Search “Agents”

English translation is being prepared.

View story:English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

View story:English translation is being prepared.

Anthropic Publicly Submits S-1 Registration Statement for Proposed IPO on Nasdaq

Anthropic publicly filed its Form S-1 registration statement on October 26, planning to list on Nasdaq under the ticker ANT. The prospectus shows that its revenue grew from $41 million in 2023 to $16.3 billion in the first half of 2026, with $11.6 billion in the second quarter alone and operating profit of $974 million; about 48% of first-half revenue flowed through Bedrock, Vertex and Azure Marketplace, and it listed NVIDIA compute commitments of up to $120.5 billion. Subsequently, on October 27, World Trade Securities estimated that its cash reserves could approach $400 billion by the end of 2027; in the cover letter of its IPO filing, Anthropic reiterated that Claude will not display ads, and on October 28 it acknowledged that its life sciences business lacks physical laboratory experience. On October 28, Goldman Sachs secured the lead-left seat on the prospectus, and TechCrunch reported that public shareholders have almost no say over board nominees. On October 29, the prospectus disclosed four previously unreported Agent safety incidents. On October 30, Reuters reported that its IPO fundraising cap could reach $100 billion; Steve Eisman accused the prospectus of deliberately omitting third-quarter data, and Anthropic has not yet responded.

Mistral releases Large 4 weights on Hugging Face

On October 29, 2026, Mistral released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter native multimodal model that continues reinforcement learning on the basis of the preview version, with improved coding and defensive cybersecurity capabilities. According to data published by Mistral AI, the model scores 19.7% on FrontierCode 2.0 and 48% on the Artificial Analysis Coding Agent Index, roughly on par with GLM-5.3, but uses an average of 47% fewer tokens per task; on Terminal-Bench 4.0 it improves by 13 percentage points to 41%. The model supports three reasoning modes, uses a modified MIT Mistral Ecosystem Open-Weight license, and is priced at $0.70 per million input tokens, $0.07 for cached input, and $2.10 for output via API. On the same day, Unsloth released a 1-bit GGUF quantized version of the model, stating that it can run locally on a Mac Studio, retains 91% top-1 accuracy after quantization, and provides a download link and usage guide.

swyx says Gemini ranking first on AAv5 prompted him to retry Antigravity

On October 29, 2026, swyx reposted Artificial Analysis's evaluation results and said that Gemini 4 Carbon topped the Intelligence Index v5 with a score of 59, ranking second only to Claude Opus 5.5 on the Coding Agent Index. He said this result made him open Antigravity for the first time in a long while; he is currently gathering early evaluations and will report back afterward. swyx also noted that compared with the last time saranormous took the aie stage, the number of frontier labs has roughly doubled, which he considers a peculiar counterexample to the bitter lesson.

Quinn and Thorsten Ball explain how the Amp team works

On October 29, 2026, Quinn and Thorsten Ball explained how the Amp team works. According to a report by Shao Meng that day, the Amp team has about 21 people and holds only one meeting a week, with no quarterly goals, no mandatory PRs or code reviews, and pushes directly to main, averaging about 90 releases per day, with about 7.7 minutes from entering main to being available to everyone. The report says its quality relies on first having Agents provide end-to-end evidence, everyone trying the product, and feature flag controls, rather than on process. The two founders emphasized that this form depends on having few people, having worked together, and being motivated by revenue and product success, and is not a template that can be replicated by simply multiplying a large organization by AI.

Google releases the Gemini 4 Carbon model

On October 28, 2026, Google DeepMind released Gemini 4 Carbon, calling it its strongest model, ranking first with 59 points on the Artificial Analysis Intelligence Index v5.0, and scoring 88.2% on SEC-bench Pro, 54.8% on ExploitGym, and 71.0% on CWE-bench. The model targets long-horizon engineering tasks and was first opened to some cybersecurity defenders, then expanded through the Fairwind Program, and afterward made available to developers, enterprises, and consumers. On October 29, Arena.ai said Gemini 4 Carbon ranked second in Agent Arena with a net improvement of 16.2% and pushed the Pareto frontier forward; the same day, Yuchen Jin cited Artificial Analysis data to raise reservations, noting that the evaluation used the Antigravity CLI. On October 30, the Artificial Analysis Intelligence Index showed it in first place with 59 points, 1 point higher than Claude Opus 5.5. On October 31, Perplexity CEO Aravind Srinivas shared an early evaluation, saying it set a new SoTA on the WANDR benchmark.

