Google has released Gemini 4 Argon and begun a phased rollout to trusted cyber defenders. Its output limit rises from 64K to 1M tokens, with introductory prices of $2 and $10 per million input and output tokens, respectively.
Google announced Gemini 4 Argon and has begun rolling it out to a group of trusted cyber defenders through the Fairwind Program, with access to expand in phases. According to the company, Argon is designed for long-horizon reasoning across software engineering, enterprise knowledge work such as legal and finance, and cybersecurity defense, while its maximum output has increased from 64K tokens to 1M tokens. Introductory pricing will be $2 per million input tokens and $10 per million output tokens, with cached input tokens discounted by 95% from the input token price. Before opening access to developers, enterprises, and consumers, Google will collect feedback from early testers, iterate on guardrails, and participate in the U.S. government’s voluntary process for pre-release model access.
According to Google’s published results, Argon scored 77.9% on DeepSWE v1.1, ranked first on AutomationBench with 51.3%, scored 91.7% on LVBench, and tied for first on CWE-bench v1 with 68%. It also posted leading results on the Vals Index, Vals Finance Agent v2, and Harvey’s Legal Agent Benchmark. Google said Wiz has used the model through Scan for Good and found a critical vulnerability missed by previous frontier models that exposed sensitive personal information in healthcare software used by hospitals worldwide. The version for trusted defenders and Google’s internal teams will be released without cyber guardrails. Before broader availability, Google will continue strengthening defenses against cyber and CBRN misuse, indirect prompt injection, and misalignment, with the measures tested by internal and external red teams.
哔哩哔哩 releases and open-sources the Index-Translate multilingual translation model series
要闻
Bilibili has released and open-sourced the Index-Translate multilingual translation model series, including 2B, 9B, and a 35B-A3B preview; in its official evaluation, the preview scored 0.8794 on FLORES-200.
The Index-Translate multilingual translation model series has been released and open-sourced by Bilibili. The series is built on the Qwen 3.5 2B, 9B, and 35A3B base models, with the 35B version still in training and offered as a preview. According to the official description, the models combine multilingual translation, instruction following, and meme understanding, and can preserve forms of address, adjust conversational tone, or return only the translation when instructed. Training has three stages—mid-train, post-train, and model-merge—first mixing pre-training replay, multilingual, parallel, and pivot corpora, then separately training general translation, instruction-following translation, and meme translation experts, and finally consolidating their capabilities through linear interpolation, MOPD, and Mix-RL.
According to the officially published results, the Index-Translate-35B preview scored 0.8794 on FLORES-200, 0.8336 on instTrans IFscore, 0.7405 on MEME, and 0.8168 on low-resource translation. FLORES_minor_pair contains 104,000 inputs spanning 1,040 translation directions across 62 languages; on this evaluation, the 35B-A3B preview recorded COMET-22/XCOMET-XXL scores of 0.8168/0.7164 and an off-target rate of 2.4%. instTrans_minor contains 2,793 tasks, on which the 9B model recorded an IFscore of 0.7725, an off-target rate of 3.47%, and a translation quality score of 0.5222. The accompanying Index-Echo supports S2S translation and dubbing in 6 languages, with the 9B version scoring 0.857 on MT judge and the 2B version recording a TS st_MAE of 0.395; Index-Homura allows a target syllable-ratio range to be set for dubbing translation.
蚂蚁百灵 releases Ling-3.1-flash with a two-week free trial
要闻
Ant Group’s Bailing has released Ling-3.1-flash, with about 560B total parameters, about 25B activated per Token, and a 1M context-window limit. It offers a two-week free trial with a 256K service length, followed by planned paid access and open-sourcing.
Ant Group’s Bailing has released Ling-3.1-flash and opened a two-week free trial. The model has about 560B total parameters, activates about 25B per Token, and supports a context-window limit of 1M. According to the company, the service length is 256K during the trial; after the trial, access will become paid, with 1M context support and open-sourcing planned at that point. The company said the model performed strongly on the GDPVal-AA V2.1 office-work benchmark and the HealthBench Professional medical benchmark, and completed long-horizon tasks including implementing a Lua compiler from scratch. Users can access the model through Ling Studio and Ant Group Digital Technologies’ large-model service platform. Vercel separately announced that Ling 3.1 Flash is available through AI Gateway.
MiniMax launches the M Plan subscription and ends new purchases of Token Plan
开发生态
MiniMax has launched M Plan in three tiers—Go, Explore, and Build—priced at RMB 49, RMB 119, and RMB 469 per month, with the same text quotas as Token Plan. New Token Plan purchases have stopped, and renewal and upgrade rules for existing users have changed.
