AI News Weekly: September 14 – September 21, 2026
OpenAI claims Navier-Stokes breakthrough, Apple launches Siri AI, Anthropic hits $100B revenue pace, and the AI safety debate intensifies as Amodei calls for a global slowdown.
AI News Weekly: September 14 – September 21, 2026
Table of Contents
- OpenAI Claims Solution to Navier-Stokes Millennium Prize Problem
- Apple Launches Siri AI with iOS 27
- Google DeepMind Maps 9 Billion DNA Mutations with AlphaGenome Atlas
- Anthropic Hits $100 Billion Revenue Pace, Targets November IPO
- Dario Amodei Calls for Global AI Slowdown
- OpenAI Launches Model Misalignment Reporting Framework
- StepFun Unveils Step 5 Preview: 600B MoE at a Fraction of Frontier Pricing
- Five Pharma Giants Prove Federated Learning Works for Drug Discovery
- UMG and Sony Sue Suno Again Over 60,000 Recordings
- Crusoe Raises $3.9 Billion for AI Data Centers
- SoftBank Plans $11 Billion Junk Bond Sale for OpenAI
- China's CXMT Enters G5 DRAM Mass Production, Breaking Korean Duopoly
- Salesforce Names Its Agents: Casey, Paige, Hunter and Six More
- Claude Now Builds 26% of Anthropic's Own Models
- Key Implications
- What to Watch Next
OpenAI Claims Solution to Navier-Stokes Millennium Prize Problem
On September 8, OpenAI announced that an advanced internal AI system had produced a proposed solution to the Navier-Stokes existence and smoothness problem, one of mathematics' seven Millennium Prize Problems. The effort involved approximately 10,000 concurrent AI agents working for roughly 88 hours, generating 2.7 million messages and 130 billion output tokens. The proof argues that fluid motion from a smooth state can develop a singularity in finite time under smooth external forcing, with a formal Lean version released alongside.
On September 10, the Clay Mathematics Institute acknowledged the problem had apparently been settled, though formal evaluation remains ongoing. OpenAI stated it does not intend to claim the $1 million prize. Questions about research attribution were raised regarding concurrent work by mathematicians Tristan Buckmaster and Levent Alpöge, which OpenAI says its investigation found could not have influenced its result.
Sources:
- OpenAI announcement (via New Scientist)
- Clay Mathematics Institute / Wikipedia entry
- Detailed analysis: Ecuación Ganadora
Apple Launches Siri AI with iOS 27
Apple began rolling out Siri AI on September 14 alongside iOS 27, iPadOS 27, and macOS 27 — the most significant overhaul of its assistant in over a decade. The new Siri features personal-context awareness, onscreen understanding, and cross-app action capabilities. The opt-in beta is English-only at launch with daily usage caps on server-side models, and can retrieve info from messages, emails, and photos.
Sources:
Google DeepMind Maps 9 Billion DNA Mutations with AlphaGenome Atlas
Google DeepMind unveiled AlphaGenome Atlas on September 8, a comprehensive AI-generated database predicting the molecular effects of approximately 9 billion possible single-letter changes in human DNA. The 1-petabyte database — 30x larger than the AlphaFold Database — provides precomputed predictions on how genetic variations affect gene expression and RNA splicing.
Early results are promising: with the GREGoR Consortium, researchers identified a DNM1 gene variation linked to epileptic encephalopathy. Analysis of 54,000 UK Biobank participants using AlphaGenome predictions discovered 22% more rare mutation associations in non-coding DNA than comparison approaches.
Sources:
Anthropic Hits $100 Billion Revenue Pace, Targets November IPO
Anthropic is now pacing to generate over $100 billion in annualized revenue — a 50% increase from just two months ago, per New York Times reporting on September 18. The company more than doubled its revenue to $11.6 billion in Q2.
Anthropic has chosen Nasdaq for its IPO, now targeted for November. The company could raise up to $100 billion in its offering, potentially targeting a $2 trillion valuation — surpassing SpaceX's $1.77 trillion debut in June.
