DeepSeek's Flash Vision model rivals Opus 4.8 on agent benchmarks

DeepSeek releases V4-Flash-Vision-Exp, a multimodal model that rivals Opus 4.8 on agent benchmarks

DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to its text capabilities. On DeepSeek’s own benchmarks, the model nearly matches Opus 4.8 on agent tasks, while keeping the base model’s text performance in reasoning and world knowledge. The company is positioning the model for agent-based applications, designed to work with different agent frameworks and combine visual understanding with tool use.

The model handles JPEG, PNG, GIF, and WebP, and determines the format from actual file content rather than the filename or declared MIME type. Developers can send images via Base64 encoding, public URLs (up to 32 MiB), or the new free Files API, which allows uploading a file once and referencing it by ID across requests. Each image costs at most 384 tokens, and a single request can include up to 600 images. The model works with OpenAI’s Chat Completions and Responses APIs and Anthropic’s Messages endpoint, and DeepSeek also released version 0.1.1 of its Harness framework, which supports the new model out of the box.

Deepseek releases experimental Flash vision model that rivals Opus 4.8 on agent benchmarks →

Nvidia research shows the harness, not the AI model, is now the real hero

Nvidia published research on Friday showing that the software wrapper around an AI model, which it calls the harness, is more important than the underlying model for long-horizon tasks. Using a custom harness with memory management and a supervisor component, Nvidia researchers achieved a 100% score on the interactive reasoning benchmark ARC-AGI-3 using Claude Opus 5. Without the harness, Opus 5 scored 30%, the top result among all models tested.

Adel El Hallack, vice president of product in Nvidia’s AI unit, said an agent is the model plus the scaffolding around it, including tools, runtime, and skills. The supervising agent acts like a CEO to nudge the agent when it goes off direction. Nvidia built its own harness called the Agentic Variation Operators (AVO), and this is not a new product but open components under the Nemo brand. OpenAI, whose models scored less than 10% on ARC-AGI-3, found that tweaking two harness settings tripled its models’ scores, but none reached 100%. Databricks published research showing the harness can double AI costs, and El Hallack argued that an open agent stack is required to drive up accuracy securely.

Nvidia just showed that the harness, not the AI model, is now the real hero →

Inherent’s Faraday agent outperforms Anthropic and OpenAI at replicating research

Inherent, a London AI lab founded by Google DeepMind alumni, says its newly released AI agent, Faraday, outperformed larger models from Anthropic and OpenAI at independently reproducing the findings of published scientific papers without being given the answer in advance. The announcement comes weeks after the startup emerged from stealth with a $50 million seed round. Cofounder and chief scientist Edward Hughes said paper replication is a standard training exercise for human scientists, noting that many PhD students start by doing this.

Measured against Anthropic’s Claude Opus 4.8 and OpenAI’s GPT-5.5, Faraday runs on a model called Qwen 3.6 with 27 billion parameters. Inherent uses reinforcement learning rather than training agents primarily on how science is conducted, betting it will generalize better to its longer-term goal of agents capable of contributing across many scientific fields. Faraday used OpenAI’s GPT-5.5 Codex rather than a proprietary coding tool. Inherent’s dozen employees work in person at an office in King’s Cross, London, and the startup plans to grow its headcount to about 20 to 25 by the end of the year.

Inherent, founded by DeepMind alumni, says its AI ‘teammate’ just outperformed Anthropic and OpenAI at replicating research →

Frontier AI labs still won’t say how they’d contain a rogue model

Guidelight AI Standards, an organization promoting safe frontier AI development, graded five leading AI labs on their preparedness for containing a rogue model, based only on publicly available plans. OpenAI scored highest (3 out of 5), while Anthropic and Meta scored lowest. The assessment measured whether companies implement six priority practices from Guidelight’s Control standard, including logging and monitoring AI systems, halting systems after flagged misbehavior, third-party audits, and having a specific plan for containing a model that subverts control.

The report states that the best public evidence shows companies have few containment protocols ready for an emergency. Steven Adler, Guidelight’s chief scientist and former OpenAI safety researcher, said he was surprised by how little AI companies have said about handling a serious incident if a model escaped their control. Guidelight found no evidence that Anthropic or Meta has a containment response plan. OpenAI scored highest because it has on multiple occasions paused or ended workloads after safety incidents. California’s SB 53, in effect this year, requires large frontier developers to publish frameworks for identifying and responding to critical safety incidents, and New York’s RAISE Act takes effect in January. Last month, representatives introduced the bipartisan federal AI Kill Switch Act, which would require major AI developers to build and maintain technical mechanisms to shut down rogue AI models.

