Reasoning

OpenAI data suggests that advanced models hide their reasoning because developers optimize for capability and compute costs OpenAI researchers examined why advanced models are increasingly solving tasks without leaving a clear trace in their chains of thought. Looking at models from GPT-5 to GPT-6 Astra, they identified four factors: the base…

Meta has introduced Muse Ember, its smallest and most cost-efficient model. It is designed for routine AI-agent work: calling tools, handling small coding tasks, routing requests, extracting data and processing long documents. Meta says Muse Ember delivers more capability per dollar than Claude Haiku 5.5 and GPT-6 Luna at every reasoning level. In an Artificial Analysis test, a task costs about $0.14: the larger Muse Spark 1.3 scored nearly as much — 37 versus 36 — but costs $2.39. In Meta’s own tests, Muse Ember leads in most cases. But on real-world coding tasks, Haiku outperformed it on the FrontierCode 2.0-AA benchmark: 34.9% versus 27.2%.

Can AI agents do more than solve physics problems and actually carry out research? The authors introduce a test built from real research challenges contributed by physicists. AI agents must review previous work, plan an investigation, run calculations or experiments, and write up their findings, with checks for both correctness and scientific quality. The results suggest that today’s agents are held back less by missing facts than by shaky reasoning and a lack of the instincts experts develop through experience. In this review we look at how the test works, what the agents were able to do, and what still stands between them and credible discoveries in physics.

A new open benchmark has found that reasoning is the main bottleneck for AI in cloud operations. CloudOpsBench is an open benchmark for testing AI agents on common cloud DevOps tasks, from fixing services and deleting resources to restricting access, installing updates and closing projects. Version 1.0 includes 45 AWS tasks prepared by…

OpenAI says GPT-6 models perform better than their predecessors on two tests involving older programming languages and code migration. The results are striking on the COBOL benchmark: GPT-6.1 Sol reaches 86.6% accuracy at low reasoning, compared with 5.4% for GPT-5.6 Sol. But the tests measure narrow tasks, not whether the models can safely replace the people maintaining legacy systems.

Anthropic says it has cut the number of Claude reasoning tokens extracted in distillation campaigns by 10–100 times, using a mix of tighter access controls, classifiers and account enforcement. The result matters because reasoning traces can teach a smaller model how a stronger one reaches an answer, not just what answer to give. But Anthropic’s account also describes a costly, ongoing contest: the company says it has blocked tens of thousands of accounts, while acknowledging that attackers have not yet mounted their most persistent effort to defeat the new defenses.

Two German courts have treated Google’s AI-generated search answers as the company’s own statements, not neutral links to other people’s material. One ruling concerns false claims about a doctor; the other supplied the reasoning behind it. Together, they point to a legal risk that reaches beyond individual errors: if Google is responsible for what its AI says, the economics of AI search—and potentially publishers’ claims over the material used to produce those answers—could change.

A team of AI models can solve a problem none of its members can solve alone—but only if the models learn how to challenge and correct one another. A new framework called SAT trains fixed groups of models to develop reusable collaboration strategies, then applies those strategies to unfamiliar tasks. Its strongest results came on math and physics benchmarks, where a diverse team outperformed single-model and multi-agent baselines. On knowledge benchmarks, however, the team often generated a correct line of reasoning but struggled to identify it.

Gemini 4 Argon scored 97.5% on ARC-AGI-3 in a test setup that supports persistent reasoning and context compaction, putting Google’s model behind only GPT-6 Astra among the results reported here. The score brings Google closer to the current leaders than any of its models since Gemini 3.1 Pro in February. But Argon’s performance changes sharply under a more restrictive harness, making the conditions behind the headline number as important as the number itself.

What if an AI agent could do some of its thinking without putting every thought into words? The authors introduce a system that lets an agent reason in a compact internal format, then switch to plain text when it needs to use tools or communicate with the outside world. This approach aims to keep the agent’s thought process inspectable while making its work faster and less costly. In this review we look at how the system blends silent reasoning with tool use, and what its results suggest about building more capable AI agents.

OpenAI’s math results are beginning to look less like a story about one field and more like a test of how far AI can reach into theoretical science. A new study by Thompson and colleagues evaluated six OpenAI models on research problems published after those models’ training cutoffs. The problems were drawn from areas where researchers could assess reasoning without simulations or experiments, making the results a limited but useful check on whether models can do more than reproduce familiar answers.