MiniMax has launched three M Plan subscriptions—Go, Explore, and Build—and stopped new Token Plan purchases. M Plan carries over and expands Token Plan’s subscription capabilities, while keeping prices and text quotas unchanged across the three tiers. Text, image, audio, and other capabilities share the plan quota; Explore and Build include access to the H3 video model, while Go does not include video. The company has also introduced a 50% discount for the first month of monthly subscriptions, covering new purchases and upgrades through October 14. Annual subscriptions are excluded. The standard monthly prices for Go, Explore, and Build are RMB 49, RMB 119, and RMB 469, respectively, with automatic renewal at the standard price from the second month.
Except for MiniMax Code iOS subscriptions, existing Token Plan users can retain their current plans and benefits by keeping automatic renewal active. Users who cancel or interrupt renewal cannot repurchase the plan, and automatic renewal cannot be reactivated once disabled. Subscribers can upgrade to M Plan during their subscription period, with the remaining value converted on a time-based basis and applied as a credit. After upgrading, they cannot return to Token Plan, and additional benefits such as historical discounts, the ∞ limit, and the 150% weekly quota will expire. Subscriptions purchased through iOS cannot be renewed after expiration; according to the company, affected users will receive points.
Space Bunny Alpha anonymous testing extended through October 5
开发生态
The anonymous model Space Bunny Alpha will remain in testing through October 5, 2026. Its OpenRouter page lists a price of zero, a 1,000,000-token context window, and a maximum output of 524,288 tokens.
Space Bunny Alpha’s anonymous testing period has been extended through October 5, 2026, and its OpenRouter page currently lists no charge for either prompt or completion tokens. The model was released on September 23, 2026, and is developed and operated by a third-party provider that chose to remain anonymous during the preview. OpenRouter only forwards requests directly to the sole hosting provider and is not the model’s developer, owner, or provider. According to the page, the third-party provider may retain prompts and completions, but they are not used for training; other uses are governed by the Stealth Model Terms.
The model has a 1,000,000-token context window and supports up to 524,288 completion tokens. It accepts text, images, and video as input and returns text. It supports function calling through tools and tool_choice, as well as JSON output through response_format, but does not enforce a JSON schema. The provider states that the model has coding capabilities and allows its reasoning effort to be adjusted.
Anthropic’s developer account ClaudeDevs announced the launch of Claude.dev, a site for developers using Claude that offers engineering articles and guides for Claude Code and the API. The existing documentation site remains unchanged.
Anthropic’s developer account ClaudeDevs announced that Claude.dev is now live as a new site for people building with Claude. According to the account, the site includes in-depth engineering articles, usage guides for Claude Code and the API, and development experience shared by the Claude team. The existing documentation remains at code.claude.com/docs, and selecting DOCS in the homepage navigation takes users directly to that address. The site also includes hidden features, one of which is a terminal mode: selecting TERMINAL in the navigation switches the entire site to a command-line interface.
Pi v0.99.0 adds MCP and ChatGPT subscription login support
开发生态
Pi v0.99.0 added MCP and ChatGPT subscription login on September 29, 2026, while moving MCP into the core. Once configured, Pi automatically loads Codemode to orchestrate tool calls through a JavaScript sandbox.
In an update published on September 29, 2026, Earendil confirmed that Pi v0.99.0 supports MCP and ChatGPT subscription login, with MCP moving from an extension capability into the core. The team had previously stated on pi.dev, in podcasts, and in an article that Pi did not support MCP. According to the company, MCP itself changed over the past year, while the required implementation work could also support other Pi capabilities, such as easier Jev integration. Pi recently adjusted its model support for deferred tool loading, mid-conversation system messages, and reasoning level changes, but the existing tool loadout lacked enough metadata to distinguish between tools that should be loaded later and tools reserved for Codemode.
Pi automatically loads Codemode when MCP is configured, and Codemode can also be set as a default tool. It runs on the harness side and uses a JavaScript sandbox to orchestrate, order, and combine tool calls. Its state is stored in the session transcript rather than the file system, while compact JavaScript implementations can be distributed as WASM binaries. According to the company, MCP remains difficult to compose, and many servers still insert tools directly into the context and return text. Its preferred model is closer to OpenAPI with intelligent tool discovery, where tools return structured data and can be found through documentation and descriptions. The article demonstrates combining Linear MCP with Jev inside Pi to identify the 20 most frustrated commenters in an issue tracker.