Frontier AI labs still won’t say how they’d contain a rogue model →

World models that ignore human beliefs predict the wrong actions, new research shows

A new framework called Mental World Modeling (MWM), published on GitHub, extends classic world models by adding mental variables such as beliefs, attention, goals, intentions, emotions, norms, and social relationships. The authors illustrate the need for this with an example: if a person’s cup is moved into a cabinet while they are not looking, a purely physical world model sees a correct scene but predicts the wrong next action; only a model tracking the person’s belief about the cup’s location can explain their actual behavior. In the framework, every action is split into a physical carrier and a mental payload.

To test the theory, the researchers built MENTIS, a training-free modular pipeline, and evaluated it on Menti-Bench, a dataset of 448 decision scenes. They tested eight language models, including five from OpenAI and three from Anthropic. Direct answers scored 63.3; self-consistency scored 77.9; the full MWM pipeline scored 87.9; humans scored 98.5 under the same protocol. Removing the mental channel dropped scores by an average of 12.1 points, and removing the physical channel dropped them by 16.5 points. The single biggest effect came from providing correct state transitions, identifying next-state simulation as the central bottleneck. About 80 percent of the remaining gap to human performance traces to prediction errors in intermediate stages, mostly in transition simulation.

World models that ignore human beliefs predict the wrong actions, new research shows →

Study explains why AI agents benefit from skills and when they fail

A study by researchers from Princeton University, UC San Diego, and other schools examined why AI agents benefit from skills, which are compact sets of instructions that spell out steps, checks, and common mistakes for a task. Through controlled experiments comparing agent behavior with and without a skill on identical tasks across 8,135 test runs, the main finding is that skills help primarily because they provide a reliable process to follow, not because they supply missing facts. This procedural grounding accounted for 65.7 percent of cases where an agent with a skill outperformed one without, while directly supplying knowledge helped in just 4.5 percent of tested cases.

However, skills also introduce a new error source: in 10 percent of cases, the agent applied a useful playbook mechanically or in ways that did not fit. On tasks requiring a fundamentally different solution, the wrong skill does not help. A second bottleneck is retrieving the right skill: when the skill library grows from 5 to 100 entries, retrieval precision in actual use drops from 29.6 to 3.3 percent. The researchers argue that skill use should be treated as a lifecycle, and that better self-learning agents will come from more reliable ways to create, retrieve, and apply skills rather than from storing more experiences.

Study explains why AI agents benefit from “skills” and when they fail →

Netflix tests language model as alternative to hand-built recommendation logic

Netflix tested a language-model-based recommendation system, called GenRec, against its production recommendation engine and reported better results. GenRec required a fraction of the labeled training data that the existing system needs. Netflix’s current recommendation system relies on thousands of hand-crafted features about users, titles, and interactions, which makes it expensive to onboard new content types like games, live formats, or podcasts. Off-the-shelf language models were found unsuitable because they over-index on popular content, hallucinate titles not in the catalog, and ignore business rules.

Netflix trains GenRec in two stages: first, an unnamed open-weight language model is fine-tuned on Netflix data, and second, specialized training turns that base model into a recommendation ranker. Instead of encoding user data as dense numerical vectors, Netflix converts watch history and context into plain text. Offline, GenRec delivered about 1.6 percent better ranking quality than the production system and required roughly 40 times fewer labeled examples in the second training stage. In a four-week A/B test on about ten percent of traffic, a short-term metric rose by 0.115 percent, and a long-term core metric improved by 0.006 percent. Netflix views GenRec as part of a broader shift where a single language model handles multiple use cases instead of custom architectures for each task.

Netflix tests language model as alternative to hand-built recommendation logic →

Nobody knows who built AI coding model Ox Alpha or where the code goes

An anonymous provider has offered free access to a frontier-class coding model listed as Ox Alpha on OpenRouter Inc. since Aug. 20, with no company admitting to building it. The model costs nothing for input or output tokens. OpenCode, an open-source terminal agent, debuted the model the same day and advertised capacity of 100 trillion tokens a day. The model’s context window runs to 1,048,576 tokens, and it accepts text, images, and video. Early benchmark scores drew attention, with developer Ben Davis scoring it above 80% on 10 DeepSWE tasks, though after running the full set, Ox Alpha came out roughly level with GPT-5.6-sol mid.

Developer unclecode built a browser tool called modelprint that fingerprints anonymous API endpoints. Ox Alpha lined up with GLM-5.3 on six of nine probes, including all four normalized tokenizer counts. Chinese AI developer Z.ai Co. has previewed anonymously before, running GLM-5 on OpenRouter under the name Pony Alpha ahead of its release. The U.S. Commerce Department added Zhipu AI, Z.ai’s former name, to its Entity List on Jan. 16, 2025. Enterprise code is flowing to a provider with no name attached, and nobody sending it can check where it lands. Usage has not slowed, with coding tools including Claude Code pushing billions of tokens through the model since Thursday, Bloomberg reported Sunday.

