The Financial Times reports that Anthropic's third-quarter revenue reached $16.8 billion, up nearly 50% quarter over quarter, with an adjusted operating margin of 11%, excluding stock-based compensation costs. Modal announced that its annualized revenue run rate surpassed $1 billion and that it raised $750 million in Series D funding. In AA's inference energy-efficiency tests, the MI355X outperformed the B200 by roughly 15–40% on V4 Pro, but delivered only one-third of the B200's throughput per megawatt on Kimi K3. AA-Robotics connected models to a real robotic arm, but the tested models were still, overall, about 100 times slower and 10 times more expensive than humans.
Model companies and compute platforms are both expanding, but revenue scale, profitability, and hardware efficiency are distinct measures.
The Financial Times reports that Anthropic told investors its third-quarter revenue reached $16.8 billion, up nearly 50% quarter over quarter. During the quarter, each $1 spent on compute generated $1.70 in revenue. Its adjusted operating margin was 11%, but that measure excludes stock-based compensation costs.
Read original →Modal announced that its annualized revenue run rate surpassed $1 billion. The company also completed a $750 million Series D led by Accel, at a valuation of $15.75 billion. Growth has also been rapid: Modal says usage has tripled since its Series C and grown more than 15-fold year over year.
Read original →AA used AA-AgentPerf to replay agentic coding sessions and test NVIDIA and AMD hardware. Running DeepSeek V4.1 Flash at 100 tokens/s per user, the GB300 achieved throughput of up to 261 million tokens/s per megawatt. The MI355X outperformed the B200 by roughly 15–40% on V4 Pro, but on Kimi K3 it delivered about 8 million tokens/s per megawatt, below the B200's 24 million.
Read original →🔍 Analysis: Modal's usage has tripled since its Series C, reflecting expanding demand. Anthropic's $1.70 in revenue per $1 spent on compute shifts the question toward whether growth translates into profit, but that ratio is not a profit margin, and its 11% adjusted operating margin also excludes stock-based compensation costs. AA's tests further show that there is no hardware winner independent of the model: the MI355X leads the B200 on V4 Pro but falls well behind on Kimi K3. For businesses, revenue growth and compute procurement require separate accounting, and hardware selection should be based on the models and workloads actually being run.
Nemotron and Ling put the focus on specific tasks and deployment constraints when choosing models.
NVIDIA released the 120-billion-parameter Nemotron 3.5 Super VL and the on-device model Bolt 74B-A9B. Super VL can recognize text and charts in web pages and PDFs, with open weights, datasets, and training recipes that let enterprises conduct their own post-training. Bolt uses native 2-bit quantization and can run in 24GB of memory. Both models are available now. Bolt E4B, a 4GB version, is scheduled for release on October 23.
Read original →Ant open-sourced Ling-3.1-flash, with about 560 billion total parameters and roughly 25 billion activated per token. The model supports a context window of up to 1 million tokens and focuses on interface and prototype generation for real-world work. Ant says it can turn rough requirements into web pages, mobile interfaces, or interactive prototypes for team review and iteration.
Read original →🔍 Analysis: Nemotron 3.5 Super VL targets web page and PDF understanding, while Ling-3.1-flash focuses on interface and prototype generation. They address different stages of a workflow and are not interchangeable based on parameter count alone. Ling has about 560 billion total parameters, with roughly 25 billion activated per token, while Bolt specifies a 24GB memory requirement. Active parameter counts and memory requirements are different deployment metrics. Developers choosing models need to distinguish task fit from resource constraints: sparse activation does not automatically mean a model can run on-device.
Minos provides scientific simulations for training, while AA-Robotics tests the efficiency costs of real-world action.
Leash Bio released Minos, a sandbox for scientific agents, alongside a preprint, blog post, and open dataset. It transforms protein–small-molecule interaction data collected by the team over the past several years into tens of thousands of closed-loop experimental simulations. These simulations support verifiable scientific reasoning by agents and can be used for reinforcement learning on scientific tasks.
Read original →Artificial Analysis introduced the AA-Robotics v0.1 research preview to evaluate frontier language models' ability to understand, plan, and act in the physical world. Models control a six-axis robotic arm over HTTP and use four cameras. GPT-6 Astra scored 69 and GPT-6.1 Sol scored 65, but the tested models were overall about 100 times slower and 10 times more expensive than humans.
Read original →🔍 Analysis: Minos turns protein and small-molecule data into tens of thousands of closed-loop experimental simulations, while AA-Robotics lets models control a six-axis robotic arm over HTTP. Both move evaluation beyond text responses to the outcomes of actions, but one provides training feedback in simulation and the other measures performance on real equipment. In AA-Robotics, models were overall about 100 times slower and 10 times more expensive than humans, showing that task completion and practical efficiency remain separate hurdles. For agent development, verifiable training environments need to be complemented by evaluation criteria that account for time and cost.
Grok Voice 3.0 tops the voice interaction ranking with a score of 86.9, with input audio priced at $4.80 per hour.
SpaceXAI released Grok Voice 3.0, a speech-to-speech model focused on handling customer service calls. Artificial Analysis gave it a Voice Interaction Index score of 86.9, placing it first, above GPT-Live-1 Astra's 83.2. Its time to first audio was 0.81 seconds. Input audio costs $0.08 per minute, or $4.80 per hour.
Read original →A SentinelOne researcher says OpenAI's cybersecurity capabilities can cut comparable reverse-engineering work from four to six weeks to about eight hours.
SentinelOne's Gregor Stewart says OpenAI's cybersecurity capabilities can complete human-level reverse-engineering work in about 8 hours. Top researchers need 4 to 6 weeks for comparable work. He believes this enables security teams to investigate more malware, connect related threats, and gain a clearer picture of how attacks evolve.
Read original →