AI 竞赛重燃

上周，Anthropic 和 OpenAI 看起来几乎独占了 AI 竞赛：它们推出的模型愈发强大，将竞争对手远远甩在身后。Google 似乎已落后得毫无希望；Meta 则尽管投入数十亿美元以重振其陷入困境的 AI 雄心，仍未能证明自己是一名真正的竞争者。即使是来自中国的最新模型，尽管表现令人印象深刻，看上去也仍在追赶。

然而，仅仅几天时间，一切都变了。Google 和 Meta 都发布了前沿模型，并声称这些模型能够与 OpenAI 和 Anthropic 最好的产品正面竞争。AI 竞赛突然又回来了——而且如今谁都有机会胜出。

直到本周，Google 看上去都像是已经永久失去了自己在 AI 领先者之列的位置。自去年晚些时候发布 Gemini 3 以来，这家科技巨头带来的大多只是渐进式改进。Gemini 3 刚推出时实力强劲，但其领先优势转瞬即逝；此后，它难以跟上 OpenAI 和 Anthropic 的后续产品，甚至难以匹敌一些来自中国的开放权重模型。

一系列内部调整进一步加深了外界的印象：Google 正在从 AI 前沿后撤。资深高管 Jeff Dean 与多名 AI 研究人员离开公司，创办了一家新创业公司。曾多年执掌 Google AI 重镇 DeepMind 的诺奖得主 Demis Hassabis，也不再负责日常领导工作。其他与 AI 相关的团队被从 DeepMind 移出，并入 Google 的其他部门。这些变化共同表明，Google 的 AI 重心正在发生更广泛的转移：从伦敦转向加州，并引发外界猜测，这家科技巨头正开始将 AI 商业化置于研究之上。

不过，周二 Google 以发布 Gemini 4 Argon颠覆了这一叙事。这是一款新的前沿模型，公司称其在多项基准测试中能够媲美甚至超越 OpenAI 和 Anthropic 的旗舰产品——尽管基准测试如今究竟还有多大价值，仍值得商榷。更重要的是，Google 为 Argon 制定了极具攻击性的价格：在推广期内，其每 token 价格最多可比竞争对手低 80%，即使促销结束后，价格仍显著更低。

Meta 的登场或许更具戏剧性。与凭借雄厚财力和 DeepMind 长期被视为 AI 领域标杆的 Google 不同，Meta 从未真正确立其在前沿 AI 领域的有力竞争者地位。这并非因为它没有努力，而是因为其模型始终未能产生显著影响。去年，该公司启动了一项雄心勃勃且成本高昂的 AI 业务重组，在人才和基础设施上投入数十亿美元，其中包括收购 Scale AI 及引入其联合创始人 Alexandr Wang。今年 4 月，其新成立的 Superintelligence Labs 发布了首个重要模型 Muse Spark。它令人印象深刻，但还不完全属于前沿模型。

这一局面在周一发生改变，Meta 终于发布了一款能够与业内顶尖产品竞争的模型。根据公司的测试，Muse Wonder Full可在多项基准测试中与 OpenAI 和 Anthropic 的领先模型抗衡，其中包括 GPT-6 Astra、Claude Opus 5.5 和 Claude Fable 5.1。换句话说，Mark Zuckerberg 的公司终于抵达了 AI 前沿。

Google 的回归——以及 Meta 的加入——令 AI 竞争变得更加活跃，也表明 AI 世界的力量平衡能够以多快的速度发生变化。两家公司都拥有极其雄厚的资金实力，以及被数十亿人使用的庞大产品生态系统；这是 OpenAI 和 Anthropic 这样的 AI 新兴公司无法比拟的优势。Meta 的 Muse Agent 助手取得成功，正说明了这一优势。它在如此短的时间内获得流行，尽管围绕其安全与隐私问题的担忧不断增加，仍展现出触达数十亿现成用户的力量。与此同时，OpenAI 必须更加努力地建立自己的生态系统，并将自己定位为 Meta 的替代选择。

这一切发生之际，AI 竞赛似乎正进入新阶段。开发能力越来越强的模型，正变得愈发棘手。日益增多的安全担忧和 AI 失控事件，使要求放缓开发速度的呼声更加强烈。开放模型也在不断进步，让任何一家公司都更难长期保持领先。

