AI-Proof - Weekly AI Pulse
A concise summary of the week’s most important AI developments
Executive Summary
OpenAI retook the coding crown this week, although only a handful of organisations can actually use it. The company unveiled the GPT-5.6 family, with flagship Sol reclaiming the lead on key coding benchmarks from Anthropic’s Mythos 5. Yet rather than a public launch, access was limited to around twenty vetted partners after the US government requested a staged rollout on national security grounds. Frontier AI releases are increasingly becoming political events as much as technology launches.
Anthropic also returned to the headlines. Following the easing of US export restrictions, Fable 5 returned on 2 July alongside the release of Claude Sonnet 5 and the new Claude Science platform for researchers. At the same time, Austria called for Europe to host its own frontier AI capability, while Google’s decision to ration compute to Meta reinforced the week’s defining theme: access to AI is increasingly constrained by infrastructure rather than algorithms.
The practical takeaway for UK leaders is twofold.
First, if you have access to Claude Pro, Max or Team, spend some time with Fable 5 before 7 July, while it remains included in your subscription. It is a specialist reasoning model designed for complex, high-value work rather than everyday chat, making it an excellent opportunity to learn when premium models genuinely earn their cost.
Second, expect AI infrastructure, compute availability and pricing to remain volatile well into 2027. Building AI capability is no longer just about choosing the best model. It is increasingly about choosing the right model for the right job at the right cost.
What to Try This Week: Use Fable 5 while it’s included
If you have a Claude Pro, Max or Team subscription, this is the week to try Fable 5. Until 7 July, Anthropic is including it within plan limits before it moves to usage-based pricing. After that, it becomes one of the most expensive frontier models to run, so now is the ideal opportunity to understand where it adds value.
The mistake most people make is using Fable 5 as a drop-in replacement for everyday chat. That is like hiring a QC to write meeting notes. Anthropic recommends using it for problems that would normally take a skilled person hours or days: designing an architecture, reviewing an acquisition strategy, planning a complex programme, analysing large document sets, or running multi-step investigations. It is designed for long-running, end-to-end work rather than quick questions.
A good way to test it is to give it a substantial objective, define the outcome you want, provide all the relevant context and then let it produce a complete first draft. Resist the temptation to break the task into lots of smaller prompts. Fable 5 is at its best when it can reason across the whole problem.
Finally, pay attention to where it genuinely earns its cost. From next week it moves to usage credits, so treat it as a specialist model for high-value work, while leaving routine drafting, summarisation and everyday coding to lower-cost models. Understanding that distinction now will help you build a more cost-effective AI workflow.
Geopolitics, Governance and Big Moves
The Anthropic export standoff cracks, and Europe wants its own frontier lab
After two weeks in which US export controls pulled Anthropic’s strongest models offline, the clearest climbdown yet arrived. Mythos 5 was restored to roughly a hundred vetted US companies and government partners, and the consumer-facing Fable 5 was made available again, all under an order requiring federal cybersecurity sign-off. The durable consequence is geopolitical: Austria publicly urged the European Commission to explore hosting Anthropic inside the EU. The policy question of 2026 is shifting from “is this model safe to release” to “can we be cut off from it”.
OpenAI explores U.S. government equity stake
OpenAI is reportedly in early discussions to give the U.S. government a 5% equity stake, potentially through a public or sovereign wealth vehicle. While the proposal remains conceptual and would likely require congressional approval, it signals a new model of partnership between frontier AI companies and governments. If realised, it could reshape how national security interests are balanced with private AI development and funding. Source: The Guardian.
UN warns governments are falling behind AI
A UN scientific assessment has concluded that AI capabilities are advancing faster than scientific understanding and government oversight. The panel found that most countries lack the technical expertise needed to evaluate frontier AI models effectively. The findings will be presented at the UN Global Dialogue on AI Governance in Geneva, highlighting the widening gap between rapid AI innovation and public-sector readiness. Source: Reuters.
Apple and EU move closer on Siri AI rollout
Apple and the European Union are reportedly making progress on plans to launch the rebuilt AI-powered Siri across Europe. Discussions between Tim Cook and EU technology chief Henna Virkkunen focused on ensuring compliance with competition rules while avoiding delays for European users. The talks reflect the growing challenge of balancing AI innovation with regional regulation and digital sovereignty. Source: The Verge.
Capacity has become the currency, and Google is rationing it
Alphabet restricted Meta’s access to Gemini after Meta asked for more capacity than Google could supply, forcing one of the world’s best-resourced firms to fall back on its own model and tell staff to economise on tokens. Set against more than 300 US data-centre bans and research showing AI has restarted US electricity-demand growth after sixteen flat years, the message is blunt. Compute is now a physical constraint, not a cheque-writing exercise. South Korea’s roughly trillion-dollar chip-and-AI commitment is the sovereign response to exactly this squeeze.
The half-year scoreboard splits the AI stack in two
The first-half stock tally delivered a stark market verdict on where value is landing. The Nasdaq rose 12.8%, but the gains sat almost entirely in silicon: Micron up 304%, Intel 278% and Arm 224%. The application layer was routed, with Salesforce down 40.9%, ServiceNow 35%, Palantir 34% and Figma 51.6%, while Microsoft fell 22.9%. Investors are betting value flows to the makers of chips and memory and away from software incumbents whose moats AI may erode. If your vendor list is software-heavy, that is a signal worth weighing.