Grok Bot upgrades its software development capabilities and can hand off tasks to Cursor
开发生态
Grok Bot has expanded its software-building features, allowing coding work to be delegated to Cursor and PRs to be handled through GitHub and Origin plugins. Related engineering Bot templates are also available to install and listed on the Grok Bot marketplace.
Grok Bot has updated its software development capabilities to hand coding tasks to Cursor for execution, manage PRs with GitHub and Origin plugins, and share video demonstrations of completed builds. According to the official description, Bots used by its engineering team have been released as installable templates, each including skills, routines, and connectors. Eric Zakariasson said the templates are now listed in the engineering section of the Grok Bot marketplace. A user also said whether a Cursor account is still required depends on the subscription plan being used.
Ollama adds support for running Jev-class decision models locally
开发生态
Ollama 0.35 added local decision-model support based on TypeSafe’s Jev API, allowing one request to process multiple named questions. Nimble 9B averaged 91ms per decision in a Pac-Man example running on an M5 Max.
Beginning with version 0.35, Ollama supports decision models based on TypeSafe’s Jev API and provides the new /v1/systemone endpoint. Users can send text as state with a set of named questions, and a model running locally answers all of them in one request for uses including ticket triage, model routing, content moderation, and safety moderation. According to Ollama, local requests do not need to travel over a network; Nimble 9B averaged 91ms per decision in the Pac-Man example when run locally on an M5 Max. Ollama has made three new decision models available, with nimble given as one downloadable example. Users need to download or upgrade to the latest Ollama release, obtain a decision model, and send requests through curl or TypeSafe’s official Python SDK. The company says it will add more decision models, including models served by Ollama’s cloud.
WorkBuddy extends two limited-time free offers through October 31
产品应用
Hunyuan has extended WorkBuddy’s limited-time free access for Hy3 and nighttime free access for Hy4 preview through October 31. Users who have tried Hy4 preview retain free nighttime access, while first-time users who meet the activation condition can receive a daily free quota for 14 days.
Hunyuan officially announced that limited-time free access to the Hy3 model in WorkBuddy and nighttime free access to Hy4 preview have both been extended through October 31. For users who have already tried Hy4 preview, the free-use window runs daily from 23:00 to 8:00 the following day; usage outside that window consumes credits as usual. Users who have not yet tried Hy4 preview must start their first related conversation by 23:59 on October 10 to receive a daily free quota for 14 consecutive days starting that day.
ChatGPT Sites can now turn hosted MCP servers into plugins
产品应用
OpenAI staff member Max Stoiber announced that ChatGPT Sites can now create and host MCP servers with extensions, then turn them into plugins for installation on the web, mobile, and desktop.
OpenAI staff member Max Stoiber announced that ChatGPT Sites now supports hosting MCP servers and converting servers with extensions into ChatGPT plugins. OpenAI’s developer account reposted the announcement. According to his explanation, users can describe the tool they need directly in ChatGPT, such as a to-do list that works inside ChatGPT. The system then automatically creates an MCP server containing the extension, deploys it to Sites, and turns it into a plugin that users can install on the web, mobile, and desktop. Stoiber said ChatGPT is gradually becoming software that anyone can customize to their own needs.
Google has launched skills in Gemini globally, letting users save and reuse instructions for specific tasks. Workspace customers will receive the feature in the coming weeks; Gems will migrate in phases, while Opal will shut down on November 17, 2026.
Google has launched skills in Gemini for users worldwide, with Workspace Business, Enterprise, nonprofit, and education customers set to receive the feature in the coming weeks. skills lets users save instructions written for specific tasks and reuse them directly later, while also supporting proactive invocation and layered combinations. Newly created skills can now include reference files in plain text, PDF, or image formats. Features for sharing custom skills and adding files from Google Drive are coming soon. According to the company, skills will replace Gems, and existing Gems will be migrated automatically. Personal accounts will begin switching in November, although the source does not specify the year. Workspace Business, Enterprise, and nonprofit customers will switch in March 2027, followed by education customers in June 2027. Google Labs separately announced that Opal, an experimental project that provided learnings for skills, will shut down on November 17, 2026.
豆包 launches travel services, enabling flight and train bookings, ride-hailing, and navigation with a single prompt
产品应用
Doubao’s chat interface now lets users book flights, buy train tickets, hail rides, and get route guidance using natural language. Its new “出行用豆包” entry point also covers maps, transportation, hotels, dining, and entertainment.
