Anthropic pointed 950 Claude agents at a DNA database and came back with an enzyme system nobody had named. Cognition said its annualized revenue run rate crossed $1 billion, roughly double where it stood in May. Akamai committed $11.6 billion to Anthropic's CPU workloads over seven years. Alibaba unveiled the Zhenwu V900, a training-and-inference chip it calls the strongest in China. Meta put an agent called Muse into glasses, a keychain-sized device, and a 100-gram headset. Alibaba's Qoder turned the messy conversation between a developer and an agent into a product surface. And two arXiv papers argue the agent harness should be baked into model weights rather than bolted on.
Anthropic's 950 agents found a CRISPR-like enzyme system
Anthropic published the first result from its new life sciences group and, for the first time, disclosed a Bay Area wet lab that has been running quietly since spring 2026. The lab is staffed by human scientists and handles only BSL-1 and BSL-2 work, so nothing that infects humans. The task handed to Claude was narrow: point it at a large DNA database and hunt for uncharacterized reverse transcriptases, the enzymes that copy RNA back into DNA.
About 950 Claude agents ran for 21 hours, burned roughly 210 million tokens, and scanned more than 200,000 reverse transcriptases. They consolidated 3,500 new candidate systems and narrowed that down to 20 reports for human review. Humans touched the work in exactly two places: the opening prompt and the wet-lab validation at the end.
The system is named ART, for array-associated reverse transcriptases. In phage genomes it sits as a trio: a reverse transcriptase, a neighboring partner gene of unknown function, and a long run of evenly spaced DNA repeats. The CRISPR comparison holds because CRISPR's arrays store the RNA sequences that make the system programmable. The reverse transcriptase itself, from a jumbo phage, had been described before, and a Stanford team independently described a similar system. Anthropic's claim is that Claude noticed the combination first: the adjacent repeat array plus the uncharacterized helper protein. One agent's working notes read "that's a CRISPR-like ... repeat array?!"
The first wet-lab result came when scientists expressed the system and the array was transcribed into a defined set of short RNAs, which is how CRISPR arrays behave. That is an early signal, not a mechanism proof. What ART does in nature is still unknown, and Anthropic chose to publish a preprint rather than wait for peer review. Dario Amodei said on X that the exact function, the biotech uses "if any," and the significance all remain unclear, "but at least this is work I would have been proud of in my PhD." Feng Zhang, the CRISPR pioneer at MIT and the Broad Institute, reviewed it and called it a fine example of what AI agents can contribute to biology.
Eric Kauderer-Abrams, who leads the life sciences division, told Reuters the lab looks no different from an ordinary molecular biology lab. The difference is upstream, where Claude scans sequences and surfaces candidates before a human picks up a pipette. Anthropic says it will not run clinical trials and will not compete with pharma, and lists Roche, Genentech, and Bristol Myers Squibb as partners. The Verge's Robert Hart reported that the practical applications remain unclear and that publishing this early is unusual by ordinary scientific standards; his piece also reads the push as part of Anthropic's pre-IPO positioning and its recruiting of scientists. Six named scientists worked with Claude on the finding.
— Anthropic · The Verge
Devin's maker crossed a $1 billion run rate
Cognition, the company behind the autonomous coding agent Devin and the Windsurf IDE, said in a September 25 blog post that its annualized revenue run rate passed $1 billion. In May the figure was about $492 million, so it more than doubled in four months, and by early September it had already reached "close to $900 million." The company was founded in January 2024, Devin hit general availability at the end of 2024, and Windsurf was acquired in July 2025.
Growth rests on two legs: an autonomous agent that takes tickets, writes tests, and opens PRs, and an IDE built for human-agent collaboration. The capital moves came fast. A $1 billion Series D in May was led by Lux Capital, General Catalyst, and 8VC, with Founders Fund and Ribbit participating, at a $26 billion post-money valuation. On September 8 came a $2 billion Series E led by a16z and Accel at a $48 billion valuation. The Information reported the company expects $4 billion to $5 billion in annualized revenue by the end of 2026.
Named customers include GE Aerospace, Rivian, Rohlik, and Exa; TechCrunch also lists Mercedes-Benz, NASA, Goldman Sachs, and Citi. Pricing mixes seat subscriptions with usage-based Agent Compute Units. Cognition claims 89% of the code its own engineers commit is written by Devin. Costs are climbing in step. The Nvidia server clusters the company rents cost hundreds of millions of dollars a year, and total cash burn in 2026 could reach $800 million, per The Information as relayed by TechCrunch.
The comparison stings. Anysphere's Cursor had already reached a $4 billion annualized run rate by early June, up from $100 million in about 16 months, and expects to pass $6 billion by the end of 2026. SpaceX bought Anysphere in an all-stock deal valued at $60 billion that closed August 14, folding Cursor into a SpaceXAI division; SpaceX had earlier approached Cognition about an acquisition, and Cognition chose to stay independent. The run-rate trap deserves spelling out: it multiplies the most recent month's billings by twelve, and it is not booked revenue. Neither company discloses gross margin. The wider backdrop is about $1.9 trillion in AI infrastructure commitments this year against less than $1 trillion in recognized AI-related revenue.
