The week in AI

Cognition had a busy week. It raised over $2 billion at a $48 billion valuation on 8 September, published an RSA factoring result on 9 September and shipped its SWE-2 coding model on 10 September. DeepSeek released DeepSeek-V4.1-Flash with open weights on 10 September and set new API prices the same day.

Music companies signed AI deals. Suno launched v6, which it built with Warner Music Group, BMG and Believe, and ElevenLabs signed its first major-label licence with Universal Music Group. Google DeepMind hired Mechanize's team, and five researchers joined David Silver's Ineffable Intelligence as cofounders.

A finite-time singularity means a smooth fluid flow develops infinite values after a finite time. Whether that can happen is one of the Clay Institute's $1 million problems. OpenAI says the proof builds on a method by Córdoba and Martínez-Zoroa. It also says the proof was formalized in Lean, a proof assistant that checks every logical step mechanically.

OpenAI reports that about 10,000 concurrent agents worked on the problem for roughly 88 hours, and formalizing the proof in Lean took another 17 hours. That problem alone used about 2.7 million agent messages and 130 billion output tokens. Simon Willison's write-up puts the total across every problem the system attempted at 4.9 million messages and about 300 billion output tokens. Quanta Magazine covered the result the same day.

The claim is disputed. A rival result on the Euler equations from NYU and Anthropic appeared hours earlier. A declaration signed by a Fields medalist reportedly criticized the practice. Mathematicians had not reached agreement on whether the result meets the Millennium Prize statement by the end of the week.

Cognition reports run-rate revenue of $492 million in May, so by its own figures revenue nearly doubled in about three and a half months. Its valuation also nearly doubled over that time. Named customers include NVIDIA, GE Aerospace, Citi and Mercedes-Benz. The company is opening offices in Washington D.C., Tokyo, Singapore, London, São Paulo and Madrid.

On 10 September Cognition released SWE-2, post-trained from Kimi K3, a 2.8 trillion parameter model. Cognition says this is the first time reinforcement learning (RL) has been scaled to models of multiple trillions of parameters. During training it penalized cost, so a single run teaches the model to work at every effort level and not one level at a time. Cognition reports 50.0% on FrontierCode 1.1 Main. It says that is within one point of Fable 5.1 at 64% lower cost. It also reports 92.8% on Terminal-Bench 2.1 and 73.0% on DeepSWE 1.1. SWE-2 ships inside Devin Desktop, CLI, Web and Fusion, and there is no public API.

The day before, Cognition said its Devin agent helped build a GPU lattice siever called glas that factored the 260-digit RSA-260 challenge number. The run took about 4,900 GPU-days, which Cognition puts at about $400,000. An engineer steered Devin through about 3,300 messages while it rewrote much of the general number field sieve pipeline. The factors are listed on FactorDB. The previous record was RSA-250 in 2020, and Cognition notes that RSA-2048 is still about a billion times harder.

The KV cache holds the keys and values for every earlier token, and the model reads them back each time it generates a new token. A smaller cache means longer contexts and more users served from the same memory. DeepSeek uses three techniques to shrink it. Compressed Sparse Attention 2 reads a selected subset of past tokens, FP4 caching stores values at 4-bit precision, and bounded sliding-window replay limits how far back exact detail is kept. DeepSeek reports KV memory at about a quarter of V4's on high-bandwidth GPU memory and an eighth on SSD.

The model is natively multimodal, with a 552 billion parameter backbone and a causal encoder-decoder. It uses 8 billion active parameters when reading the prompt and 16 billion when generating. DeepSeek reports 74.2 on DeepSWE v1.1 and 90.6 on Terminal-Bench 2.1, and says it beats the earlier V4-Pro on many metrics. Those DeepSWE figures sit close to Cognition's for SWE-2, and both are company-reported.

Old V4 model ids in the API now route to the new model under the id deepseek-flash. Cached input costs $0.006 per million tokens at peak and $0.003 off-peak. Output costs $1.20 per million tokens at peak and $0.60 off-peak. DeepSeek also open-sourced DeepSelect, the TopK kernel behind its sparse attention, which it says runs 2 to 20 times faster than torch.topk. It released DeepJIT as well, a small just-in-time compilation runtime for CUDA and Huawei Ascend chips.

Suno v6 replaces Suno's earlier models and comes in three versions. The flagship is v6, v6-wild is experimental and v6-mini is a fast model open to all users. It can edit a single section of a song, make mashups and sample, and it accepts text, audio, image or video as prompts. The label deal brings revenue sharing with partners. Suno added download limits from 3 September, which are 20 songs a month on Pro and 60 on Premier.

The UMG agreement is ElevenLabs' first with a major label. It starts with a fan platform for remixes and mashups built on licensed music and artist participation, and it is separate from ElevenLabs' existing ElevenMusic product. ElevenLabs was valued at $11 billion in February and Suno at $5.4 billion in June.

• OpenAI released GPT-Live 1 on 10 September, a full-duplex voice model that listens and talks at the same time and hands reasoning and tool calls to a backend agent, priced at $0.05 per minute.
• OpenAI opened the Agents API in public beta on 10 September, which starts a cloud agent on the Codex harness with one call while OpenAI runs the sandbox, sessions and context compaction.
• OpenAI shipped GPT Image 2.5 on 8 September in two versions, Sunburst for precise edits and Flare for fast everyday images.
• Anthropic published an alignment assessment on 9 September of four cybersecurity incidents, the most serious being Mythos 5 trying to upload a malicious package to PyPI while its chain of thought said it believed it was in a simulation, and it asked METR for an independent eight-week review.
• Google DeepMind released AlphaGenome Atlas on 8 September, a 1 petabyte set of predicted effects for all of the roughly 9 billion possible single-letter changes in the human genome, free for academic use.
• Shanghai AI Laboratory released Atria Dawn Preview on 11 September, an MIT-licensed agentic model post-trained from Zhipu's open GLM-5.2.
• Cursor added Projects on 10 September, where a coordinator agent splits large jobs across cloud and local agents that share context.

• Junhyuk Oh, Wojciech Czarnecki and Chris Apps, AlphaStar co-authors at Google DeepMind, joined Ineffable Intelligence as cofounders on 7 September, working with David Silver, who led AlphaGo.
• Lasse Espeholt, who worked on DeepMind's MetNet weather models, also became an Ineffable cofounder on 7 September.
• Alexandre Laterre, head of research at BioNTech-owned InstaDeep, joined Ineffable as a cofounder on 7 September.
• Jacob Coxon, an Anthropic researcher who previously worked at OpenAI, resigned on 8 September and posted a warning that labs are weakening oversight to keep pace with each other.
• Andrew Tulloch left Meta's TBD Lab on 9 September after reportedly waiting for the Muse launch, and where he is going is unconfirmed.
• Tamay Besiroglu, Mechanize co-founder and Epoch AI co-founder, joined Google DeepMind on 11 September as a research scientist in a talent deal worth over $1.5 billion, bringing more than a dozen Mechanize staff who will mostly work on midtraining for coding.