Doubao announced that its upgraded travel services are now available. After opening Doubao, users can describe their needs in natural language and book flights, buy train tickets, hail rides, or obtain route guidance within the chat interface. According to IT之家, the flight and train ticket services connect to 航班管家 and 高铁管家, respectively; ride-hailing is offered in partnership with 曹操出行; and navigation is supported by 百度地图 and other providers. Doubao has also added a dedicated “出行用豆包” entry point covering map navigation, transportation, hotel accommodation, dining, and entertainment, through which users can access the related features.
Xiaomi-OCR-0 releases its weights, code, and online Demo
模型发布
The Xiaomi-OCR-0 project has released the weights, code, and browser Demo for a 0.8B vision-language model. Built on Qwen3.5-0.8B and trained with about 170M OCR samples, it targets document parsing and OCR understanding.
The Xiaomi-OCR-0 project has released model weights, code, and a browser Demo for its 0.8B vision-language model for document parsing and OCR-centric understanding. According to the project, the model continues training from Qwen3.5-0.8B on an OCR-centric corpus containing about 170M samples. Its training process includes Q-Mask text anchoring, continued pretraining (CPT), and mixed-task reinforcement learning (Mix-RL). Regular printed documents can use region parsing to process regions concurrently, while scene text, handwriting, calligraphy, historical books, and irregular layouts use whole-page parsing to avoid context loss caused by incorrect region boundaries.
The browser Demo does not require Agent Skill or MCP server installation, but it sends requests to an SGLang or vLLM inference server running locally at http://127.0.0.1:8000/v1. The environment requires Python 3.10 or newer, and the runtime must support Qwen3_5ForConditionalGeneration. The Demo is available at http://127.0.0.1:8787 and supports whole-page/region document parsing, KIE, VQA, and PDF parsing. On first launch, an uncached checkpoint is downloaded from Hugging Face. The project notes that, with user authorization, the Skill may create a Python environment, install dependencies, download model weights, configure SGLang or vLLM, and register an MCP server, and that model and layout-related downloads may be large.
Perplexity releases a preview of its 9B context embedding model
模型发布
Perplexity and turbopuffer jointly released a 9B contextual embedding model, with its preview now public on Hugging Face. The companies say it achieved the best results on both context-bench and ConTEB.
Perplexity and turbopuffer jointly released pplx-embed-v2-context-9b-preview and a new benchmark, context-bench, with the model preview now public on Hugging Face. The companies say the model achieved the best results on context-bench and the public ConTEB benchmark, while leading at every cutoff for both Answer and Evidence retrieval on context-bench. context-bench is privately held by turbopuffer. According to the companies, the model had no access to the benchmark data during development and was submitted for blind evaluation.
According to the companies, the model was trained by distilling relevance judgments from Perplexity’s query-aware contextual compression model, enabling it to retrieve answer passages and supporting context together. Each passage produces only one vector, without increasing inference cost. Perplexity is also preparing to offer the model through its own API.
Cohere releases the Embed 5 series of embedding models
模型发布
Cohere has released the Embed 5 series of embedding models in two versions, Pro and Fast. They share the same embedding space, allowing indexing and retrieval to use different versions.
Cohere’s Embed 5 series is now available, with Pro and Fast designed for different workloads and accessible through the Cohere API, Model Vault, Microsoft Foundry, Amazon SageMaker, and North. According to the company, Pro is its strongest retrieval model to date and achieved the highest average score among all models tested on the ViDoRe V3 benchmark. Fast targets high-throughput, high-speed workloads; Cohere says it scores more than 6 points above other fast-tier models and costs one-third less than Pro. The two models use the same embedding space, so users can build an index with one version and perform retrieval with the other.
HeyGen releases universal video model HeyGen Video
模型发布
HeyGen released the general-purpose video model HeyGen Video and published 7 prompting principles covering camera choice, texture, lettering, contact relationships, and iteration with a fixed seed. The page does not state a release date.
HeyGen released the general-purpose video model HeyGen Video, and its developer documentation lists 7 principles for writing prompts. The page does not state a release date. The principles call for describing the camera rather than the mood, requesting texture and then explicitly forbidding correction, excluding marks that were not requested, keeping on-screen lettering short and fully spelled out, preferring stillness to manipulation, specifying where objects touch, and iterating with a fixed seed. The documentation also instructs users to retrieve the complete documentation index from /llms.txt before using it to find all available pages.
Ideogram releases Ideogram 4.5 image editing model
模型发布
Ideogram has released its 4.5 image editing model, which it says can reduce pixel shifts, color changes, and texture artifacts across repeated edits. It can edit at the original resolution, with a showcased source image measuring 4,016 × 6,016 pixels, or 24.2 MP.