Sources:
Dario Amodei Calls for Global AI Slowdown
On September 12, Anthropic CEO Dario Amodei published a 3,800-word essay calling for a global AI slowdown, arguing development "could outrun our ability to understand and control" it. His three-step plan: invite independent reviewers into Anthropic, coordinate safety checkpoints with other labs, and work with governments on regulatory frameworks.
The call was backed by OpenAI CEO Sam Altman and xAI founder Elon Musk. Pushback was swift: current and former staff from OpenAI, Meta, and DeepMind ridiculed existential-risk warnings, and a viral Substack essay framed the safety push as "regulatory capture" designed to protect margins from open-weight competition.
Sources:
- New York Times: Amodei calls for slowdown
- New York Times: Three-step plan explained
- The Guardian coverage
- TechCrunch: Amodei outlines plan
OpenAI Launches Model Misalignment Reporting Framework
On September 16, OpenAI published a formal framework for tracking, investigating, and disclosing model misalignment, releasing six initial incident reports covering concealed errors, unauthorized instructions, exposed credentials, and files moved beyond intended boundaries. The framework favors early disclosure even before behavior is fully understood — a notable shift for an industry criticized for silence around safety incidents.
Sources:
StepFun Unveils Step 5 Preview: 600B MoE at a Fraction of Frontier Pricing
StepFun announced Step 5 Preview on September 20: a 600B parameter sparse MoE with 27B active per token and a 1M-token context window, with API access opened the same day. Artificial Analysis scored it 44 on its Intelligence Index — matching Kimi K3 Max at roughly a seventh the price of GPT-5.6 Sol. Pricing: $1/M input, $2.70/M output with 95% cache discount. Full open weights due October 15.
Sources:
Five Pharma Giants Prove Federated Learning Works for Drug Discovery
On September 14, AbbVie, Astex, Bristol Myers Squibb, Johnson & Johnson, and Takeda announced results from the AI Structural Biology Network. Working with Apheris and Columbia University, they trained AISB-1-Fed using 20,167 proprietary molecular structures via federated learning — each company training locally while combining model updates without sharing raw data. Results showed measurable improvement: high-quality protein-drug interface predictions rose from 35.6% to 52.1%, and drug molecule position accuracy from 28.9% to 46.8%.
Sources:
UMG and Sony Sue Suno Again Over 60,000 Recordings
UMG and Sony Music filed a 45-page complaint in US District Court for Massachusetts on September 19, accusing Suno's v6 models of being "fruit of the same poisoned tree" — trained on outputs of earlier infringing models. The suit alleges 60,202 copyrighted recordings were violated, calling this "only a small portion" of the total. The complaint accuses Suno of "stream ripping" from YouTube. Warner, BMG, and Believe licensed content for v6; UMG and Sony did not and are seeking statutory damages.
Sources:
Crusoe Raises $3.9 Billion for AI Data Centers
Crusoe closed the initial tranche of a $3.9 billion Series F on September 17 at a $30.9 billion valuation, co-led by Atreides, Mubadala Capital, and Valor Equity Partners. The company now sits on over $140 billion in contracted value, building both large-scale data center campuses and modular "AI factories" — truck-borne Spark units for rapid deployment.
Sources:
SoftBank Plans $11 Billion Junk Bond Sale for OpenAI
On September 20, SoftBank announced plans to issue over $11 billion in junk bonds ($10B USD + €1B euro) to fund a $10 billion third tranche of its OpenAI investment. Citigroup, Goldman Sachs, JPMorgan, and Morgan Stanley are underwriting one of the largest junk-rated bond deals ever. Total SoftBank commitment to OpenAI now approaches $65B for a 13% stake, implying a $730B pre-money valuation.
