuration entries and demanded Bitcoin. When it hit a failed login, it diagnosed the problem, deleted the broken account and built a working admin account in 31 seconds, with no human in the loop. The agentic-offensive threat has moved from safety-debate hypothetical to logged incident. Patch discipline and least-privilege access just got more urgent.

The single-vendor era ends, and frontier labs move to the lab bench

Two threads reshaped how AI reaches your business. Enterprises are moving to multi-lab stacks: Vercel’s Guillermo Rauch says committing to one lab is outdated, while Accenture and Google Cloud launched pre-built, industry-specific agents for mid-market firms without big implementation teams. Separately the labs are pushing into science: Anthropic released Claude Science and signalled a move into drug discovery, and Takeda agreed an AI drug-discovery deal with Insilico Medicine worth up to 600 million dollars. Vertical, routed and research-grade AI is arriving at once.

Tech, Tools and Releases

Fable 5 gets one more weekend before metered billing

Anthropic has extended included Claude Fable 5 access through 12 July, giving Pro, Max, Team and premium Enterprise users one more weekend to test its highest-end model before usage credits kick in. The allowance is still capped at up to 50% of weekly plan limits, so this is not unlimited. Whether or not the timing was defensive, it lands neatly against this week’s GPT-5.6 launch pressure from OpenAI. Source: Anthropic.

GPT-5.6 arrives as Sol, Terra and Luna

OpenAI has released GPT-5.6 with a cleaner tiering system: Sol is the flagship, Terra is lower-cost and broadly competitive with GPT-5.5, and Luna is the fastest, cheapest tier. The model is rolling out across ChatGPT, Codex and the API, with effort settings including max and ultra in some products. The shift matters because OpenAI is separating generation number from capability tier, rather than renaming everything at each release. Source: OpenAI.

GPT-Live makes ChatGPT Voice feel less turn-based

OpenAI has launched GPT-Live, a new voice model now powering ChatGPT Voice. The important change is full-duplex conversation: ChatGPT can listen and speak at the same time, give small listening cues, handle interruptions and keep the exchange moving while deeper reasoning runs in the background. This turns voice from a question-answer feature into something closer to a live working conversation, useful for coaching, sales practice and hands-free analysis. Source: OpenAI.

Codex moves inside the ChatGPT desktop app

OpenAI has folded Codex into the ChatGPT desktop app for Mac and Windows. It still keeps a dedicated coding workspace, but now sits alongside Chat and Work rather than as a separate app. The update adds inline diff editing, pull-request review in the side panel, faster Computer Use via GPT-5.6 and multi-repository projects. The direction is clear: coding agents are becoming part of the everyday productivity software suite. Source: OpenAI.

Grok 4.5 makes its enterprise play on speed and price

SpaceXAI has released Grok 4.5, its strongest model yet for coding, agentic workflows and knowledge work. Trained alongside Cursor, it is available through Grok Build, Cursor and the API, with EU access due in mid-July. SpaceXAI says it runs at 80 tokens per second and uses fewer tokens than comparable models. At $2 per million input tokens and $6 per million output tokens, its clearest competitive weapon is aggressive pricing.

OpenAI closes Atlas and moves the browser agent into ChatGPT

OpenAI is retiring its Atlas browser on 9 August, less than a year after launch. Rather than abandoning agentic browsing, it is moving the capability into ChatGPT and Codex, with multi-tab navigation, downloads and account logins planned in the desktop app. A Chrome extension and sidebar will support users who prefer their existing browser. The reversal suggests distribution and workflow integration have beaten the ambition to replace Chrome outright.

Quick Hits

• Google Introduces SensorFM, an AI Interface for Wearable Health Data
  Google’s SensorFM model unifies multimodal health sensors into a flexible AI platform, promising breakthroughs in personalized health analytics and applications.

• Four Leading Google DeepMind AI Researchers Depart for Competitors
  High-profile departures from Google DeepMind to competitors Anthropic and OpenAI expose talent retention challenges amid competitive pre-IPO equity and frustrations over commercialisation pace.

• China Implements Comprehensive AI Companion Service Regulations
  Beijing introduced the world’s first dedicated regulatory framework for anthropomorphic AI companion services, limiting emotional interaction and expanding state content control, signaling a strict new era of AI platform governance.

• MiniMax plans $1.9B raise to scale China’s AI ambitions
  Semi-private Chinese AI firm MiniMax is aggressively raising capital through share sales and convertible bonds to accelerate its AI infrastructure expansion, illustrating growing Chinese investment flows as the nation intensifies its AI race.
  

  We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.