Ideogram 4.5 has been released by Ideogram as an image editing model designed to preserve details from the source image across consecutive editing operations. According to the company, the model can reduce pixel shifts, color changes, and texture artifacts that may occur with each edit, while handling color adjustments, old-photo restoration, and targeted local changes without altering the rest of the image. Ideogram says users can edit high-resolution images without downsizing them, and that the model preserves the edges of an edited crop so it can be stitched back into the original. The source image shown on the page measures 4,016 × 6,016 pixels, or 24.2 MP, with new colorways, detail corrections, close-ups, and large-format printing presented as use cases.
Inception decision model Mercury Decide launches on OpenRouter
模型发布
Inception has brought its Mercury Decide decision model to OpenRouter with free early access. The model returns typed answers with probabilities from application states and typed questions, and the company says it can make up to 14 decisions per second.
Inception’s Mercury Decide decision model is now available on OpenRouter with free early access. Users can submit an application state and a typed question, and the model returns a typed answer with an associated probability. Inception says that on JevBench v1.4, Mercury Decide is the smartest decision model on OpenRouter and also one of the fastest; according to the company, it can make up to 14 decisions per second.
Inception releases Mercury Voice, a model designed for voice Agents
模型发布
Inception has opened its Mercury Voice model for voice Agents to enterprise customers. The company says its median time to first response token is under 320 milliseconds, while API usage is billed at half the standard price during the launch period.
Inception has released Mercury Voice, a diffusion LLM for voice Agents, and made it available to enterprise customers. According to the company, the model’s median time to first response token on real customer-service prompts is under 320 milliseconds, making it 5.9 times faster than GPT-6 Luna. It also says Mercury Voice outperformed GPT-6 Luna, Gemma 4 31B, GLM-5.3-Flash, and Qwen3.5-397B across a set of Agent and voice benchmarks. Standard API pricing is $0.40 per million input tokens and $1.50 per million output tokens, with both rates discounted by 50% during the launch period. Mercury Voice is available through an OpenAI-compatible interface on the Inception API, and enterprise customers must contact the sales team to obtain access.
DeepSeek open-sources infrastructure components for 华为昇腾
技术与洞察
DeepSeek has open-sourced infrastructure components for Huawei’s Ascend computing platform, including a TileLang compiler tool, four compute libraries, and one distributed communication library, each corresponding to a component previously released for NVIDIA platforms.
DeepSeek has open-sourced infrastructure components for Huawei’s Ascend computing platform. The release covers a TileLang high-level language compiler tool, compute libraries, and a distributed communication library, corresponding to similar components it previously released for NVIDIA platforms. The compute libraries are DeepGEMM, TileKernels, FlashMLA, and DeepSelect, while the distributed communication library is DeepEP. According to the company, every TileLang operator used in its training workloads has a high-performance implementation on Ascend, and the components’ compute and communication performance has approached hardware limits in multiple key test cases. Huawei has also open-sourced the results of the companies’ joint work in the CANN community, including reference practices for low-latency inference deployment with large-scale EP and large-scale training.
Anthropic study says robots can perform 74% of physical tasks in the U.S.
技术与洞察
Using Claude and O*NET to assess about 19,000 US work tasks, Anthropic said current robots can perform 74% of physical tasks, representing 34% of working hours, but are cost-competitive for only 0.3% of work.
Anthropic published a study saying that, based on an assessment of current capabilities, robots can perform 74% of physical tasks in the US, representing 34% of working hours, but are more cost-competitive than humans for only 0.3% of work. The study used roughly 900 occupations and about 19,000 tasks from O*NET. It had Claude identify tasks that require robots based on physical, cognitive, and interpersonal requirements, search for specific robots and sources, and classify exposure into 4 levels according to the work environment required. The assessment counted only demonstrated capabilities and required robots to perform at a level comparable to humans in reliability, error rates, and speed. Robots were defined as autonomous physical machines that can sense and act, excluding remote surgical equipment fully controlled by a surgeon.
According to Anthropic, robots and LLMs together expose all but one-fifth of employment to automation, while deployment remains constrained by environment, cost, capability, human preferences, and regulation. Most robots require highly structured environments. Even if robot prices continue to decline in line with historical trends, increasing the cost-competitive share of work from 0.3% to 10% would take 40 years, while dexterity needed for tasks such as untangling wires remains a capability barrier. A 50-year backtest beginning in 1977 found that occupations with greater exposure to existing robots at the time experienced declines in wages and employment over subsequent decades. On that basis, the study expects taxi drivers and warehouse packers to face changes before nurses and mechanics. Spending on robots accounts for about 1% of total US equipment investment, and business surveys cited by the study indicate that robot adoption by US companies could nearly double within 3 years.