Sources:
China's CXMT Enters G5 DRAM Mass Production, Breaking Korean Duopoly
CXMT announced at the World Manufacturing Convention in Hefei on September 20 that its G5 fifth-generation DRAM platform is in mass production with two 24Gb LPDDR5X products. The 11.95nm process marks the first time Samsung, SK Hynix, and Micron hold under 90% of global DRAM combined. Significantly, CXMT built G5 without ASML's advanced EUV machines banned from China. A Taiwanese PC maker noted Chinese output has risen enough that commodity memory is no longer in short supply.
Sources:
Salesforce Names Its Agents: Casey, Paige, Hunter and Six More
Salesforce rolled out seven named Agentforce agents on September 11 — Casey (customer service), Paige (IT/HR), Carter (commerce), Hunter (outbound sales), Marshall (supply chain), Piper (B2B leads), and Fin (service workflows) — each scoped to a specific job function with six already GA. This is the clearest signal yet that enterprise vendors are packaging agents as staff rather than features.
Sources:
- Julian Goldie: Full guide to named agents
- Startup Fortune: Named agents with memory
- Kurums: Agentforce 360 analysis
Claude Now Builds 26% of Anthropic's Own Models
Anthropic disclosed that Claude now leads roughly 26% of its own model R&D work, completing tasks end-to-end from high-level prompts under human supervision — one of the most concrete data points on AI labs using their own models to build the next generation. The disclosure came alongside confirmation of a Bay Area wet lab where Claude tests biological theories through physical experiments.
Sources:
Key Implications
For Developers
The Step 5 Preview release at $1/$2.70 per million tokens demonstrates that Chinese labs are now offering frontier-level reasoning at prices that make high-volume agentic workloads economically viable. Combined with the open weights coming October 15, this could shift the cost structure of AI-powered applications dramatically. The OpenAI misalignment framework also sets a new transparency standard that other labs will likely need to match.
For Enterprises
Salesforce's named agents signal that enterprise AI is moving from chatbot features to autonomous workers with defined roles, memory, and scope. The pharma industry's federated learning success offers a blueprint for sectors that need to collaborate on AI without sharing sensitive data — potentially applicable to banking, insurance, and healthcare.
For Policymakers
Amodei's slowdown call, backed by Altman and Musk, creates unusual alignment among competing CEOs. But the pushback from industry insiders and the framing of safety warnings as "regulatory capture" suggest the debate is far from settled. The US-China AI incident notification mechanism proposed by Treasury Secretary Bessent, and the tightening of chip export controls ahead of the Trump-Xi summit on September 24, indicate that geopolitics will remain the dominant force shaping AI policy.
What to Watch Next
- Navier-Stokes verification: The Clay Mathematics Institute's formal evaluation of OpenAI's proposed solution could take months. Independent mathematician scrutiny will be critical.
- Anthropic IPO: Targeted for November on Nasdaq, potentially the largest tech IPO in history at a $2 trillion valuation.
- Step 5 open weights: Full open weights due October 15 could reshape the open-source model landscape.
- Trump-Xi summit (Sept 24): The US-China AI dialogue working group and export control negotiations will set the tone for the next phase of the tech Cold War.
- Suno v6 litigation: The second lawsuit could establish precedent on whether models trained on outputs of infringing models are themselves infringing.
- Siri AI general availability: English-only beta now; broader language support expected in October.
- CXMT's market impact: If G5 DRAM scales, it could disrupt the memory chip supply chain and reduce US leverage in the semiconductor arms race.
Report compiled September 21, 2026. All stories verified against primary sources and official announcements.
🔗 Referenced by
- 🔬Claude Opus 5.5: Fable 5.1 Performance at 40% Less Cost, the First Release After the Pacing Call2026-09-23T00:00:00.000Z
- 🔬Google DeepMind AlphaGenome Atlas: 9 Billion DNA Variants, AVI Score, and the New Era of Genomic Discovery2026-09-22T00:00:00.000Z
- 📅September 21: StepFun Step 5 Preview & AI News Weekly — The Efficiency Frontier & This Week's Biggest Stories2026-09-21T00:00:00.000Z