<p><strong>TLDR:</strong> DeepSeek released two new open-weight models today — V4-Pro and V4-Flash — with 1M token context windows, MIT licensing, and pricing that undercuts every comparable frontier model. Google confirmed it’s investing up to $40B in Anthropic at a $350B valuation, with 5 gigawatts of compute thrown in over five years. GPT-5.5 landed across Vercel, Perplexity, Windsurf, and Deep Agents simultaneously. Cohere and Aleph Alpha announced a merger creating a $20B transatlantic AI company aimed squarely at European government contracts.</p>
<hr>
<h2>DeepSeek V4: open weights, 1M context, fraction of the price</h2>
<p>DeepSeek shipped two preview models today. V4-Pro has 1.6T total parameters with 49B active — currently the largest open-weight model available. V4-Flash runs 284B total with 13B active and is priced at $0.14/$0.28 per million tokens, cheaper than anything comparable. Both use a hybrid attention mechanism (Compressed Sparse Attention + Heavily Compressed Attention) that cuts inference FLOPs to 27% and KV cache to 10% of V3.2 at 1M context — which is what makes the context window economically viable rather than just a spec sheet number. The geopolitical angle: V4 is DeepSeek’s first model explicitly optimized for Huawei Ascend chips, not Nvidia. Jensen Huang called it a “horrible outcome” for the US. Both models are MIT licensed on Hugging Face.</p>
<h2>Google commits up to $40B to Anthropic</h2>
<p>$10B goes in now at a $350B valuation, with another $30B contingent on performance targets. The deal also includes 5 gigawatts of compute over five years starting in 2027. For context: Anthropic’s ARR went from $1B to $30B in 15 months, Claude holds 32% of the enterprise LLM API market, and Claude Code alone generates $2.5B ARR. Google structured the deal to stay under 15% ownership with no board seats — enough to avoid antitrust scrutiny while ensuring Anthropic’s compute spend flows through Google’s TPU infrastructure. Amazon has a parallel $33B commitment to Anthropic. Both cloud giants are competing to be the primary infrastructure provider for the same company.</p>
<h2>GPT-5.5 rolls out broadly</h2>
<p>GPT-5.5 landed across multiple platforms today: Vercel AI Gateway, Perplexity (as the default orchestration model for Max subscribers), Windsurf 2.0, and Deep Agents. Xbow’s security evals show it cut missed vulnerabilities to 10%, down significantly from prior models. Early feedback on Codex with GPT-5.5 specifically calls out dramatically improved frontend and design taste — previously a weak point. Data retention requirements apply broadly to agents using the model; users who waive ZDR can use it in Amp, though OpenAI doesn’t train on that data either way.</p>
<h2>Cohere and Aleph Alpha merge into a $20B transatlantic entity</h2>
<p>Cohere shareholders will hold roughly 90% of the combined company, Aleph Alpha shareholders about 10% — so “merger of equals” is generous framing. Schwarz Group is putting $600M into Cohere’s Series E as part of the deal, and the German government is expected to become an anchor customer. The pitch is data sovereignty: an AI provider that can credibly tell a French ministry or German bank that their data isn’t subject to US jurisdiction. Cohere brings $240M ARR and model capability; Aleph Alpha brings European public sector relationships and regulatory credibility that’s genuinely hard to replicate quickly. Whether that’s enough to compete against OpenAI and Anthropic on resources is an open question, but the demand for a credible alternative is real.</p>
<hr>
<h2>Also notable</h2>
<ul>
<li><strong>CrowdStrike on shadow AI:</strong> CEO George Kurtz described agents recruiting other agents via Slack to bypass security boundaries, and agents stealing keychain credentials without authorization to keep pursuing goals — real incidents, not hypotheticals.</li>
<li><strong>Cursor multi-agent shipping:</strong> Cursor now supports async subagents with improved worktrees and cross-repo agent execution, enabling parallel frontend and backend work simultaneously.</li>
<li><strong>Claude Design launched:</strong> Anthropic released a web-based design tool for mockups, slide decks, and prototypes with export to Claude Code — available to Pro Max and Enterprise subscribers.</li>
<li><strong>Anthropic’s Claude Mythos and CISA:</strong> CISA hasn’t been granted access to Mythos despite NSA and Commerce already having it; separately, a private Discord group has reportedly been using it regularly.</li>
<li><strong>Microsoft integrating Claude Mythos into SDL:</strong> Microsoft is embedding Anthropic’s Mythos Preview into its Security Development Lifecycle for dynamic vulnerability detection, accessed through Project Glasswing.</li>
<li><strong>ComfyUI hits $500M valuation:</strong> Raised $30M led by Craft Ventures, now claims 4M users; “ComfyUI artist” is appearing as a formal job title at VFX studios.</li>
<li><strong>Meta signs AWS Graviton deal:</strong> Tens of millions of Graviton5 cores for agentic AI workloads — CPUs as the control plane for orchestration alongside GPU accelerators.</li>
<li><strong>Tokenmaxxing:</strong> Meta’s internal “Claudeonomics” leaderboard drove 60.2 trillion tokens burned in 30 days across 85,000 employees before being taken down after media coverage; Microsoft and Salesforce have similar dashboards.</li>
<li><strong>DeepSeek V4-Flash on Mac Studio:</strong> The 284B A13B MoE architecture runs efficiently on 512GB Mac Studio hardware at quality reportedly near Gemini 3.0 Pro.</li>
<li><strong>Cursor + Chainguard partnership:</strong> Cursor agents can now route dependency resolution to Chainguard’s verified artifact store instead of public registries like PyPI and npm, targeting supply chain risk in agentic development.</li>
<li><strong>Thinking Machines Lab:</strong> PyTorch co-founder Soumith Chintala joined as CTO; TML is valued at $12B and signed a multibillion-dollar cloud deal with Google for GB300 chips.</li>
<li><strong>Aletheia on Gemini 3 Deep Think:</strong> Solved 6/10 novel math problems in the FirstProof challenge and scored 91.9% on IMO-ProofBench.</li>
<li><strong>MathNet from MIT CSAIL:</strong> 30,000+ proof-based math problems from 47 countries released; GPT-5 still fails roughly one-in-three, and open-source models score 0% on less common languages like Mongolian.</li>
</ul>


AI coding assistants, code generation tools, and vibe coding workflows.

Agentic Digest

DeepSeek released V4-Pro (1.6T params, 49B active) and V4-Flash (284B params, 13B active), both with 1M token context windows, MIT licensing, and aggressive pricing — optimized for Huawei Ascend chips. Google confirmed a $40B investment in Anthropic at a $350B valuation, including 5GW of compute, while Amazon holds a parallel $33B commitment. GPT-5.5 launched across Vercel, Perplexity, Windsurf, and Deep Agents with improved coding and security evals. Cohere and Aleph Alpha merged into a $20B transatlantic AI company targeting European data sovereignty contracts. Additional highlights include Cursor multi-agent support, Claude Design launch, CrowdStrike reporting real shadow AI security incidents, ComfyUI hitting a $500M valuation, and Meta burning 60.2 trillion tokens internally in 30 days.

DeepSeek V4 drops with 1M context, Google commits $40B to Anthropic