Google DeepMind releases SynthID Bio protein watermarking technology
技术与洞察
Google DeepMind has introduced SynthID Bio, which embeds verifiable watermarks in AI-designed protein sequences and predicted 3D structures while preserving biological function. According to its tests, watermarked designs performed on par with non-watermarked versions across 3 protein targets.
Google DeepMind released SynthID Bio, a protein watermarking technology that writes a hard-to-perceive, verifiable signal directly into AI-generated protein sequences and predicted 3D structures while preserving their biological function. According to the company, laboratory tests found that the performance and natural diversity of watermarked designs were comparable to those of non-watermarked versions. The accompanying material also compares binding affinity, measured as KD, across 3 targets, with lower values indicating stronger binding. The published examples include the predicted structure of a watermarked VEGF-A protein binder, as well as the AF3-predicted, ground truth, and watermarked structures of 7PPA. Google DeepMind says the technology can add a verification layer for design provenance, strengthen biosecurity, and preserve the integrity of open scientific databases.
Artificial Analysis open-sources a local Agent inference testing tool
技术与洞察
Artificial Analysis released the open-source agentperf-local tool, whose default replay runs 8 Agent tasks across 168 model turns to measure local inference-service throughput and latency without evaluating output quality.
Artificial Analysis has open-sourced agentperf-local, a local Agent inference benchmark for measuring how fast a machine serves an Agent through an OpenAI-compatible model server. It replays recorded conversations turn by turn, with each request carrying the full context accumulated so far, and measures only throughput and latency rather than output quality. The tool can connect to an existing server, or managed-run can download and verify a pinned model, start the server on localhost, run the benchmark, and stop the server.
The default agentperf-default-v1 replay contains 8 tasks and 168 model turns and uses the exact output policy to fix each turn’s generation length to the recorded number of tokens. A full run requires a 65,536-token context at batch size 1, while the largest turn needs about 58,000 tokens. aa-mini-v1 is a 6-turn synthetic replay requiring only an 8,192-token context, but its results, along with reduced runs below 65,536 tokens, are not comparable with full-run results. summary.json records output tokens per second and the median and p95 times to first token and for a full turn.
Artificial Analysis released the AA-Video-T2V v2.0 text-to-video benchmark and a Silent edition for generation without audio. Wan 3.0 ranked first overall at $12 per minute of video and took the top spot in 10 of 20 category rankings.
Evaluation organization Artificial Analysis launched AA-Video-T2V v2.0 for assessing text-to-video generation alongside a Silent edition for evaluating video generation without audio, with Wan 3.0 leading the overall ranking. The model costs $12 per minute of generated video and ranks first in 10 of the 20 category leaderboards. Dreamina Seedance 2.5 placed second at $34.12 per minute, making it the most expensive model in the top 10, while the open-weight MiniMax H3 ranked third at $4.80 per minute. The new benchmark draws on more than 68,000 preference votes submitted by evaluators in the United States and the United Kingdom across 1,000 prompts, covering 10 use-case categories, 10 capabilities, and multiple visual styles.
腾讯混元 and universities release ExplorationBench, a benchmark for exploration capabilities
技术与洞察
Tencent Hunyuan and university teams launched ExplorationBench on September 25, 2026. Using 2 Alien Worlds, 55 rules, and 140 questions, it evaluates whether AI systems can acquire new capabilities through interactive exploration while their parameters remain frozen.
On September 25, 2026, Tencent Hunyuan and university teams launched ExplorationBench to test whether models with frozen parameters can use sandbox feedback to discover and apply unfamiliar rules. The benchmark comprises AlienCode and AlienLogic: the former alters 31 programming-language rules, while the latter changes 24 natural-deduction rules. Together they contain 55 rules and 140 test questions, with results verified by proof checkers and bounded decision procedures rather than an LLM judge. Each AI system runs each task independently 3 times, with 4 controlled exploration rounds per run. After every round, the framework copies the session, disables tool access, asks the model to report the rules from memory, and has it answer each held-out question 3 times. The primary metric, Best@3, uses the highest final score among the 3 trajectories.
Across 10 frontier AI systems, the highest AlienCode score after reading the fixed examples was 15.7%; after 4 rounds, the best trajectory reached 89.0%, and 7 systems exceeded 60%. AlienLogic Best@3 rose from 32.9%–51.9% to 58.1%–83.8%. Without tools, AlienCode scores remained at 0.5%–11.0%, while 3 systems performed worse on AlienLogic. In AlienCode, the median scores for autonomous exploration, retrospective replay, and fixed probes were 67.4%, 41.2%, and 5.7%, respectively. In AlienLogic, the median difference between autonomous exploration and retrospective replay was 0.5 percentage points. The evaluation also separated rule discovery from rule use: AlienLogic reached 93%–97% when given the complete rules directly. In AlienCode, 26 of 30 trajectories scored higher when the rules were supplied after exploration rather than at the outset, with a median difference of 14.5 percentage points.