Anthropic Executive Discusses Opus 5.5 Release and Claude Platform Engineering

On October 28, 2026, Anthropic engineering lead Katelyn Lesse and product lead Angela Jiang gave an interview on the Latent Space podcast, discussing the release of Opus 5.5. The interview covered adjustments to the model's writing style, pricing per dollar of intelligence, and API changes such as requiring thinking blocks to be retained across turns to prevent distillation. The two also described how Claude Managed Agents separate conversations and the harness from the sandbox that executes code, mount credentials separately through a proxy, and ensure that a sandbox failure does not take down the entire conversation, and explained that Dreaming is implemented as one managed agent inspecting another agent's history and writing back to memory. The team said that as safety work for model releases has grown heavier, recent efforts have shifted more toward safety, with progress on features such as skills and memory slowing relatively.

OpenAI launches GPT-6.1 Terra on ChatGPT Work and Codex

On October 27, 2026, OpenAI released and launched GPT-6.1 Terra, touted for its speed and efficiency, visual reasoning, and precise computer operation, with clearer and more natural writing as well, rolling it out that day to Plus, Pro, Enterprise, and Edu users in ChatGPT Work and Codex. On the same day, Arena announced that the model had entered Agent Arena and Code Arena; OpenRouter launched the model, saying it matches GPT-6 Astra on Agents' Last Exam at one-tenth the price per token; OpenClaw also announced that the model is now selectable. OpenAI co-founder Greg Brockman said Terra's computer use has crossed the threshold of a qualitative change and that using it is faster than operating a computer on his own. fal, meanwhile, used the fal Codex plugin to test turning Blender 3D content into Seedance 2.5 cinematic footage. On October 28, Arena announced that Terra ranked 14th in Code Arena: WebDev with a score of 1671, at a median cost of $0.40 per task, 28% lower than GPT-5.6 Terra. OpenAI's system card confirms that Terra is overall safer than older models, but its flag rates on items such as concealing uncertainty, circumventing restrictions, data exfiltration, and destructive operations

Claude engineering lead proposes giving models full data and tools, using adversarial review

In an interview, Claude engineering lead Felix Rieseberg proposed giving crash data, business tools, and computers to the model, letting the model decide for itself what to fix, rather than having a human act as a relay and pass along each item one by one. He suggested opening multiple threads aimed at the same output to specifically prove it wrong, then opening another thread to act as a judge and filter out the criticisms that hold up; rather than slicing off a small piece of data for each agent, it is better to give a panoramic view of the business and tools with clear documentation. He also mentioned that the half-life of Skills is extremely short, that both the official team and individuals recommend using them with caution, and that self-evaluation within the same session is still unreliable, so external signals should be sought as much as possible.

Claude engineering lead proposes giving models full data and tools, using adversarial review

In an interview, Claude engineering lead Felix Rieseberg proposed giving crash data, business tools, and computers to the model, letting the model decide for itself what to fix, rather than having a human act as a relay and pass along each item one by one. He suggested opening multiple threads aimed at the same output to specifically prove it wrong, then opening another thread to act as a judge and filter out the criticisms that hold up; rather than slicing off a small piece of data for each agent, it is better to give a panoramic view of the business and tools with clear documentation. He also mentioned that the half-life of Skills is extremely short, that both the official team and individuals recommend using them with caution, and that self-evaluation within the same session is still unreliable, so external signals should be sought as much as possible.

View story:Claude engineering lead proposes giving models full data and tools, using adversarial review

Itamar Friedman argues that engineering teams must build and own bounded L4 systems

On October 24, 2026, QodoAI founder Itamar Friedman put forward the view that engineering teams must build and own bounded L4 systems. He stated that having agents at every stage does not mean a software factory already exists; the goal of a factory is to autonomously complete the authorized lifecycle on defined classes of work, produce verifiable evidence, and provide a safe fallback when failures occur, forming a bounded L4 system. Borrowing from autonomous driving levels, he devised his own L0 to L5 framework, pointing out that L4 does not mean omnipotence—operational boundaries are themselves part of the capability. Teams can first push routine changes to mature internal services to L4, while still tightly controlling areas such as payment authorization. He emphasized System > Agents > LLMs, and that a team's advantage lies in building and owning the boundaries, evidence, permissions, and feedback, rather than stacking models or buying black boxes. Currently, this view is reported by a single source, and there are no subsequent factual developments.

DeepSeek integrates Harness with the web, App, and open platform teams

On October 16, 2026, DeepSeek established a product department, integrating its four teams—Harness, Web, App, and Platform—under the leadership of Cui Tianyi, who comes from a quantitative trading background, to oversee product, design, and R&D. Related reports noted that DeepSeek Harness was open-sourced under the MIT license on August 13, and by mid-October the repository had accumulated about 250,000 GitHub stars; the official website has posted job openings for two directions: AI product manager and Agent Harness. The reports also mentioned that Guo Daya, a core contributor to R1, left in March 2026 to join ByteDance Seed because he believed the Agent direction was not a high priority within DeepSeek. Currently, this integration is coordinated by Cui Tianyi, and there are no further developments with supporting evidence at this time.

