Chinese artificial intelligence developer DeepSeek has released a new experimental large language model, touting efficiency gains and steep price cuts that could intensify global competition in generative AI — in a manner reminiscent of how China has disrupted other industries, from solar panels to 5G equipment, by driving down costs.

The Hangzhou-based company said its DeepSeek-V3.2-Exp model is an “intermediate step toward our next-generation architecture,” describing it as more efficient to train and better at processing long sequences of text compared with earlier iterations. The announcement was made in a post on the developer platform Hugging Face.

DeepSeek, which rattled Silicon Valley in January with the launch of its V3 and R1 models, is once again signaling ambitions to undercut global rivals. The firm said the V3.2-Exp incorporates a new mechanism called DeepSeek Sparse Attention, designed to cut computing costs while boosting certain types of model performance.

In a parallel move aimed squarely at market adoption, DeepSeek said in a post on X on Monday that it was slashing API prices by “50%+.”

A Step Toward Next-Gen AI Architecture

While the V3.2-Exp is labeled “experimental,” analysts believe the release points toward a much larger leap. DeepSeek itself described the model as a bridge to its forthcoming next-generation architecture, which it framed as its most important product release since V3 and R1.

Those earlier rollouts drew extraordinary attention in January, not just from China’s domestic AI rivals like Alibaba’s Qwen, but also from U.S. giants including OpenAI, whose models dominate enterprise AI deployments. Investors and policymakers alike were startled at the pace with which a Chinese firm appeared able to close the gap with the West’s leading labs.

Although DeepSeek acknowledged that V3.2-Exp will not “roil markets” on the scale of those January debuts, the company suggested it could still mount meaningful pressure on competitors if it sustains a record of offering high capability at dramatically lower cost.

Winning by Efficiency and Price

The release draws parallels with how China has previously leveraged low-cost efficiency and scale to dominate global industries. Solar panel manufacturers, for example, flooded global markets in the 2010s with products priced far below Western rivals, reshaping the renewable energy supply chain. Telecoms giant Huawei took a similar path in 5G, offering cheaper infrastructure than European or U.S. firms.

DeepSeek’s move to halve API prices positions it in the same mold — a challenger using affordability and streamlined design to win share in markets where U.S. companies are charging premium rates.

The release puts renewed pressure on China’s domestic AI ecosystem, where companies like Alibaba, Baidu, and Tencent have poured billions into developing their own large models. Alibaba’s Qwen suite has been aggressively marketed to enterprise customers, while Baidu’s Ernie Bot remains a key government-backed project.

DeepSeek, by contrast, has staked its strategy on efficiency and affordability — an approach that resonates in both Chinese and global markets at a time when AI training costs are soaring.

For U.S. rivals, the implications are equally notable. OpenAI, backed by Microsoft, has positioned its GPT-4 and GPT-5 models as premium products, betting that enterprises will pay for frontier performance. Google’s Gemini and Anthropic’s Claude models are pursuing similar paths, with heavy investment in infrastructure and enterprise tooling. DeepSeek’s promise of comparable performance at half the cost could force pricing pressure in both consumer and enterprise segments.

Global Stakes of a Cheaper AI Alternative

What sets DeepSeek apart is not just its model architecture, but its pricing strategy. If the firm can demonstrate reliable performance in enterprise-grade tasks while cutting API costs by more than 50%, it could accelerate adoption in emerging markets — from Southeast Asia to Africa — where price sensitivity has slowed the uptake of U.S.-made AI tools.

Analysts warn that this could also sharpen the geopolitical dimension of the AI race. Washington has sought to curtail China’s access to advanced semiconductors, fearing Beijing’s rise in AI capabilities. But efficiency-driven architectures like DeepSeek’s could allow Chinese firms to do more with less, reducing their dependence on cutting-edge hardware.

While DeepSeek has cast V3.2-Exp as a stepping stone, if its next-generation release builds on the efficiency and cost trajectory it has outlined, analysts expect it to once again challenge assumptions about where the frontier of AI development lies — and at what price point it can be deployed globally.

Microsoft is betting that the same “vibe coding” trend that made app-building accessible to non-programmers can do the same for office work.

The company today introduced Agent Mode in Excel and Word and a new Office Agent in Copilot chat, both designed to turn prompts into complex documents, spreadsheets, and presentations.

“Today we’re bringing vibe working to Microsoft 365 Copilot with Agent Mode in Office apps and Office Agent in Copilot chat,” said Sumit Chauhan, corporate vice president of Microsoft’s Office Product Group. “In the same way vibe coding has transformed software development, the latest reasoning models in Copilot unlock agentic productivity for Office artifacts.”

At the product level, Agent Mode uses OpenAI’s GPT-5 to handle multi-step planning in Excel and Word, showing users a transparent breakdown of its reasoning in a sidebar. Microsoft stressed that it built in validation loops to ensure spreadsheets remain “auditable, refreshable, and verifiable.” Early benchmarks place Agent Mode at 57.2 percent accuracy on SpreadsheetBench, better than Shortcut.ai and Claude Files Opus 4.1 but still below the 71.3 percent human benchmark.

Agent Mode in Word pushes beyond summarization tools into what Chauhan calls “vibe writing”—a conversational drafting system where Copilot drafts, suggests refinements, and highlights what’s missing. The new Office Agent in Copilot chat, powered by Anthropic’s models, takes on PowerPoint and Word creation, offering structured slide decks with live previews and built-in web research.

“Productivity is our DNA, we’re Office,” Chauhan said. “While others will try to replicate us, there is no substitute for the real thing.”