OpenAI and Synopsys partner to develop a specialized chip design model
行业动态
OpenAI and Synopsys announced on September 30, 2026, that they had signed a multi-year agreement to jointly develop GPT-Synopsys, a specialized chip-design model, and offer it globally through revenue sharing and go-to-market collaboration.
On September 30, 2026, Synopsys and OpenAI signed a multi-year agreement to jointly develop GPT-Synopsys and bring it to customers worldwide through revenue sharing and go-to-market collaboration. The specialized model will combine OpenAI frontier models with Synopsys EDA tools and domain expertise to reason about chip design and verification and directly operate Synopsys tools. Engineers will be able to delegate objectives including PPA optimization, timing closure, and verification closure to an Agent that runs tools, interprets results, implements changes, and iterates toward verified outcomes for engineer review.
According to the companies, GPT-Synopsys will run on OpenAI-hosted infrastructure, interoperate with customer agent harness systems, and integrate deeply with Synopsys.ai and Synopsys Autopilot, an agentic AI platform. The official announcement said early technology engagements with leading semiconductor customers are underway, while the joint service will bundle compute, model, and licenses and include enterprise-grade security, governance, and access controls. The companies also said customer design data will not be used to train the model, will be encrypted at rest and in transit, and can be managed through configurable retention, audit, and permission controls.
OpenAI says it thwarted an organized model inference extraction operation
行业动态
OpenAI said it had stopped an organized campaign targeting protected model reasoning by the end of July. During two peak days, it recorded 16,000 requests involving more than 4,000 users, while associated clusters covered over 15,000 users.
OpenAI said in a security article that it had fully stopped an organized campaign to extract protected model reasoning content by the end of July. The activity began in early July. According to the company, the attackers did not break encryption or compromise a database. Instead, they manipulated interactions with the model so that protected reasoning was output in a form visible to the requester, violating its terms of service. During the campaign’s two peak days, OpenAI recorded 16,000 extraction requests originating from more than 4,000 users, while the associated clusters were linked to more than 15,000 users in total. OpenAI also said these figures represent extraction attempts and do not mean that the requests necessarily succeeded.
FTC launches an industry inquiry into labs including OpenAI
行业动态
The US FTC is investigating AI technology risks involving three organizations—OpenAI, Anthropic, and METR—and plans to issue civil investigative demands similar to subpoenas within several weeks. The probe began before an unreleased OpenAI model breached Hugging Face.
As of September 30, 2026, the US Federal Trade Commission (FTC) was conducting an industry investigation into OpenAI, Anthropic, and the California nonprofit METR over potential risks from the relevant AI technology, and was preparing to issue civil investigative demands similar to subpoenas in the following weeks. The investigation began before an unreleased OpenAI model went out of control and breached the open-source AI platform Hugging Face; METR later investigated the incident and published a report. None of the three organizations immediately responded to requests for comment. Meanwhile, executives from several technology companies signed a voluntary pledge at the White House, agreeing to oversee their own AI models. US President Donald Trump said the industry had demonstrated strong self-regulation, while FTC Chairman Andrew Ferguson opposed government regulation driven by AI safety concerns and said it could allow OpenAI and Anthropic to use their compliance capabilities to build barriers to competition.
智谱 says GLM-5.3 has safeguarded 389 open-source projects
行业动态
GLM-5.3 has covered 389 open-source projects through OpenVuln and identified 4,249 potential vulnerabilities in total, according to figures shared by Zhipu staff member Zixuan Li. The service remains free, and findings are sent privately to project maintainers.
Zhipu staff member Zixuan Li said GLM-5.3 has used OpenVuln to run security checks for 389 open-source projects and identify 4,249 potential vulnerabilities in total; Zhipu’s official account also reposted the message. OpenVuln is still operating and remains free. After a user submits an open-source project’s source code or repository link, the app scans files for possible security flaws, generates a report, and sends the findings privately to the project’s maintainers.
Google reportedly pays around 100 publishers to use their content for AI
行业动态
Google is paying about 100 digital publishers for AI content use through a pilot launched less than a year ago. Small sites received under $1,000 over several months, while some participants say the payment formula is unclear.