Western mathematicians are debating whether AI-generated proofs weaken understanding or simply flood the field with claims to check. In China, mathematician Shing-Tung Yau and a group of leading Chinese mathematicians have formed an alliance across eight institutions, aiming to build a network of more than 1,000 researchers. The contrast is not just about whether to use AI: the alliance plans to develop specialized models, produce proofs that can be checked automatically and make AI reasoning transparent, even as Yau warns that China is behind the United States.

Anthropic says it has cut the theft of Claude’s reasoning traces by 90–99% using new access controls and automated detection. The company named Chinese firms Zhipu and Moonshot among those behind what it calls industrial-scale distillation attacks. The measures have already led to tens of thousands of account bans, but Anthropic says legitimate users have been caught in the process too.

Research: More AI agents spend tokens for barely better results Vals AI's benchmark shows cost per app versus score on Vibe Code Bench. Arrows point from each model's single agent to its team at the same reasoning effort. Teams cost far more but barely improve scores. | Image: Vals AI[ Anthropic’s tests suggest that adding agents often buys…

Voice AI still has not had its ChatGPT moment, according to leaders from Otter and another voice-AI company. At the HumanX conference last month, Wen argued that models can already listen while speaking; the harder step is fast reasoning, so replies arrive without long pauses. For Otter, whose meeting-notes service is also developing digital twins to represent people in meetings, the test is broader: voice systems must understand what was said, preserve the meaning and earn users’ trust.

Google appears to be testing Gemini 4 Carbon, a model that reports say matches Anthropic’s Opus 5.5 at programming. The more consequential signal is not the comparison, which remains unverified in the announcement, but Google’s emerging model hierarchy: Argon is positioned for advanced reasoning, while Flash handles speed and volume. Carbon may be a new Argon release—or a separate model—and Google has not said which.

Meta has proposed giving AI agents a separate reasoning loop to decide what to do next while a task is underway. The approach is meant to help agents spend additional inference compute where it is useful, without changing or fine-tuning the underlying model. In tests across four benchmarks and three models, the framework beat a version that made decisions in a single step in all 12 comparable comparisons at the highest tested budgets. The result points to a different way to scale agents: not just giving them more compute, but letting them decide how to use it.

Reflection has introduced Beam, a 501-billion-parameter text model that it says can match China’s GLM-5.2 on difficult reasoning benchmarks while using three to four times less compute at inference. The model is due to become open later this month, with its weights and full technical documentation. Those claims have not been independently verified, making Beam’s release less a settled challenge to closed labs than a test of whether Reflection can turn a striking efficiency pitch into a model others can reproduce and use.

OpenAI says it stopped a campaign to extract hidden reasoning from its models, but a follow-up test found the same attack still worked through Microsoft Azure. The episode exposes a gap between securing a company’s own API and securing the wider cloud platforms that serve its models. OpenAI says it blocked more than 15,000 accounts and fixed a vulnerability; researchers later extracted reasoning from every tested OpenAI model on Azure in a single attempt.

Anthropic’s Claude Sonnet 5.5 is designed to make each completed task cheaper, not each token. The company says the model works faster and needs fewer tool calls, while keeping API prices at $2 per million input tokens and $10 per million output tokens. Its benchmark scores now sit close to Opus 5.5 on several tests, though Anthropic still recommends Opus for work that needs sustained reasoning and judgment.

Anthropic engineer Jackson Kernion says Claude’s writing got worse for a counterintuitive reason: training made it better at mathematics, programming and reasoning, while also teaching it to produce explanations tuned to other AI models. That can yield technically dense prose that a language model handles easily but a person finds overloaded. The problem matters beyond Claude because it exposes a tradeoff in reinforcement learning: rewards that improve machine understanding can pull writing away from human-readable explanation.

AI agents can sound fluent yet still struggle to find their way through a jumble of tables, files, and databases. The authors propose EvoOntology, a layer that gives an agent a living map of what data exists, what it means, and which tools can help it use that data. Instead of relying on fixed instructions or forcing the agent to inspect every source from scratch, the system builds and continually improves this map as the agent works, adapting it to different tasks and ways of reasoning. This helps turn scattered, hard-to-navigate information into something an AI agent can actually explore and use. In this review we look at how EvoOntology is built, how it learns to refine its own understanding of data, and why this approach may make AI agents more capable across mixed data sources.