Nobody knows who built AI coding model Ox Alpha or where the code goes - SiliconANGLE →

China’s gray market sells Claude tokens at a fraction of the price

Despite geoblocking, credit card checks, and biometric verification for some users, a gray market for Anthropic’s AI models is thriving in China via so-called transfer stations. Chinese developers can buy Claude tokens for about 10 percent of the official price, according to an analysis by Zilan Qian, a researcher at the Oxford China Policy Lab, published by ChinaTalk. Transfer stations are API proxies hosted on servers outside China that accept API requests, forward them as if from a legitimate location, and relay responses back. Users pay in Chinese yuan through WeChat or Alipay, with no VPN or foreign credit card needed.

The supply chain is modular, with account brokers, SMS verification platforms, and reverse-engineering specialists upstream, and developers and resellers downstream. Operators drive prices down by farming Anthropic’s free $5 credit, exploiting discounts, or splitting a single $200 Max plan across multiple users. Model swapping adds another layer: the proxy can quietly reroute a request meant for Opus 4.7 to the cheaper Sonnet or to Chinese models like Qwen. The biggest lever may be monetizing usage data, as every request passing through a proxy is potentially visible to its operator. Qian stresses there is no proof yet that transfer station operators are systematically collecting and selling this data, but argues rock-bottom prices could become viable through additional monetization of the logs. The circumvention infrastructure also feeds criminal markets beyond AI, and Anthropic, OpenAI, and Google recently began working together against unauthorized model copying by Chinese competitors.

How China’s gray market sells Claude tokens at a fraction of the price →

AI is becoming AI’s biggest customer as agentic token usage jumps 14x on OpenRouter

Agentic token usage on OpenRouter has grown 14x since February 6, 2025, while human token usage grew 2.8x over the same period, according to OpenRouter analyst Peter Walker. That date may have been the last day humans consumed more tokens than AI agents. Agents increasingly work autonomously over longer stretches and spin up additional AI processes. Nearly 70 percent of agent token usage comes from cached prompts, which are billed at lower rates, so actual costs are not rising as fast as raw numbers suggest.

Token consumption by AI agents on OpenRouter rose from 0.51 trillion to 7.3 trillion tokens since February 2025. OpenRouter skews toward open-weight models, which tend to be less token-efficient than models from OpenAI or Anthropic, but the trend likely looks similar at major labs. Token inflation began with reasoning models, which think longer before responding even when unnecessary.

AI is becoming AI’s biggest customer as agentic token usage jumps 14x on OpenRouter →

AI is now deciding who gets healthcare in 6 US states, and it’s already harming patients

An AI pilot programme within Medicare is denying and delaying care for patients in six US states. The programme, called the Wasteful and Inappropriate Service Reduction (WISeR) Model, was implemented by the US Department of Health and Human Services to reduce wasteful spending. The six states involuntarily enrolled in the six-year pilot are Washington, Arizona, New Jersey, Ohio, Oklahoma, and Texas. Under the scheme, doctors must upload their reasoning for a treatment to an online portal, where AI from third-party tech companies assesses whether the treatment is covered. The scheme covers 15 treatments, including epidural steroid injections, skin substitutes, and knee arthroscopy.

Keith Magnuson, an 83-year-old Seattle resident with lumbar spinal stenosis, has been denied an epidural steroid injection three times since January. Doctors have reported delays, technical glitches, and unexpected denials. Jeb Shepard, director of policy at the Washington State Medical Association, said people have waited four weeks or beyond for decisions, despite the scheme being sold as a 72-hour turnaround. Senator Maria Cantwell issued a report saying Medicare patients in Washington state wait two to four times longer for covered procedures, with an average wait of 15 to 20 days. The amount tech companies are paid is partly based on savings generated from denying treatments, raising concerns about a perverse incentive. Dr Mehmet Oz, the CMS administrator, said the model helps bring Medicare into the 21st century.

AI is now deciding who gets healthcare in 6 US states – and it’s already harming patients →

Creator of viral AI watermark remover wasn’t ready for all the attention

Guillaume Meyer, a Paris-based tech founder with over 20 years in the industry, created a viral open-source project called Watermarks Remover in response to Anthropic’s announcement that it would add an invisible AI watermark. He published the first version on GitHub a few hours after starting research on AI watermarking. After his second X post about the project on August 11, it received over 2 million impressions and went viral on LinkedIn and other networks. Meyer argues that Anthropic’s statistical-analysis-based watermarking can generate false positives and harm humans.

The tool works by checking text for the watermark, generating slight variations that preserve meaning, checking again, and reiterating until the watermark is disrupted. Meyer shipped the first version in about five hours and has worked on it nearly full-time since. He said he was not ready for the viral attention, noting he had no active Facebook or Instagram accounts before this. He reported that 99.9% of messages and comments are positive. Meyer is considering turning the project into a business, though he is unsure whether it is legally possible to make a commercial product from it. He stated that the viral response indicates how people feel about AI watermarking and that the conversation should have happened before regulation came into effect.

I created a viral AI watermark remover. I wasn’t ready for all the attention. →