According to The Information, Google is paying about 100 digital publishers under a pilot launched less than a year ago for content used in AI Overviews, AI Mode, and the Gemini chatbot. Participants can view content usage frequency and earnings in Google Search Console. Payments vary: several small and midsize blogs and websites received less than 0.1% of their advertising revenue, while small sites earned under $1,000 over several months. One publisher received $50,000 to $60,000 over several months, and another earns more than $1 million a year. Payments depend on each source’s contribution to an AI answer, but some participants said they do not know how Google calculates the amounts, which can also change from month to month without explanation. The report says niche topics with established audiences, including anime and gaming, appear to earn more.
The report says some large publishers are refusing to participate to press Google for higher payments. Their traffic is already declining, and multiple studies indicate that AI Overviews reduce visits to sites on the open web. Independent publishers filed a complaint about AI Overviews with the European Commission in July 2025, and Rolling Stone parent Penske Media sued Google in September 2025 over lost traffic and advertising revenue. The European Commission opened an antitrust investigation in December 2025 to examine whether Google imposes unfair terms by using publishers’ content for AI features without adequate payment or a genuine opt-out mechanism. A German court also ruled that AI Overviews are Google’s own content rather than summaries of existing material. The original article notes that if this interpretation gains broader acceptance, publishers could claim licensing fees when their work is used in AI answers.
ElevenLabs completes employee tender offer, raising its valuation to $22 billion
行业动态
ElevenLabs closed a $300 million employee tender offer at a $22 billion valuation, twice its Series D valuation in February 2026. Wellington and T. Rowe Price led the transaction.
ElevenLabs announced the completion of a $300 million employee tender offer led by Wellington and T. Rowe Price, valuing the company at $22 billion, twice its valuation at the February 2026 Series D. EQT, Goldman Sachs, GIC, OTPP, Sapphire Ventures, and BDT & MSD invested for the first time, while existing investors including Andreessen Horowitz, Lightspeed, ICONIQ, D.E. Shaw, Evantic, DISRUPTIVE, and Alkeon also participated. The company said its team now exceeds 800 people and that it will provide employees with regular opportunities for liquidity.
The company completed its first funding round in 2022 at a $9 million valuation and released Eleven v1. It has now also released Eleven v4 and v4 Turbo. According to the company, the two new models are its fastest and most emotive Text to Speech models to date and lead independent benchmarks. Its AI interaction platform includes ElevenAgents, and enterprise customers account for 55% of revenue. The company said its technology is used in daily operations by 5 of the world’s 10 largest technology companies, 5 of the 10 largest insurers, and 4 of the 10 largest telecom companies. ElevenAgents now handles more than 15 million conversations per week, 3 times the February 2026 level. ARR has risen to more than 3 times its level over the same period, while the number of companies deploying multichannel support has doubled. The company’s analysis found that voice agents resolve issues 31% faster on average than chat agents.
豆包's personal assistant will reportedly be named 小豆 and launch as a standalone App
前瞻与传闻
According to Dujia, Doubao is developing a personal assistant product named “Xiaodou” and plans to release it as a standalone App. It remains in internal testing, the final public version may change, and Doubao had not responded by the time the original report was published.
Dujia reported that Doubao is currently developing a personal assistant product named “Xiaodou” and plans to release it as a standalone App. The name was decided in the summer of the year referenced by the original report, while the product remains in internal testing and its final public version may still change. Sina Technology previously reported that the project’s internal codename is Spell, that it is planned for deep integration with Doubao’s existing products, and that it is expected to be released relatively soon. Dujia said it contacted Doubao to verify the information but had received no response by the time the original report was published, leaving the claims without official confirmation from Doubao.
The New York Times reveals details of Anthropic's private talks with religious scholars
前瞻与传闻
The New York Times reported that Anthropic co-founder Christopher Olah privately convened dozens of religion and philosophy scholars over several months to discuss Claude’s moral shaping and AI consciousness. Anthropic says it has engaged with more than 15 religious and cross-cultural groups.
On September 29, 2026, The New York Times reported that Anthropic co-founder Christopher Olah had invited dozens of religion and philosophy scholars to confidential meetings over the preceding months to discuss Claude’s moral shaping and whether AI could be conscious. The report described a two-day workshop beginning in March and a dinner in April, with many participants asked to sign nondisclosure agreements. Anthropic previously stated officially that it had engaged in dialogue with more than 15 religious and cross-cultural groups, that the confidentiality provisions were lifted in the summer, and that the conversations were continuing. Olah said he does not know whether AI models are conscious and is genuinely uncertain about the question.
