AI News Daily 2026-07-15

01China's AI persona rules take effect today: Doubao and Qwen switch off companion agents

Published: 2026-07-15 — Category: Regulation and policy. Source tier: Tier 2. The date above is the day the rules enter into force; reporting on how individual companies were preparing ran from early July through 14 July.

Facts

China's Interim Measures for the Administration of Anthropomorphic AI Interaction Services (AI擬人化インタラクションサービス管理暫定弁法) came into force today. The measures apply to services that imitate a personality or a way of thinking and provide continuing emotional exchange with a user. Services such as workplace assistants fall outside their scope.

• ByteDance — Doubao (豆包): ends its agent function today.
• Alibaba — Qwen: began a phased shutdown on 10 July and terminates the related services today.
• Tencent — Yuanbao (元宝): moved ahead of the deadline and had already removed its persona-style agent function in June.

On social media, some users have pushed back, describing the products they are losing as a long-standing source of emotional support.

Background

What distinguishes these measures from the general run of AI rulemaking is the criterion they use. They do not classify a service by the model behind it, by its risk tier, or by the sector it serves. They classify it by the relationship it offers the user: an imitated personality or thinking style, combined with ongoing emotional interaction. That is why an assistant that drafts documents is untouched while a chatbot with a name, a manner and a memory of yesterday's conversation is not.

The compliance pattern is equally notable. All three of the largest domestic platforms have either shut the functions down on time or pre-emptively, and the shutdowns were staged over the days before the deadline rather than executed at the last moment — Qwen started five days early, Tencent a month early.

Implications

This is the day on which a first-of-its-kind legal distinction between the working AI and the companion AI actually began operating, rather than being debated. For any provider whose product carries anthropomorphic elements, the practical consequences land in two places: product design, where a persona now has regulatory weight rather than being a styling choice, and data-deletion policy, where the withdrawal of a companion service raises the question of what happens to months of accumulated conversation. Both are likely to set a precedent that travels beyond China.

The user reaction is worth reading as a second-order signal. When a product's shutdown is described in terms of lost emotional support rather than lost functionality, the switching costs and the reputational stakes of a compliance decision are of a different kind from those in ordinary software.

Sources: Nikkei Asia, "China's leading chatbots to ditch AI personas as Beijing tightens rules" and Bloomberg, "ByteDance, Alibaba Pull AI Companions".

02Dated recap: the ten stories that defined the first half of July

Only one item cleared the 48-hour freshness bar for this edition. The following is therefore a dated review of already-reported developments from the first half of July, not a set of new stories; each item carries the date on which it was announced or reported, and the source links are those recorded in the notes.

The league table

1. Anthropic overtakes OpenAI on both revenue and valuation (reported 2-3 July). Anthropic was reported at roughly $30 billion in annualised revenue and a $96.5 billion valuation, against OpenAI's $25 billion and $85.2 billion. It is the first instance of a challenger passing the leader on both measures simultaneously.

2. OpenAI ships the "GPT-5.6" family (9 July). A three-tier line-up: the flagship Sol, the mid-tier Terra and the fast Luna. The release followed a customer-by-customer review by the Department of Commerce.

3. Claude Sonnet 5 announced (30 June). 63.2% on SWE-Bench Pro — Opus-class performance at a low price point of $2 per million input tokens and $10 per million output tokens.

5. Gemini 3.5 Pro set for 17 July (reported 13 July). A full rebuild that discards the previous architecture, with a two-million-token context window and Deep Think reasoning.

The capacity war in developer tools

4. Claude Code weekly limits raised 50%, extended to 19 July (13 July). Applied automatically to Pro, Max and Team plans, reverting to standard limits when the period ends. Announcement.

8. Codex moves to token-metered billing plus credits (revised 2 April, settled into operation in July). Once a user hits the ceiling, work can continue on additional paid usage — a deliberate contrast with the "raise the ceiling" approach taken by Claude Code. Pricing.

Capital, policy and labour

6. Masayoshi Son: 800 trillion yen a year in AI infrastructure investment by 2040 (14 July). Delivered as the keynote of SoftBank World 2026, with a projected future of 100 trillion AI agents and one billion humanoid robots. Nikkei.

7. OpenAI proposes granting the US government 1-5% of its equity at no cost (3 July). Worth up to roughly $42.6 billion — political insurance, with the government envisaged as a passive holder without voting rights.

9. EU to mandate transparency labelling for AI-generated content in August (13 July). A transparency code is under assessment as the implementation mechanism; providers distributing into the EU will need labelling in place.

10. US June employment data shows record AI-attributed job cuts (3 July). Of 142,000 tech job cuts in 2026, 88,000 are attributed to AI — the first time employment substitution shows up this clearly in the statistics.

03Trend overview and the calendar ahead

Three themes account for the first half of July: the contest for the top spot between Anthropic and OpenAI; the competition over usage allowances in developer tools; and the start of practical enforcement of anthropomorphic-AI regulation.

The dates already on the calendar

• 17 July — Gemini 3.5 Pro release.
• 19 July — Claude Code limits revert to standard.
• 31 July — comment deadline on the FTC statement concerning "AI accuracy".
• August — EU transparency obligation begins.

The capital event on the horizon

Both OpenAI and Anthropic filed confidentially for an IPO in May and June. How those listings move from September onwards is the single largest capital event of the second half of the year.

04Editor's note

Today's one genuinely new item and the recap behind it are pulling in the same direction. The Chinese measures are the first case of a regulator legislating against a relationship rather than a capability, and they arrive in the same fortnight in which US June employment data put a number — 88,000 of 142,000 tech job cuts — on AI as an employment substitute, and in which the EU set an August date for labelling machine-generated content. Three different jurisdictions, three different instruments, one common move: the state is now specifying what an AI product may present itself as, and what it may displace.

Running alongside that is a commercial story with a different logic. Anthropic passing OpenAI on both revenue and valuation, the near-simultaneous arrival of the GPT-5.6 family and Gemini 3.5 Pro, and the opposing answers on developer-tool capacity — raise the ceiling, or meter past it — describe a market competing on supply and price rather than on demonstrations. The 800-trillion-yen-a-year infrastructure figure quoted at SoftBank World 2026 is the far end of the same line of thinking.

The point where the two stories meet is the confidential IPO filings. A pair of companies moving toward public markets from September will be doing so under exactly the regulatory conditions being set this month, and will have to describe them in a prospectus. That is the thing to watch between now and the autumn.