Boris Cherny shares five practical principles for Claude Code

On October 22, 2026, Boris Cherny, the author of Claude Code, shared five practical agent principles for Claude Code in the fourth week of Stanford's CS146S course. He mentioned that he runs several thousand Claude agents simultaneously, that since last November nearly all of his team's code has been written by Claude, and that per-person output has increased by roughly 10 times. The five principles include: automating R&D bottlenecks (code review has already been automated, with approvals as the next step), building for the model six months from now, small teams being faster, scheduling agents in loops, and domain experts winning. He also mentioned that Claude Code's early revenue stayed flat for several consecutive months, and only after the release of Opus 4 did growth climb steeply. This talk is currently a single-source report, with no subsequent developments yet.

Nadella's shareholder letter proposes matching token costs to business outcomes

On October 22, 2026, Microsoft CEO Satya Nadella proposed building a frontier ecosystem in the 2026 annual shareholder letter, and put forward "cost-to-outcome" (the business outcomes obtained per unit of token spending) as a new software economics optimization metric. The letter also disclosed that Microsoft's fiscal 2026 revenue was $331.8 billion, up 18% year over year; Azure grew 41% and surpassed $100 billion for the first time; M365 Copilot paid seats exceeded 30 million, and GitHub Copilot users reached 50 million. The letter also stated that interaction is shifting from synchronous chat to asynchronous agents, with Harness separated from models; Foundry offers more than 11,000 models and has over 100,000 customers, and Microsoft Frontier Co. was established, dispatching 6,000 experts to work on-site. Currently, the content of this letter is reported by a single source, with no subsequent independent developments yet.

Jediah Katz proposes that Harness coding patches are being absorbed by models

On October 21, 2026, Jediah Katz, head of Grok Bot and Cursor Harness, offered the judgment that the part of Harness that helps models write code is being absorbed by models, and that fancy Harness no longer increases models' coding ability. He also stated that the three responsibilities of opening up new task boundaries, providing infrastructure and trust/security, and optimizing efficiency still cannot do without Harness. Citing UC Berkeley tests on SWE-bench Lite and Apple and EPFL tests on MLE-Bench, he said that score differences across different Harness are not significant, but per-task cost differences are obvious. He also predicted that within 6 months, 90% of Agent work will leave personal computers and run in the cloud. Currently this judgment is reported by a single source, with no subsequent factual developments yet.

QuestMobile releases Q3 AI application market report

QuestMobile releases the 2026 Q3 AI application market report. The report shows that in September, Doubao ranked first among AI-native Apps with 419 million MAU, while Qwen and DeepSeek ranked second and third with 158 million and 123 million respectively. In the office Agent ranking, Tencent WorkBuddy topped the list with 32.353 million MAU, while Baidu Dazi and Qwen Office ranked second and third with 9.77 million and 7.905 million respectively. In September, the AI office track reached 142 million MAU, up 50.6% year-on-year, and the PC client user scale of office Agents grew 512.2% year-on-year. Other reports say that under Analysys's desktop measurement, WorkBuddy's monthly visits increased from 20.97 million in June to 85.34 million in September.

DeepSeek releases DeepSeek-V4.1-Flash-1017

On October 17, 2026, DeepSeek released DeepSeek-V4.1-Flash-1017, which the company says delivers stronger benchmark performance, faster task completion, and shorter outputs, achieving a Pareto improvement in performance and cost; it is now available on the official website, App, and API, and the model also launched on OpenRouter the same day. The company also announced that the Search and Fetch APIs are open to everyone, priced at $0.75 per 1,000 calls per endpoint, with calls made through server-side tools in DeepSeek Harness, Claude Code, and Codex temporarily free, and billing starting at 16:00 UTC on October 31, 2026. On October 17, Thomas Wolf commented that DeepSeek called this a "minor update," yet Terminal-Bench 4.0 rose from 31.2 to 42.3. On October 19, Arena announced that the model, with a score of 1670, became the highest-ranked open-source model on Code Arena: WebDev. On October 20, OpenCode updated it to version 1017 in Go and Free; Unsloth released a GGUF that can run locally, with the 1-bit quantized version retaining about 88.2% accuracy and shrinking in size from 508GB to 250GB; Arena then announced that it topped the Agent Arena open-source model leaderboard.

English translation is being prepared.