From model races and mathematical milestones to sharp questions about trust in the lab, this has been an unusually eventful 24 hours in artificial intelligence. DeepSeek pushed its KV-cache architecture to new extremes, OpenAI claimed a first-ever AI solution to a Millennium Problem while facing fresh allegations about how it got there, and Anthropic came under scrutiny for building what critics call a predictive surveillance apparatus. Below are the five stories that defined the day.
1. DeepSeek launches V4.1 Flash: cheaper, more capable, dramatically leaner
Chinese lab DeepSeek released DeepSeek-V4.1-Flash, a multimodal Mixture-of-Experts model that the company bills as both cheaper and more capable than its own V4 Pro, and which topped the day’s Hacker News discussion with its unusually detailed technical report. The model packs 552 billion backbone parameters while activating just 8B per token during prefill and 16B during decode — a cost efficiency the team credits to a new Causal Encoder-Decoder (CED) architecture: a 20-layer causal encoder feeding a 20-layer decoder.
The headline advance is memory. A new Compressed Sparse Attention 2 (CSA2) design and FP4 KV caching shrink the global KV cache to roughly 890 bytes per token — about a quarter of V4-Flash’s footprint — with a reported cache-hit price as low as $0.003 per million tokens. The model is trained from scratch on 45T tokens, offers a continuously controllable reasoning effort (1–100), and is released under an MIT license. Independent benchmarks show it reaching a 3,471 Codeforces rating, 74.2% on DeepSWE v1.1, and 90.6% on Terminal-Bench 2.1. One side effect: the jump from 284B to 552B parameters drew mixed reactions from the local-inference community, who found the larger model far harder to run on consumer hardware.
2. OpenAI claims a first: AI solves Navier-Stokes — and a fierce credit dispute erupts
OpenAI says it has solved the Navier-Stokes problem, one of the Clay Mathematics Institute’s six Millennium Problems carrying a $1 million prize. The company recounted an 88-hour assault by some 10,000 autonomous AI agents that found a case where the fluid equations break down, later formalized in Lean — potentially the first time artificial intelligence has solved a genuinely major open problem in mathematics.
The cost was enormous: roughly a thousand-fold more than the $2,000 Astra spent on earlier math breakthroughs, with independent estimates of the inference bill landing near $1 million. Chief Research Officer Mark Chen put the total in “millions of dollars.” Yet the achievement was quickly overshadowed by controversy. Mathematician Tristan Buckmaster (NYU) and Levent Alpöge (affiliated with Anthropic) had been closing in on the same result for a month using AI tools — including OpenAI’s own Codex. Buckmaster says OpenAI offered him sole authorship of the solution, but only if Alpöge were removed and the write-up credited an internal OpenAI model. He refused, and questions persist over whether OpenAI’s agents had access to the pair’s Codex logs — fueling the day’s sharpest debate about what researchers can trust when they share unpublished work with frontier labs.
3. GPT-6 Astra: looped transformers, hidden reasoning, and the “recurrent depth” rumour
Sebastian Raschka’s deep-dive “Ahead of AI” essay on OpenAI’s freshly released GPT-6 Astra became the reference read of the day. Raschka calls it “the best model I’ve used” and notes it “leapfrogs its GPT-5.6 predecessor in practically all categories” — especially 3D rendering, animation, and computer use, backed by a standout 99.9% on ARC-AGI-3 (vs. 7.8% for GPT-5.6 Sol).
The meatiest part is architecture gossip. Following reporting in The Information, Raschka unpacks “looped transformers” and “recurrent depth” — the idea that Astra re-applies the same transformer blocks at inference time, effectively doubling depth from 22 to 44 block applications without adding new weights. That looped design sits at the center of the longstanding rumour that Astra hides its chain of thought, reasoning internally rather than emitting a transparent trace. Raschka also connects the dots to reporting that OpenAI bought tens of thousands of Mac Minis and Mac Studios — not to train on GPUs, but to expose macOS to the model during RL so Astra learns computer use on real operating systems.
4. Cognition’s SWE-2: RL at multi-trillion-parameter scale for coding
Cognition (the Devin developer) unveiled SWE-2, its most advanced coding model, claiming the first reinforcement-learning run scaled to the multi-trillion-parameter regime. Built on Kimi K3’s 2.8T-parameter backbone, SWE-2 scores 50.0% on FrontierCode 1.1 Main — within one point of Fable 5.1 while 64% cheaper — and posts 73.0% on DeepSWE 1.1 and 92.8% on Terminal-Bench 2.1. The technical headline is a Pareto-informed cost penalty: a single RL run trains all reasoning-effort levels at once by penalizing each tier’s cost to match the base model’s efficiency frontier.
Community reaction was double-edged. Skeptics pointed to a wide Terminal-Bench 4 gap (27.3% vs. GPT-6 Astra’s 57.9%) and asked why the model is not open-weights — it is closed; SWE-2 lands in Devin’s Desktop, CLI, Web, and Fusion products. Others noted that at 53 steps vs. SWE-1.7’s 127 average, the “focused exploration” gains look real.
5. Anthropic’s predictive surveillance ambitions draw scrutiny
The American Prospect‘s Daniel Boguslaw reported that Anthropic is building an extensive monitoring system to track activists opposed to rapid AI development. Drawing on job postings and interviews, the piece details a “pre-crime” approach — attempting to predict incidents before they occur, in some cases reporting suspects to police before a crime has happened. Anthropic’s security team contracts with risk-detection firm Samdesk, and executives have cited receiving about 60 minutes of advance warning about a protest to reroute themselves.
The most jarring anecdote: Anthropic reported a user to San Francisco police after he told Claude he had bought an AR-15 and had CEO Dario Amodei “in his sights” — a remark the man later dismissed as “just fucking around” — yet refused to hand police the actual messages. The reporting landed as Anthropic advertises a Global Safety, Intelligence & Security role (paid $180,000–$230,000) tasked with tracking “geopolitical instability, terrorism, crime, activism,” and critics contrasted the push with CEO Dario Amodei’s own recent words that the public’s distrust of tech is “fundamentally a crisis of trust.”
That is where AI stands this Friday: dazzling capability advances racing alongside increasingly urgent questions about cost, credit, and consent.
