GPT-6 reportedly clears 92% on agentic SWE-bench-verified, doubling last year's frontier.
Internal evals leaked to a researcher mailing list show the next OpenAI model resolving real GitHub issues end-to-end at a rate that, if confirmed, would be the first time a public-bound model crosses senior-engineer parity on contained tasks.
Operator takeaway
Coding agents are no longer the demo. Expect every enterprise AI roadmap to add an 'autonomous engineering' line item in the next planning cycle.
Anthropic closes a $9B Series F at a $190B valuation — Google and a Gulf sovereign anchor.
The round nearly doubles the company's prior mark and is structured with a mix of cash and committed GPU credits from Google Cloud, mirroring the OpenAI/Microsoft template that now defines frontier-lab financing.
Operator takeaway
Frontier AI is now a sovereign-scale capital game. Three Western labs and two Chinese labs will define the 2027 model landscape — everyone else is a customer.
JPMorgan reports $2.1B in measurable productivity gains from internal LLM rollout.
In a rare disclosure, JPM's COO attributed the figure to deployments across legal review, equity research drafting, and contact-center triage — with 60% of the gain in just two of those three workflows.
Operator takeaway
The pattern keeps repeating: enterprise AI value concentrates in two or three workflows per company. The rest is overhead. Pick the two ruthlessly.
Nvidia's Rubin Ultra ships ahead of schedule; TSMC commits an extra CoWoS-L line.
Hyperscaler allocations for Rubin Ultra are already oversubscribed through Q1 '27. The TSMC line addition signals that packaging — not silicon — has become the true bottleneck of frontier compute.
Operator takeaway
If your AI strategy assumes commodity GPU pricing in 2027, rebuild the model. Compute is going to stay scarce and priced like beachfront real estate.
EU AI Act Phase II enforcement begins — fines now uncapped for high-risk systems.
Today marks the start of full enforcement for high-risk AI categories under the Act, with the first formal investigations opened against two recruiting-software vendors operating in France and the Netherlands.
Operator takeaway
AI governance is no longer a compliance checkbox; it's a board-level liability. Any model touching hiring, credit, or healthcare in the EU needs an auditable lineage by quarter-end.
Saudi Arabia's HUMAIN strikes a $12B compute partnership with a US frontier lab.
The deal will site 1.2GW of training capacity in NEOM by late 2027, with energy supplied by dedicated solar and nuclear contracts. Equity terms remain undisclosed; technology transfer is restricted.
Operator takeaway
The Gulf is not just funding AI — it's hosting it. The geography of frontier training is fragmenting along sovereign capital lines, not technical ones.
Klarna walks back AI customer-service claims; rehires 1,200 human agents.
After two years of headlines about AI replacing the contact center, Klarna's CEO disclosed in the earnings call that customer satisfaction had dropped six points and that escalations now require a hybrid model.
Operator takeaway
The first AI wave automated answers. The next wave will have to automate judgment — and that is a meaningfully harder problem. Expect more public retreats this year.
DeepSeek V4 open-weights matches GPT-5 on long-context retrieval at 1/30th the inference cost.
The release puts a credible Chinese open-weights option within striking distance of US frontier models on every public benchmark except multimodal reasoning — and at an order of magnitude lower serving cost.
Operator takeaway
The economic floor of 'good enough' AI is collapsing weekly. Any enterprise still paying frontier prices for commodity reasoning workloads is leaving 80%+ of the cost on the table.
Sequoia closes a dedicated $4B AI-Apps fund — first cheque size raised to $25M.
The fund is explicitly oriented toward post-training application companies, with a stated thesis that frontier model commoditization will push value capture into vertical workflows over the next 24 months.
Operator takeaway
The smart money has officially called the bottom of the application layer. If you've been waiting for a sign to ship an AI-native vertical product — that was the sign.
US Treasury proposes mandatory model-card disclosure for any AI system used in lending decisions.
The draft rule would require lenders to publish architecture summaries, training-data provenance, and fairness metrics for every model materially used in credit underwriting, with 18 months to comply.
Operator takeaway
Financial services just became the first US sector with binding AI transparency obligations. Healthcare and insurance are next. Get your model registry in order now, not after the comment period closes.