Good morning,
This week we cover the great model rush of early July — five frontier launches in ten days, and what it changes about how you should buy intelligence. Plus the usual top news, including Apple suing OpenAI, the first AI-run ransomware operation, Tencent taking back Manus, and the summer's regulatory calendar around the globe.
GPT 5.6, Grok 4.5, Muse Spark 1.1…. we are back in the race
Since the launch of ChatGPT in 2022, we have repeatedly been led to think that one actor had pulled ahead and would eat the market. Each time, the field closed the gap within months. The last couple of weeks proved it again: this is still a race. What is genuinely new this cycle is the double focus — not just intelligence, but pricing and accessibility.
In the space of ten days: Anthropic's Claude Fable 5 came back online worldwide (1 July) after the US lifted its export-control order — still the most capable model on the market, and priced accordingly (pay to win as of today, quite literally). OpenAI answered with GPT-5.6 (9 July) — Sol, Terra and Luna — plus ChatGPT Work, an agent designed to turn a goal into finished work over hours, aim at Claude’s Cowork. Sol is a notch below Fable on raw intelligence, but at $5/$30 per million tokens the pricing is, frankly, remarkable in comparison. Then the surprise: Grok 4.5 (8 July), the phoenix rising from xAI's ashes under SpaceXAI ownership. Trained on Cursor's data after last month's acquisition, it ranks fourth overall on Artificial Analysis at $2/$6 per million tokens. Add Meta's Muse Spark 1.1, its first paid model (Alexandr Wang's bet finally paying off?), and Google's Gemini 3.5 Pro, which should launch soon but has now missed two deadlines. They are late, and in a moment when everyone shipped, that silence was loud.
Two nuances matter for anyone comparing models. First, price should not be measured in $/token. Models consume very different amounts of tokens for the same job: Anthropic's recent Sonnet 5 looks cheap per token, but it thinks at length and spins off sub-agents, resulting in high usage and a high bill. Others charge more per token but work efficiently, and come out cheaper overall. The better metric is cost per task — Artificial Analysis now publishes exactly this alongside its intelligence rankings, and it reshuffles the leaderboard considerably.
Second, openness. If the past six months (two worldwide shutdowns, one government-gated launch) taught us anything, it is not to depend on a single frontier lab — which means your models must be easy to swap and orchestrate across platforms. Here the intelligence leader loses: Anthropic is among the hardest to integrate, with real reasoning masked from developers and subscriptions that cannot be connected to external harnesses such as OpenCode or Pi. Winning the intelligence race and losing the openness one is a strange trade.
From a company standpoint, this is the moment to define a proper intelligence access strategy:
Which models for which use cases? Read Ethan Mollick's A Guide to Which AI to Use in the Agentic Era, a great non-technical map of the landscape.
Which platform orchestrates it all? Read The Information’s last piece on "Model Routers", a technology that Palantir or Databricks are pursuing.
When to use private US models versus open European or Asian counterparts? Read more on Chinese model: Center for Strategy and International Study (CSIS) “What to know about Chinese AI Models”.
Who gets the expensive frontier models versus the basic tiers, and with what budget (Tesla now caps AI usage at $200 per employee per week)? Read: SemiAnalysis's TokenBudgeting conversations with 50+ enterprises on where the caps are landing.
And, perhaps most importantly: where do we not need AI? Read a quick piece from
A story I quite like on this last point: Microsoft's new Head of Xbox, Asha Sharma, who previously led CoreAI, one of Microsoft's main AI divisions, refused to put AI features into Xbox. If she doesn't want it everywhere, neither should you. It is time to retire the "everyone, everywhere, as much as they want" approach.
Why it matters: Intelligence is strategy decision more than just a tech procurement one. The winners of the next 18 months won't be the companies using the smartest model, but those routing the right intelligence, at the right price, to the right task — with a hedge against any single lab (or government) switching the lights off.
More top news
Nvidia teams up with its rivals. After years of building the best alone, Nvidia announced a string of partnerships with competing chipmakers the most recent one being d-Matrix. This represent a strategic shift from "outrun everyone" to "coach the whole ecosystem," keeping every roadmap under its guidance (and, conveniently, its control).
Geneva hosted the UN's first all-nations AI dialogue. The Global Dialogue on AI Governance (6–7 July) convened all 193 member states. The scientific panel co-chaired by Yoshua Bengio and Maria Ressa warned that science "cannot guarantee" frontier AI won't cause catastrophic harm, and quantified the concentration: the US holds 75% of the world's top AI supercomputing power, China 15%. The forum was follow by another AI summit in Geneva: AI for Good.
Tencent re-takes control of Manus. After Beijing signalled clear displeasure at Meta — an American firm — acquiring Chinese AI know-how, Tencent moved to bring the agent startup back under Chinese ownership.
Apple sues OpenAI. Filed 10 July in California: trade-secret theft and breach of contract, alleging a coordinated campaign to poach Apple staff (400+ former employees now at OpenAI) and extract confidential hardware knowledge for the Jony Ive-linked device.
The summer regulatory roadmap. From 2 August the EU can fine general-purpose AI providers up to 7% of global turnover, retroactive to August 2025. China's anthropomorphic-AI rules bite on 15 July — ByteDance and Alibaba are shutting down user-created companion agents entirely. A great tracker to follow it all from Law Firm White&Case.
JADEPUFFER: the first AI-run ransomware. Sysdig documented the first end-to-end ransomware operation driven by an LLM agent, exploiting a known vulnerability, harvesting credentials, moving laterally, and recovering from a failed step to a working fix in 31 seconds, unaided.
SK Hynix listed on Nasdaq — and hit a $1tn market cap. The Korean memory maker raised $26.5bn in the largest US IPO ever by a foreign company, closing up 13% on day one, with its chairman telling investors that AI memory demand "is enormous."
The Fed put AI on the monetary policy dashboard. The Federal Reserve appointed VC a16z co-founder Marc Andreessen to co-lead a new task force studying AI's impact on jobs, productivity, and rates.
The China race amplifies. Chinese open-weight models now handle up to 46% of enterprise API tokens on US developer platforms (average last year: 11%), at prices 60–90% below Anthropic and OpenAI. Anthropic responded by closing the loopholes — Singapore subsidiaries, VPNs — that let Chinese firms access Claude, weeks after its own models were export-blocked.
Tools to try: Artificial Analysis
Given this week's theme, the tool to try isn't a model but a scoreboard. Artificial Analysis (artificialanalysis.ai) independently benchmarks every major model on intelligence, speed, and cost per task rather than cost per token, accounting for how many tokens each model actually burns to finish a job. Their Coding Agent Index is where the Sol-vs-Fable and Grok-value claims from above come from. For anyone drafting the "intelligence access strategy" discussed earlier, twenty minutes on their comparison pages is the fastest way to ground the debate in numbers rather than vendor slides. Fair warning: the rankings move weekly. That, of course, is rather the point.
Have a good week,