Alexis Ohanian, co-founder of Reddit, posted that the industry needs to come up with a new way of talking about data centers; Axios picked it up. Matt Higgins, co-founder and chief executive of the investment firm RSE Ventures, was more specific. He proposed renaming them outright: "fraud prevention centers," "clean energy efficiency centers," "precision agriculture centers." His reasoning is…

Nvidia has released Alpamayo 2 Super, an open-weights driving model that takes a 360-degree view from seven cameras and returns future trajectories, causal reasoning chains, high-level meta-actions, grounded answers about the scene and automatic reasoning labels from one input. The weights are on Hugging Face and the inference notebooks on GitHub, under OpenMDW-1.1, the Linux Foundation's…

The European Commission announced in Brussels on Monday that ChatGPT will be supervised as a very large search engine under the Digital Services Act. The reasoning is narrow and worth reading closely: the service was classified this way because it searches the internet and answers user queries, and because it clears the threshold of at least 45 million monthly users in the EU. Reddit and…

Salesforce and Nvidia have built Koa, a reasoning model trained for sales and customer support, and it will sit inside Agentforce next to the models Salesforce already pays for. Until now, when an Agentforce agent hit a long or multi-step reasoning task, Salesforce's AI gateway routed the prompt out to a frontier model — Claude or ChatGPT. Koa is the in-house answer to that, and the pitch is…

PrismML, a Caltech spinout with a $22.25 million seed round behind it, released Bonsai 2 27B on Thursday. The model is a compressed version of Alibaba's open Qwen3.8 27B that fits in 5.9 GB, nine to ten times less memory than the original needs, and scores 98% of the original's aggregate benchmark results. The number to watch is not 5.9 GB. It is 98 — up from 95% for the first Bonsai, which…

OpenAI has built a technique known as opaque recursion into Astra: instead of writing out its steps, the model runs a single query through a loop several times, leaving fewer readable traces and largely bypassing the chain-of-thought record that safety teams depend on. Buck Shlegeris, CEO of Redwood, wrote that he is extremely concerned about it. Ryan Greenblatt, chief scientist at Redwood…

The working vocabulary of AI gained a new center of gravity this year, and it is not a capability. OpenAI's Astra, released in September 2026, is known for early use of opaque recurrence: a method in which a model pushes the same prompt through its own internal layers over and over instead of reasoning step by step in language a person can read. OpenAI says Astra preserves a legible chain of…

OpenAI says its new model, GPT-6 Astra, is artificial general intelligence — by the company's own definition, "autonomous systems that outperform humans at most economically valuable work." The same model is the first OpenAI has ever placed in the "critical" category for cybersecurity capability, and the company has confirmed it shows a "substantial reduction in chain-of-thought…

Researchers at KAIST and Naver AI Lab report that the discrete steps a reasoning model writes out — pulling data, decomposing the problem, recalling a formula, computing — correspond to separable patterns inside the model's numeric representations. The separation is strongest in the middle layers, holds across three different models, and survives the case that matters most to anyone hoping to…

Researchers who profiled 13 large language models across more than 4 million responses report that the frontier systems have very nearly stopped forgetting: GPT-5 and Gemini 3 encode 95–98% of the facts tested against them. What those models cannot reliably do is get the facts back out. Asked directly, without reasoning, they fail to produce 26–34% of what already sits in their parameters.…

Conversations about AI agents give almost all their attention to models. Which LLM is stronger, whose reasoning is better, who writes more accurate code. But in real engineering work, an agent's success rarely rests on the model alone. There is another layer that assembles prompts, holds state, calls tools, and keeps the steps in order. The…

A near-cult of engineering complexity has grown up around modern LLM agent systems. Orchestrators, sub-agents, memory, skill libraries, tool calls — it all looks impressive, and it leaves the important question unanswered: what actually produces the gain in quality? The authors of HeavySkill: Heavy Thinking as the Inner Skill in Agentic Harness…

Agents do well on short tasks. Over a long horizon they are undone by memory, inconsistency and an inability to stick to a strategy. AI agents have gotten decent at problems that take a dozen actions, a couple of tool calls and an answer. Stretch the task to hundreds of steps — the length of real work — and it gets interesting. Early mistakes…

Multimodal LLMs have learned to recognize objects, but how do you give them visual imagination? A look at the Cognitive Supersensing idea. Over the past few years, multimodal LLMs (MLLMs) have learned to recognize objects, read captions, answer questions about an image and even give a decent account of what is happening in a frame. But they have…