— Cognition · The Machine Herald
🔗 Cognition · The Machine Herald
Akamai signed $11.6 billion for Anthropic's CPU workloads
Akamai Technologies (NASDAQ: AKAM) announced an $11.6 billion, seven-year commitment from Anthropic, a large expansion of the $1.8 billion seven-year commitment disclosed in May, when the customer was not named. The deal runs through two new project schedules under an existing master services agreement signed May 5, 2026. This is a CPU contract, not a GPU training contract: Anthropic is using Akamai Cloud's distributed infrastructure and software to support fast-growing CPU workloads.
There is room to grow. The commitment can rise by up to $9 billion more, for a potential total near $20 billion. Akamai also issued Anthropic warrants tied to non-voting convertible Series B preferred stock, converting to 7.7 million common shares, about 5% of Akamai's outstanding common stock, at a strike price of $111.33. About 2% vests with this $11.6 billion commitment; the remaining 3% vests in tranches over the seven years as the partnership grows, roughly 1% for every additional $3 billion in purchases.
On Akamai's side, total capital expenditure tied to the commitment is estimated at about $5.5 billion; capex in 2026 rises by about $1.7 billion to lock in supply-chain components including memory; and the 2026 revenue guidance is unchanged. Revenue from the buildout starts in 2027, reaches full contract run rate by the end of 2028, and then runs about $1.7 billion a year through the tail of the contract. For scale, Akamai has announced more than $2.8 billion in multi-year cloud infrastructure services commitments this year, and its Q2 cloud infrastructure services revenue was about $99 million, up 39% year over year. After the news, the stock rose more than 16% in after-hours trading. The company was founded in 1998 by MIT applied math professor Tom Leighton and Danny Lewin, starting from a 1995 question Tim Berners-Lee raised at MIT about how to move web content at scale, and it is best known for its CDN business.
— Akamai IR · Voice&Data
🔗 Akamai IR · Voice&Data
Alibaba's Zhenwu V900 chip and a 10-trillion-parameter Qwen roadmap
At the Apsara Conference in Hangzhou (September 22 to 24, themed "Intelligence goes beyond"), Alibaba unveiled the Zhenwu V900 from its semiconductor unit T-Head. It is the third generation of the Zhenwu line and handles both training and inference. CEO Wu Yongming called it "the strongest AI chip in China today."
The vendor's own specs: 216 GB of HBM (up from 144 GB on the previous M890), 1,200 GB/s of inter-chip interconnect bandwidth (up from 800 GB/s), native FP8 and FP4 support, about three times the performance of the M890, and scaling to 500,000 cards in a single cluster. Mass production and commercial availability are targeted for Q1 2027, with the next generation, J900, placed between Q3 2027 and Q3 2028. The chip ships as a supernode server alongside Alibaba's in-house ICN Switch, Panmai SmartNIC, and Zhenyue SSD controller. The right way to read it: Alibaba is selling a stack, not a die.
On models, Qwen 4 is in training, and Qwen 4.5 and Qwen 5 target 5 trillion to 10 trillion parameters, against roughly 2.4 trillion for the current flagship Qwen 3.8 Max. Alibaba also says it has made real progress on recursive self-improvement (RSI) aimed at superintelligence, which is a statement of direction rather than a benchmark number. On infrastructure, Alibaba Cloud aims for more than 20 GW of global data center capacity by 2032. The market read the day as a full-stack bet: Hong Kong shares rose 5.1% to a one-month high, and the US ADR rose about 3% to $118.50, still about 38% below its 52-week high of $192.67. Every performance number is vendor-reported and untested by third parties, and the chip is still a quarter away from shipping.
— Alibaba Cloud · Pandaily
Meta put an agent into glasses, a keychain, and a 100-gram headset
Meta Connect 2026 ran at the company's California headquarters and put the personal AI agent Muse at the center of the entire hardware line, with prices from $249 sport glasses up to a $1,299 headset.
Muse Charm is the strangest object of the set: a standalone handheld about the size of a headphone charging case, with a 2-inch touchscreen, built-in 5G, a microphone, a speaker, and a corner fingerprint sensor. It is meant to hang on a keychain or a bag, not to replace a phone, and to let you talk with Muse's real-time avatar. It ships in December at an undisclosed price. The Ray-Ban Meta Audio is the first model in the line without a camera: 43 grams, open-ear audio, 12 hours of battery (plus 48 more from the case), Clubmaster and Burbank styles, 23 frame and lens combinations, prescription support, $349, shipping October 13. Dropping the camera is a direct answer to privacy backlash, and it lets the glasses into offices and bars.
