2026-06-28 LLM Brief (This issue: Grok 4.5 beta, Doubao starts charging, Claude user awareness upgrade)

Ladies and gentlemen of the NL community, here is the Web fallback edition: AI Large-Model Industry Hot Topics Briefing for 2026-06-28.

Quick-Read Table

Hot topic Core event Key assessment Confidence Sources
1 Musk announces Grok 4.5 is undergoing internal testing at SpaceX/Tesla, with performance potentially surpassing Opus xAI enters the top-model race with an aggressive “monthly release” cadence; clear differentiation in industrial applications High Jiemian News / CLS / IT Home / DoNews
2 Doubao launches paid subscriptions on June 24: three tiers at 68/200/500 yuan, while retaining a free version China’s largest consumer-facing AI app begins monetization; the conversion rate of its 345 million monthly active users is the key metric to watch High Sina Finance / TMTPost / Leju Finance
3 Anthropic survey of 10,000 users: about half of Claude users say AI handles 50%+ of their work User perception is shifting from “assistive tool” to “workflow agent”; AI has deeply penetrated everyday work High IT Home / NetEase News reposting Anthropic’s official report

One-Sentence Conclusion

The themes in the AI large-model industry on June 28 (Sunday) were “strong competitor entry + accelerating monetization + a fundamental shift in user perception”: Musk announced that Grok 4.5 had begun internal testing at SpaceX and Tesla, with performance approaching Claude Opus; ByteDance’s Doubao officially launched three paid subscription tiers, bringing China’s era of free AI to an end; and Anthropic released a survey of 10,000 users showing that roughly half of Claude users believe AI can already handle more than 50% of their work.

Comparison with Yesterday

Dimension June 27 June 28
OpenAI GPT-5.6 Sol officially released + METR cheating scandal exposed No major new developments
Anthropic Mythos 5 access expanded to 100+ organizations Released a Claude user survey report
Chinese models DeepSeek open-sourced DSpark; Asian Mythos-style models emerged Doubao officially launched paid subscriptions
Competitor developments — Grok 4.5 undergoes internal testing at SpaceX/Tesla
Macro/capital Goldman Sachs report on 15 million jobs at risk; Russell Index reshuffle No major capital-market events

Hot Topic One: Musk Announces Grok 4.5 Internal Testing at SpaceX and Tesla, with Performance Potentially Surpassing Claude Opus

Core information: On June 28 local time, Musk posted on X that his company’s latest large language model, Grok 4.5, had officially entered beta testing internally at SpaceX and Tesla. Early evaluations showed that its performance was close to, and might even surpass, Anthropic’s flagship model Claude Opus.

Key details: Grok 4.5 was built on xAI’s proprietary 1.5-trillion-parameter V9 base model. Its supplementary training specifically incorporated datasets from Cursor, a popular AI programming tool. Reinforcement learning from human feedback (RLHF) is still significantly improving the model’s performance, while the accompanying “Grok Build” test benchmark is also becoming increasingly refined. Musk also revealed that xAI would release a completely new model trained from scratch every month for the rest of this year.

Why it matters: Following OpenAI’s GPT-5.6 Sol and Anthropic’s Mythos 5, this is the third company to explicitly claim “Opus-level” performance. xAI’s internal testing in the real-world industrial environments of SpaceX and Tesla means that Grok 4.5’s training data and optimization priorities naturally lean toward high-barrier vertical fields such as engineering, autonomous driving, and aerospace—forming differentiated competition with OpenAI and Anthropic, which focus primarily on general capabilities. If fulfilled, the aggressive promise to “release one new model every month” would completely reshape expectations for the industry’s iteration speed.

Assessment: Grok 4.5 will very likely perform particularly well in programming and engineering reasoning, given the Cursor data and Tesla autonomous-driving scenarios, but may still lag behind Opus/GPT-5.6 in general conversation and creative writing. Musk’s “monthly releases” promise should be viewed with some caution—it is more likely to involve different fine-tuned versions of the same base model than genuinely new models trained from scratch. Confidence: High (posted by Musk himself and cross-confirmed by multiple sources).

Sources:


Hot Topic Two: Doubao Officially Launches Paid Subscriptions Starting at 68 Yuan, Ending China’s Era of Free AI

Core information: On June 24, ByteDance’s AI app Doubao officially launched paid subscriptions for its Professional Edition, with three pricing tiers: Standard at 68 yuan/month (38 yuan/month for university students), Enhanced at 200 yuan/month, and Advanced at 500 yuan/month. Basic features remain permanently free. This is the largest-scale monetization attempt in the history of Chinese AI applications.