如果单纯的能力和性能已不足以确保领先，那么 Google 和 Meta 拥有 Anthropic 与 OpenAI 所不具备的大量优势。它们或许花了很长时间追赶，但过去一周表明，任何人若将这两家公司排除在竞争之外，都将是极其不明智的。

Last week, Anthropic and OpenAI looked like they had the AI race to themselves, pulling ahead with increasingly powerful models that left their rivals in the dust. Google seemed hopelessly far behind, while Meta, despite spending billions to revive its floundering AI ambitions, had yet to establish itself as a serious contender. Even the latest models coming out of China, impressive as they were, appeared to be playing catch-up.

Then, in the space of a few days, everything changed. Google and Meta both unveiled frontier models they claim can go toe to toe with the best from OpenAI and Anthropic. Suddenly, the race is back on — and it’s anyone’s game.

Until this week, Google looked like it had permanently lost its place among the AI front-runners. Since releasing Gemini 3 late last year, the tech giant had delivered little more than incremental improvements. Gemini 3 was formidable when it launched, but its lead was fleeting, and it struggled to keep up with later offerings from OpenAI and Anthropic, or even some of the open-weight models emerging from China.

A series of internal shake-ups only added to the sense that Google was retreating from the AI frontier. Veteran executive Jeff Dean departed the company with a number of other AI researchers to launch a new startup. Demis Hassabis, the Nobel laureate who spent years at the helm of Google’s AI powerhouse DeepMind, stepped back from day-to-day leadership. Other AI-related teams were moved out of DeepMind and into Google proper. Together, the changes signaled a broader shift in Google’s AI center of gravity, moving away from London and toward California, fueling speculation the tech giant was beginning to prioritize commercializing AI over research.

Then, on Tuesday, Google upended that narrative with the release of Gemini 4 Argon, a new frontier model the company claims matches or beats flagship offerings from OpenAI and Anthropic on numerous benchmarks, whatever value benchmarks are still worth. What’s more, Google has priced Argon aggressively: Per token, it’s undercutting its rivals by as much as 80 percent during its introductory period, and is still significantly cheaper once that promotion ends.

Meta’s arrival was arguably even more dramatic. Unlike Google, which had long been the one to beat in AI thanks to its deep pockets and DeepMind, Meta had never really established itself as a serious contender at the frontier. It was not through lack of trying, but its models struggled to make a mark. Last year, the company embarked on an ambitious and expensive overhaul of its AI operations, spending billions on talent and infrastructure, including Scale AI and its cofounder Alexandr Wang. In April, its new Superintelligence Labs released its first major model, Muse Spark. Impressive, but not quite frontier.

That changed on Monday when Meta finally unveiled a model capable of competing with the industry’s very best. According to the company’s testing, Muse Wonder Full can compete with leading models from OpenAI and Anthropic, including GPT-6 Astra, Claude Opus 5.5, and Claude Fable 5.1, on numerous benchmarks. In other words, Mark Zuckerberg’s company has finally arrived at the AI frontier.

The return of Google — and the arrival of Meta — makes the AI competition considerably more dynamic and shows just how quickly the balance of power can shift in the AI world. Both have extraordinarily deep pockets and vast ecosystems of products used by billions of people, advantages that AI upstarts like OpenAI and Anthropic simply cannot match. The success of Meta’s Muse agent assistant illustrates that advantage. Its popularity in such a short period of time, and despite the growing list of security and privacy concerns, shows the power of having billions of users within ready reach. OpenAI, meanwhile, is having to work harder to establish an ecosystem of its own, positioning itself as an alternative to Meta.

All this comes as the AI race appears to be entering a new phase. The drive to develop ever more powerful models is becoming increasingly fraught. The growing array of safety concerns and rogue AI incidents means calls to slow development are growing louder. Open models are also getting better and better, making it harder for any one company to stay ahead for long.

If raw power and performance are no longer enough to stay ahead, Google and Meta have plenty of advantages not enjoyed by Anthropic and OpenAI. They may have spent a lot of time playing catch-up, but the past week shows it would be incredibly foolish to count either of them out.

Follow topics and authors from this story to see more like this in your personalized homepage feed and to receive email updates.

• Robert Hart

• AI

• Anthropic

• OpenAI

• Tech