The Ray-Ban Meta Gen 3 keeps the camera: a 12MP sensor shooting 3K video, a six-microphone array that filters out more than 90% of ambient noise, Dolby Atmos spatial audio, 9 hours of battery, and a dedicated AI button, from $449. An FDA-cleared hearing aid feature unlocks through the Meta One subscription or a one-time $149 payment. Meta VR Glasses are the engineering step that matters: about 100 grams, roughly a fifth of a Quest 3, because the processor and battery moved into a small external puck. The Micro-OLED panels run 5K resolution at 37 ppd with Dolby Vision and Atmos, up to three hours of playback, backward compatibility with Quest games and Xbox cloud gaming, $1,299.99, arriving spring 2027.
On the software side, Meta Superintelligence Labs introduced the Muse Realtime Avatar system; on the glasses it starts hands-free and can act directly on what the wearer is looking at; Muse now has its own email address; it can operate a computer on Mac; and the connector list grew to Walmart, Best Buy, Sephora, Ulta, Wayfair, PayPal, Expedia, Instacart, Notion, GitHub, and Box, on top of the existing full Shopify catalog. By year's end Meta says it will have more than 100 AI glasses styles across Ray-Ban, Oakley, and Meta Glasses, including the $249 Meta Adventurer. The market backdrop that makes the hardware viable: Counterpoint Research counts a 263% year-over-year rise in global AI glasses shipments in the first half of 2026.
— Meta · Ubergizmo
Qoder turned the human-agent discussion into a product surface
Qoder, Alibaba's agentic coding platform, announced two collaboration features at the Apsara Conference, Projects and Discussion, both in beta, alongside custom agents and custom agent teams, opening first to Qoder Teams and Enterprise subscribers. The platform says it serves more than 6 million developers across desktop, mobile, IDE, a JetBrains plugin, CLI, an Agent SDK, and cloud agents.
The problem it names is not model quality. A task starts in a meeting, moves through group chat, and then gets reinterpreted by each developer inside their own agent session, so decisions, code changes, test results, and PR links scatter across separate sessions. In founder Ding Yu's framing, people should no longer act as the "mover" for agents. Projects is a shared space where people and agents join as members and work around the same codebase. Work is tracked as Issues with a description, an owner, a status, and linked discussions.
Discussion hangs off an Issue. Members add documents and context, reply to specific points, and have agents analyze material or compare options. The mechanism that carries the most weight is the Key Information marker: anyone can flag a message as a confirmed requirement, constraint, or decision, and those markers travel into execution instead of sinking into the scroll. Custom agents store instructions, model settings, permissions, and tools for repeatable work such as implementation planning and code review. A custom agent team has a lead agent that splits the work, member agents that explore and verify from different angles, and a human who judges the consolidated result.
The loop is the point: an Issue, a discussion, a flagged decision, an agent that picks it up, code and test results that return to the same discussion, a new requirement confirmed in place, and the agent continues. Importing from other issue trackers is on the roadmap.
— Qoder
🔗 Qoder
Two papers argue the harness belongs in the weights
Harness-Zero, from a Peking University team led by Haoran Ye and uploaded to arXiv on September 21, asks what happens if you stop shipping a harness and bake it into the model weights instead. The method is called agent-as-harness, and it replaces the code scaffold with a reviewing model: a harnessing agent (GPT-5.6 Sol in the paper) inspects the responses of a Qwen3.5-9B student model inside the target harness's action space, corrects them, and the resulting trajectories become supervised fine-tuning data.
Across three domains, the macro-average task success rate rose from 23.3% to 44.3%. The distilled model beat the same base model carrying a specialized harness at 41.7%, and the harness was gone at deployment time. The paper reports recovering an average 82.3% of 28 harness-induced behavior patterns, and on six settings agent-as-harness averaged 81.1% against 78.1% for the code-scaffold baseline. The limits are as useful as the gains. On USPTO Retrosynthesis the distilled model reached only 30.0%, while the deployed model carrying a meta-harness reached 38.0%; the authors' explanation is that procedural harness behavior is easier to internalize than deep domain knowledge. Trajectory collection is also 2.4 times slower than a native mini-SWE-agent trial run, though that cost lands in training rather than inference.
Agent-Editing World Model (AEWM, arXiv:2609.28416, submitted September 23, by Shuang Sun and colleagues, with Wayne Xin Zhao and Ji-Rong Wen among the authors) targets a different failure mode: task-state contamination, where unfounded assumptions and stale plans linger in an agent's history and distort later decisions. Existing language world models try to predict environment observations, but when real tool feedback is already available, that path buys little. AEWM instead models how reasoning and actions shape task progress: an Action Judge sorts decisions into Critical, Exploratory, and Noisy categories, a State Revision component edits noisy reasoning-action sequences, and EditAct changes the state that later decisions rest on during real execution rather than just commenting on it.
AEWM scores 70.5% macro-F1 on the Action Judge benchmark, 10.6 points above the strongest frontier baseline (59.9% for DeepSeek-V4-Pro). EditAct improves results by an average 3.2 to 6.7 points across six benchmarks and three agent backbones; AEWM-RFT, which fine-tunes on rejection-sampled verified trajectories, adds another 2.2 to 2.6 points over Self-RFT in three domains, and it needs no online AEWM guidance at inference.
— arXiv · AI Weekly
Top comments (0)