For markets, these moves highlight a calculated hedge. Microsoft has leaned heavily on OpenAI, but by threading Anthropic’s models into GitHub Copilot, Office Agent, and Copilot Studio, the company is broadening its AI portfolio. That diversification could reassure enterprise customers wary of overreliance on a single provider. Yet it also raises questions for investors about the cost and complexity of maintaining parallel partnerships, especially as Anthropic’s models run on Amazon Web Services rather than Microsoft’s Azure.

The 57.2 percent accuracy rate in Excel is another detail financial analysts may seize on. For enterprise buyers, the number shows meaningful progress over rivals but still underscores the risk of errors in mission-critical areas like finance and compliance. Analysts suggest early adoption may skew toward use cases where errors carry lower stakes—drafting reports or presentations—before finance chiefs and auditors are comfortable entrusting balance sheets to Agent Mode.

Meanwhile, investors will likely parse the rollout strategy. Microsoft signals a controlled entry aimed at testing scale and reliability before pushing deeper into the desktop suite, where most enterprise workloads still reside, by confining Agent Mode to the web versions of Excel and Word at launch. That may slow adoption in the near term but could be read by markets as a risk-mitigation tactic designed to protect Office’s core reputation.

Agent Mode and Office Agent arrive as startups flood the market with AI-native productivity tools. Microsoft’s counter is not just integration but brand trust—positioning Copilot as a digital consultant capable of doing “what a first-year consultant would do, delivered in minutes.” Enterprise buyers weighing vendor sprawl versus bundled reliability could see this as a reason to keep their spend anchored in the Microsoft ecosystem, a factor that investors may interpret as a defensive but stabilizing moat around Office’s recurring revenue streams.

Agent Mode in Excel and Word, alongside Office Agent in Copilot chat, launches today in the Frontier program for Microsoft 365 Copilot customers and Microsoft 365 Personal/Family subscribers in the US. Desktop support will follow in the months ahead.

After nearly a decade of offering Memories as a free digital time capsule, Snapchat is drawing a line. The company said Monday that access to the feature will now be capped at 5GB, with additional storage pushed behind a paywall.

The new tiered system includes an introductory plan of 100GB for $1.99 per month, bundled as a standalone subscription. Snapchat+ subscribers ($3.99 monthly) will see their allotment raised to 250GB, while Snapchat Platinum subscribers ($15.99 monthly) will be granted a hefty 5TB of storage.

In an email to TechCrunch, Snapchat explained that when Memories launched, it was not expected to balloon into a repository of this scale. But user behavior told a different story: more than 1 trillion Memories have now been saved, underscoring the feature’s durability.

“It’s never easy to transition from receiving a service for free to paying for it, but we hope the value we provide with Memories is worth the cost,” Snapchat wrote in a blog post. “These changes will allow us to continue to invest in making Memories better for our entire community.”

To ease the transition, the company will offer 12 months of temporary storage for Memories that exceed the new 5GB limit. Users can also download their archives directly to personal devices. If no plan is chosen after that period, Snapchat says older Snaps will be preserved while newer ones exceeding the threshold will be deleted.

Snap insists the change won’t affect most of its user base, noting the “vast majority” remain under 5GB of saved Memories. Instead, the move targets what it calls “power users” — those with thousands of Snaps stored on its servers.

Investor Lens on Monetization

The policy shift is more than a product tweak — it is a signal to Wall Street. Snap has struggled to stabilize its advertising-dependent revenue model amid stiff competition from Meta and TikTok. The company is creating what investors often prize most: a recurring subscription stream tied to an emotionally sticky product by introducing paid storage.

Photos and videos, unlike premium filters or experimental AR features, are rarely abandoned by users once they’re archived. Analysts say this creates a strong incentive for subscribers to keep paying rather than face the risk of losing digital memories. That dynamic mirrors Apple’s iCloud and Google One, both of which have become reliable profit engines in their respective ecosystems.

However, the move reflects a broader trend for Snap, where platforms are monetizing infrastructure as much as content. The enormous costs of storing over one trillion Memories — a data set rivaling mid-sized cloud platforms — have become difficult to justify under a purely free model. Snap is attempting to avoid the backlash that often accompanies a universal paywall by selectively charging heavy users while keeping access free for casual ones.

Comparisons to Rivals

The structure resembles what Apple did with iCloud, initially offering free tiers before capping storage at 5GB and building out paid tiers that today contribute billions annually to its services revenue. Google Photos made a similar pivot in 2021, when it ended unlimited free storage. Both companies successfully conditioned users to see storage as a subscription commodity rather than a permanent free service.

Snap’s difference is its narrower focus: unlike Apple and Google, it does not control the hardware ecosystem. That means its leverage is primarily emotional — the attachment users have to Snaps as personal history. Analysts say that could either prove more fragile than ecosystem lock-in or, conversely, even stickier given the intimacy of Snapchat’s role in teen and young adult social lives.

Market Behavior

Some analysts believe that from a financial-market perspective, Snap’s move could shift investor narratives around the stock. Instead of being viewed purely as an ad-tech play competing against larger social platforms, Snap is attempting to reposition itself with hybrid economics: advertising for scale, subscriptions for stability.

The recurring storage subscriptions are believed to be sticky revenue, and investors tend to reward that stability. The bigger question, however, is whether Snap can scale these offerings enough to materially shift its revenue mix.

The introduction of Snapchat Platinum at $15.99 per month with 5TB of storage also places the company closer to enterprise-style pricing tiers seen in cloud computing — albeit aimed at consumers. If uptake is meaningful, markets may begin to value Snap less like a purely ad-driven social platform and more like a consumer-facing SaaS hybrid, a framing that could alter expectations on margins and cash flow.

For now, Snap is careful to stress that most users will not feel the change. But in financial markets, the focus will be on the minority who do. If even a small slice of Snapchat’s 400 million daily active users converts to paid storage, analysts say it could open a new recurring revenue channel worth hundreds of millions annually.