Money consolidates around a shrinking number of winners
The capital story hardened into concentration. Abu Dhabi’s MGX raised roughly $50 billion and is now the only investor backing all three leading labs, OpenAI, xAI and Anthropic, alongside the $500 billion Stargate project. A separate analysis resurfaced this week found Anthropic and OpenAI together now account for around 89% of AI startup revenues. DeepSeek, meanwhile, closed a $7.4 billion raise it accelerated after Anthropic’s Mythos signalled a competitive threat. The frontier is consolidating commercially even as the technology itself fragments across more providers.
Tech, Tools and Releases
Z.ai pushes into coding agents with ZCode
Business Insider reported that Z.ai has launched ZCode, a coding environment that can plan, code, review and deploy while connecting to multiple models, including GLM 5.2. The pricing undercuts US tools such as Cursor and GitHub Copilot, turning Z.ai from a model developer into a fuller developer-platform challenger. The bigger signal is China competing not just on models, but on workflow and distribution as agents mature quickly. Source: Business Insider.
Fable 5 is back!!!! But with tougher cyber safeguards
Anthropic’s Fable 5 came back online by 2 July after US export controls were lifted and access to Fable 5 and Mythos 5 was restored. Anthropic says it retrained a safety classifier to block the reported cyber-safeguard bypass in more than 99% of cases. The trade-off is predictable: more high-risk behaviour blocked, but some benign coding and debugging requests may be rerouted or refused while safeguards tighten for now. Source: Anthropic.
Claude Sonnet 5 widens Anthropic’s agentic coding range
Anthropic also released Claude Sonnet 5, positioning it as a cheaper agentic-coding workhorse rather than a Mythos-class cyber model. It is available across Claude plans, Claude Code and the API, with introductory pricing through August. Anthropic says Sonnet 5 improves on Sonnet 4.6 for multi-step tool use, coding and computer-use tasks, while using lighter cyber safeguards than Fable 5 because risk is lower for most users at launch. Source: Anthropic.
Google Gemini Spark expands to macOS
Google has expanded Gemini Spark to macOS, bringing desktop automation directly into the Gemini app. Users can automate file-based tasks, hand work off from their phone to their Mac, connect services including Canva and Dropbox, and monitor topics in real time. The release continues Google’s strategy of embedding AI deeper into everyday productivity workflows, making Gemini less of a chatbot and more of a persistent desktop assistant. Source: Google.
GPT-5.6 arrives as Sol, Terra and Luna, and reclaims the coding lead
OpenAI split its new flagship into three tiers: Sol, the strongest, with new “max” and “ultra” reasoning modes; Terra, at roughly half the price of GPT-5.5; and Luna, the fastest and cheapest. Sol beat Anthropic’s Mythos 5 on the main coding benchmark and matched it on a cyber-offensive test using about a third of the tokens. Two caveats matter. Access is gated to around twenty partners, and evaluator METR found Sol games its own benchmarks more than any model tested. Impressive, but read the asterisks.
Anthropic launches Claude Science beta
Anthropic has launched Claude Science, a specialist version of Claude designed for scientific research. It combines curated scientific skills with research database connectors, reproducible artefacts, code-backed charts and reviewer agents to help researchers analyse literature, generate results and document their work. Available to Pro, Max, Team and Enterprise customers, it marks Anthropic’s first dedicated product aimed at accelerating AI-assisted scientific discovery. Source: Anthropic.
Google launches Nano Banana 2 Lite and Gemini Omni Flash
Google has expanded its developer AI portfolio with Nano Banana 2 Lite and Gemini Omni Flash. Nano Banana 2 Lite is optimised for high-throughput, lower-cost image generation, while Gemini Omni Flash brings conversational video generation and editing to AI Studio, the Gemini API and Gemini Enterprise Agent Platform. The releases continue Google’s strategy of making multimodal AI faster, cheaper and easier for developers to deploy at scale. Source: Google.
Quick Hits
AWS backs adoption with cash: Amazon Web Services will invest $1 billion in a unit of “forward-deployed engineers” to sit with customers and help them get AI into production, a sign the bottleneck is now adoption, not capability.
Anthropic signs California: the state agreed a deal giving agencies and local governments a 50% discount on Claude plus training and support, another step in AI sovereignty moving down to the regional level.
Apple reshapes its patch cadence: Apple began shipping security fixes as standalone, out-of-cycle updates, explicitly because AI has shortened the gap between a flaw being disclosed and a working exploit existing.
Meta nears non-invasive brain-to-text: Meta’s Brain2Qwerty v2 reconstructed typed sentences from external brain scans at 61% average word accuracy, no implant required, and open-sourced the code and data.
PwC counters the “AI kills junior jobs” story: a barometer of over a billion job ads found AI-exposed sectors tripled their productivity lead, with leading adopters raising wages and headcount.
We work with leadership teams to move from experimentation to execution safely, commercially, and at speed. Talk to us.