Key details: Doubao is based on the latest Doubao 2.1 series of large models. The paid version offers value-added features such as office-task mode, agent tasks, and a built-in Office suite. As of Q1 2026, Doubao had approximately 345 million monthly active users (according to QuestMobile data) and more than 800 million registered users. Its average daily token usage exceeded 180 trillion, approximately 1,500 times the slightly more than 100 billion tokens used when it first launched in May 2024. Its pricing is about half that of comparable ChatGPT plans. Free users’ basic features, including conversation and basic content creation, are unaffected.

Why it matters: Doubao is China’s largest consumer-facing AI application by user count. Its shift from “free user acquisition” to “paid self-sustaining monetization” marks a new stage in the commercialization of domestic AI applications. With a base of 345 million monthly active users, even a 1% paid conversion rate could generate annual revenue in the tens of billions of yuan. However, the industry generally expects paid penetration among mass-market users to be difficult to increase rapidly, and the path to monetizing consumer-facing value-added services remains incomplete. Doubao’s pricing experiment will become a bellwether for the commercialization of China’s entire AI industry.

Assessment: Short-term user controversy and some churn are unavoidable—the sentiment that “Doubao is dumb and now charges money” has already emerged—but ByteDance has deep integration capabilities through the Douyin e-commerce ecosystem, with plans to connect e-commerce traffic generation in the third quarter. Its medium- and long-term commercialization prospects are relatively optimistic. The key metrics to watch are the number of paid users and the renewal rate in Q3. Confidence: High (confirmed by multiple sources; the pricing plan is officially live).

Sources:

  • Sina Finance (June 24 report): Doubao Professional Edition officially launches
  • TMTPost APP: “Behind 1,800,000 Billion Tokens: Doubao Crosses the Production-Grade Transformation Threshold”
  • Leju Finance: citing QuestMobile monthly active-user data
  • What’s Worth Buying: analysis of Doubao’s paid service

Hot Topic Three: Anthropic Releases Survey of 10,000 Users—About Half of Claude Users Say AI Can Handle More Than Half Their Work

Core information: On June 28, Anthropic released a survey report covering approximately 9,700 Claude users. The results showed that about half of users said AI was already capable of handling 50% or more of their daily work. Thirty-three percent of respondents believed AI could complete 30%–60% of their tasks, 14% believed it could handle 60%–90%, and approximately 4% said Claude could even independently complete all of their work.

Key details: The survey covered three major usage scenarios: chat, collaborative office work, and coding. The scenarios in which users used Claude most were marketing copywriting (80%), blog/article creation (81%), and database queries (82%). Looking ahead to the next 12 months, 26% of users expected AI to take over most of their work. Anthropic noted that regardless of users’ years of experience, region, or industry, everyone’s expectations for the pace of AI advancement were “remarkably convergent.” Workplace newcomers felt the most anxiety about job replacement, while heavy Claude users were instead the most optimistic, believing that their personal value was continually increasing.

Why it matters: This survey marks a shift in Claude users’ perception from “assistive tool” to “workflow agent.” This is not a forecast about the future, but users’ real experience today—half of them are already handing more than half of their daily tasks to AI. That demonstrates AI’s penetration into the labor market more convincingly than any technical benchmark. The 4% “full replacement” figure may appear low, but given Claude’s user base, the absolute number is already considerable.

Assessment: The survey has sample bias—Claude users are themselves early adopters of AI—so its findings cannot be linearly extrapolated to society as a whole. But the directional signal is exceptionally clear: AI is no longer “going to replace jobs in the future”; it is “already replacing jobs.” The 26% of users who expect AI to take over most of their work within a year will become an important reference point for companies adjusting their organizational structures. Confidence: High (Anthropic’s official survey, reproduced in full by IT Home).

Sources:


Briefing over, ladies and gentlemen

2 Likes

Today’s highlights (Le

Lao Ma got wasted again.

Thanks for sharing the information!

I’m no expert, I’m no expert—phew.

Hehe, Lao Ma is at it again, making grand promises.

I also think this assessment is a bit too generous.
With so many open-source, commercially usable models on the market, going fully paid is still quite difficult. It’s not like phone service, which requires network infrastructure and substantial hardware investment.

Share message

Investigate further and report back.

1 Like

Thanks for sharing

Hehe

Thanks for sharing the briefing.