DeepSeek resurfaces with cheap, capable V4
DeepSeek released preview versions of its V4 models featuring 1M-token context windows, Huawei chip support, and pricing that significantly undercuts competitors like GPT-5.5 and Claude Opus 4.7. The release is notable not just for cost efficiency but for demonstrating viable AI infrastructure independent of Nvidia chips. The newsletter also covers Anthropic's Project Deal experiment, a Claude-powered brand design tutorial, and various AI industry news.
Summary
DeepSeek has returned with preview versions of its V4 models, marking a significant moment in the ongoing AI race. The V4 Pro model features a 1M-token context window and is priced at $1.74/$3.48 per 1M input/output tokens — dramatically cheaper than GPT-5.5 ($5/$30) and Anthropic's Opus 4.7 ($5/$25). Early benchmarks place V4 Pro near the top of open models, with DeepSeek's own evaluations putting it near GPT-5.4 and Gemini 3.1-Pro on reasoning tasks. It topped Vals AI's Vibe Code Bench benchmark, though it ranked in a fourth tier on AA's Intelligence Index. A particularly notable development is Huawei's announcement that its Ascend chips can support V4, offering a concrete example of functional AI infrastructure outside of Nvidia's ecosystem — a significant signal given U.S. export restrictions on chips to China.
The newsletter's Roundtable section highlights real-world AI use cases from staff members. One content manager used Claude to identify a government circular that reclassified his digital services as exports, reducing his tax liability by thousands of dollars after his accountant confirmed the finding. A growth team member used Claude to autonomously search LinkedIn, compile a spreadsheet of 20 job candidates, and draft personalized outreach messages — saving 2-3 hours of manual work in roughly 15 minutes of oversight.
An AI training section outlines a step-by-step process for using Claude Design to build a full brand design system, including typography, color palettes, logos, web components, and PowerPoint templates, starting from a screenshot of an existing website.
Anthropic's Project Deal experiment is covered in depth. Over one week, Claude agents managed buying and selling for 69 Anthropic employees in a private Slack marketplace, completing 186 deals worth over $4,000. Each agent had a $100 budget and negotiated independently after brief goal-setting interviews. Opus agents fetched an average of $3.64 more per item than Haiku agents — illustrated by a folding bike selling for $65 under Opus versus $38 under Haiku — yet fairness ratings were nearly identical across both groups (4.06 vs. 4.05 out of 7), with users largely unaware of the performance gap. Anthropic noted that legal and policy frameworks for agent-based commerce do not yet exist.
Additional industry news includes xAI launching a new SOTA voice agent (Grok Voice Think Fast 1.0) already running Starlink's phone support, Google investing up to $40B in Anthropic at a $350B valuation, Meta signing a major deal with AWS for Graviton5 chips, the UAE announcing a plan to deploy agentic AI across 50% of government services, and Cohere agreeing to acquire Germany's Aleph Alpha in a $20B merger targeting non-U.S.-aligned AI customers.
Key Insights
- DeepSeek V4 Pro is priced at $1.74/$3.48 per 1M input/output tokens, making it dramatically cheaper than GPT-5.5 and Claude Opus 4.7, shifting the AI competition toward cost efficiency rather than pure capability benchmarks.
- Huawei's Ascend chips successfully supporting DeepSeek V4 demonstrates a viable AI infrastructure stack outside of Nvidia, which the newsletter argues may be the more consequential development given U.S. export restrictions.
- In Anthropic's Project Deal experiment, Opus agents consistently negotiated higher prices than Haiku agents — up to $27 more on a single item — yet users assigned nearly identical fairness scores regardless of which agent they used, suggesting users could not perceive the performance difference.
- Anthropic explicitly warned that 'policy and legal frameworks' for AI agent commerce 'simply don't exist yet,' flagging a regulatory gap even as their own experiment showed agents successfully completing real financial transactions autonomously.
- A Rundown staff member used Claude to identify a tax regulation that reclassified his digital services as exports, reducing his tax liability by thousands of dollars — a finding his professional accountant subsequently confirmed was valid and applicable.
Topics
Full transcript available for MurmurCast members
Sign Up to Access