As autonomous LLM agents take over human tasks — from negotiating with services to allocating resources inside companies — we have gotten used to judging them on solo benchmarks. We care how well a model writes code, answers questions, or plans. But in the real world they run into each other, compete for limited resources, and sometimes manufacture competition nobody needed. The paper…

We tend to assume reasoning models are stronger simply because they write longer chains of thought and burn more compute before answering. In Reasoning Models Generate Societies of Thought the authors offer a more interesting explanation: these models don't just think for longer, they start thinking differently — as though a small assembly of voices had formed inside them, putting questions to…

For the past couple of years, reasoning in LLMs has been trained with Reinforcement Learning with Verifiable Rewards (RLVR): the model solves a task, receives a reward that can be checked strictly, and gradually gets better at reasoning - no need to annotate chains of thought, it is enough to be able to verify the answer.

Today's AI agents can plan, call tools, run chains of actions, even operate inside a multi-agent system. But most of these setups share an awkward property: they are fundamentally reactive. An agent can answer well in the moment, yet after deployment it rarely changes its own habits, rarely revisits its strategies, and almost never modifies itself. When the environment shifts — new interfaces,…

In finance, the CFA (Chartered Financial Analyst) exams are a marathon run over three distances. Level I tests the fundamentals and whether you can keep the terminology straight. Level II puts you in front of cases where formulas and logic have to be applied in context. Level III asks for more than correct arithmetic: a coherent professional answer — how to build a portfolio, how to assess…

Today's LLMs can do a lot: explain hard topics, write code, hold a long thread of reasoning. But science is not just knowing the answers. It is a research cycle: work through the literature, come up with a hypothesis, test it with an experiment, then interpret the results honestly and adjust the plan. And this is where the field has long lacked a shared language and a shared yardstick: what…

Over the past couple of years, large reasoning models (LRMs) have gotten noticeably better at olympiad math. On problems at the level of AIME (the American Invitational Mathematics Examination), one long reasoning pass is usually enough: the model writes out a chain of thought, checks the answer, sometimes takes a couple of runs at it — and lands it. IMO (the International Mathematical…

When we ask a model to reason, it reasons in words if the medium is text, or over a static scene if the medium is an image. The world, though, is not static: objects move, and the rules often only become visible in how those objects behave over time. The authors propose video generation as a general-purpose channel for reasoning. Text can be written directly into the frames, visual hypotheses…

Enterprise data tends to sprawl across email, reports, databases and code repositories. Answering a hard question usually takes not one fact but many, plus the ability to synthesize hundreds of sources with checkable citations and a line of reasoning someone can follow. Ordinary agents and classic RAG systems are often a poor fit here: they give shallow answers, they are hard to steer once…

LLM agents can reason, but reasoning is not enough to solve real tasks. An agent also has to call outside tools, work through long scenarios and stay autonomous across dozens of steps. Rigid pipelines with fixed modes get in the way of that, and so do the classic approaches like ReAct and Plan-and-Solve. They impose the same action loop every time and work well on tasks that take two or three…

Put several models in a multi-agent system on the same question and they will argue a little, correct each other, and settle on a compromise that is not always right. Language is what lets them do it, and it is also the bottleneck: it is sequential, often ambiguous, and rarely a faithful record of the reasoning behind it. A new paper argues for working above the level of words and giving…

Large language models do well on short reasoning and coding benchmarks. But real work stretches over dozens or hundreds of steps, demands switching between applications, careful context management and the ability to catch your own mistakes. The core problem is that at test time most agents stay static: they accumulate no experience and get no better from one attempt to the next. The authors of…

Over the past two years one thing has become clear: many applications built on large language models learn better through careful work with the context than through fine-tuning weights. Into the context go system instructions, reasoning steps, examples, domain rules, facts, even hints on how to use tools. It is transparent, portable, and it works at runtime. On top of that, progress on long…

Large reasoning models (LRMs) today don't just answer — they unfold long chains of thought. That lets them handle harder problems, but it creates a new one: how do you describe the structure of that reasoning, and how close does it come to human thinking? The researchers propose borrowing a framework well tested in cognitive science — Schoenfeld's episode theory, built originally to analyze…

Reasoning tasks are a sore spot for many AI systems, even ones with solid factual knowledge. A new paper shows that reinforcement learning (RL) does more than push accuracy up — it rebuilds the model's internal logic into a hierarchy that runs from low-level execution to high-level planning. That explains where those aha moments come from. More usefully, it explains why the standard algorithms…