DeepSeek Just Open-Sourced a Rival to Claude Code

AI Agents

DeepSeek Just Open-Sourced an Autonomous Coding Agent — Right Before Anthropic’s IPO

DeepSeek has shipped a prototype “Super-Agent” that doesn’t wait to be asked: it proactively monitors developers’ group chats, spots coding issues, and suggests fixes without being summoned. The open-source release, announced Saturday by the DeepSeek Harness team, is aimed squarely at Anthropic’s Claude Code — the tool that has become the default for serious agent-driven development.

Why this matters: the coding-agent war is now a price war disguised as a feature war. Claude Code runs on Anthropic’s closed models, while DeepSeek’s agent is open-weight and dramatically cheaper to run. For anyone building or operating AI employees, the takeaway is simple — the cost of “hands” is collapsing. If an autonomous coder that monitors your team’s chat and fixes issues proactively becomes a commodity, the scarce skill shifts from writing code to briefing the agent well and reviewing its output. That’s a skillset shift, not a savings headline.

What to watch: the timing is the story. Anthropic is reportedly weeks from a Nasdaq listing targeting a $1.5 trillion valuation. DeepSeek landing an open, cheap rival to Claude Code at exactly that moment is a deliberate squeeze on the narrative that foundation-model companies can defend premium pricing forever. Expect more “open-source ships a cheaper X” moves as the IPO window opens.

Source: Bloomberg

Deepgram Buys Daily — and Gives Pipecat Its Independence

Voice-AI company Deepgram acquired Daily, the infrastructure behind the open-source Pipecat framework — then handed Pipecat to an independent foundation run by competitors, not by Deepgram. It’s a deliberate anti-lock-in move: voice agents are consolidating, and the winner is betting developers will pay for a neutral standard (a new “Real-Time Context Standard”) rather than a walled garden.

Source: ShiftDelete

Karpathy Ships “HackerTube” in ~2 Hours With Claude Opus 5.5

Andrej Karpathy built and shipped HackerTube — a YouTube alternative seeded with 6,454 videos from Hacker News — in roughly two hours using Claude Opus 5.5. It’s the clearest hands-on proof yet that a single skilled builder plus a strong agent can now go from idea to live product in an afternoon.

Source: AI Weekly

AI News

OpenAI Cuts Its Funding Target as Investors Demand Real Economics

OpenAI reportedly lowered its latest fundraising target — the WSJ pegged the round at a $1.25 trillion valuation, below the $1.4 trillion it sought. The “easy money” phase is over: investors are now running diligence on compute costs, inference efficiency, and retention, not just growth. For builders, the signal is that AI capability is no longer the scarce thing — sustainable unit economics are.

Source: The Tech Buzz

Reddit Plunges 29% as AI Search Eats Its Traffic

Reddit beat revenue expectations yet still fell more than 29% after disclosing Google’s AI Mode is draining about a million visits a week. It’s the clearest number yet for what AI Overviews mean to any site that depends on search — publishers like Reuters and the Washington Post have already lost 35–44% of Google referral traffic. If you publish content, this is the moment to ask where your readers will come from when the search box starts answering for you.

Source: BigGo Finance

Texas A&M’s “Aggie AI Foundry” Puts AI in a Physical Loop

Texas A&M announced a $2 million automated metallurgy lab where AI proposes new metal alloys and robots mix, melt, and test them — compressing development from months to under a week. The lesson isn’t the grant size; it’s the architecture. Models alone don’t invent materials — models wired into a loop that continuously feeds real test data back in do. Watch who builds the loops, not who ships the model.

Source: Tech Insider