Gemini Omni 1.1 Flash launches with start/end frame controls and 4K upscaling
要闻
Google has released Gemini Omni 1.1 Flash for developers. According to the company, it supports scene extension, start and end frame control, and 1080p or 4K output, while 360p previews are up to 60% faster and cost one-third as much as standard 720p generation.
Google has released Gemini Omni 1.1 Flash and says the version is ready for production use, with access now available through the Gemini API in Google AI Studio or the Gemini Enterprise Agent Platform. According to the official description, the new version can analyze up to 10 seconds of preceding context when extending a scene, whereas previous models referenced only the final second. Videos can be extended in 10-second increments to a cumulative maximum of 40 seconds. Developers can also specify the starting and ending frames of a shot, allowing the model to generate continuous video between two keyframes for camera orbits, zoom transitions, or seamless loops. Compared with standard 720p, lightweight 360p previews generate up to 60% faster based on system throughput and cost one-third as much; finished videos can be output in 1080p or 4K.
Tibo says usage limits have been reset for all ChatGPT Work and Codex users
要闻
Codex head Tibo said usage allowances had been reset for every ChatGPT Work and Codex user across both services. A user also called for the existing 5-hour usage limit to be removed.
Codex head Tibo announced that the usage allowances for all ChatGPT Work and Codex users had been reset, with the action covering every user of both services. User feedback focused on the existing 5-hour usage limit. One user said the restriction had prevented them from using their remaining allowance before the reset and asked for the limit to be removed. Tibo’s announcement confirmed only that the allowance reset had been completed; the source did not say whether the 5-hour limit would be removed or provide a timetable for such a change.
OpenAI has opened Luna Reserve to some individual ChatGPT Plus and Pro accounts. Once regular usage is exhausted, it provides a separately metered, capped reserve available only for GPT-5.6 Luna.
OpenAI has introduced the Luna Reserve backup mechanism in Codex and ChatGPT Work for some individual ChatGPT Plus and Pro accounts, providing additional usage after an account exhausts its regular allowance. Luna Reserve is metered separately from the regular allowance, has its own cap, and can be used only with GPT-5.6 Luna; it neither restores exhausted regular usage nor unlocks more capable models. The mechanism is currently unavailable in ChatGPT Business and Enterprise workspaces and does not apply to regular Chat or the API.
ChatGPT updates Temporary Chat with memory and plugin access and conversion to regular chats
要闻
OpenAI announced new controls for ChatGPT temporary chats. When creating a chat, users can enable personalization to access existing memories, custom instructions, and plugins, or save the conversation and convert it into a regular chat governed by their account settings.
OpenAI says ChatGPT is rolling out controls for temporary chats, allowing users to enable personalization or save the conversation when creating a temporary chat. Once saved, the conversation is converted into a regular chat governed by account-level settings. With personalization enabled, a temporary chat can access existing memories, custom instructions, and plugins from the account, but it does not create new memories during the interaction. Personalization can only be configured when the temporary chat is created and cannot be changed after the conversation begins. A personalized temporary chat does not appear in the sidebar or chat history until it is saved, after which it appears as a regular chat.
OpenAI joins Anthropic and others in calling for stronger cyber defenses
要闻
OpenAI joined Anthropic, AWS, Google, Microsoft, Oracle, and others in a collective call for stronger cyber defense, setting out 3 principles and asking frontier AI companies to make Agentic identities traceable and accountable.
OpenAI, Anthropic, AWS, Google, Microsoft, Oracle, and other companies jointly released an action initiative to strengthen cyber defense, and OpenAI stated that the window for responding to AI-driven cyberattacks is limited. The initiative sets out 3 principles: recognizing that current security measures are insufficient, empowering defenders with AI that has cyber capabilities, and mobilizing a collective global response. According to OpenAI, the proposed actions address organizations, cybersecurity companies, governments, and frontier AI companies. Frontier AI companies are asked to ensure that Agentic identities are traceable and accountable, and to provide model access and funding to under-resourced defenders of critical infrastructure.
Anthropic releases a research preview of Model Hardware Standard
要闻
Anthropic has opened the MHS research preview to an initial group of research labs and advanced manufacturers. The company says the specification lets AI Agents operate microscopes, liquid handlers, and robotic arms in parallel while reducing device integration from weeks or months to hours or minutes.
Anthropic has provided an early version of the Model Hardware Standard (MHS) to an initial group of partners in science, robotics, electronics, and manufacturing to test AI Agent operation of physical equipment and jointly develop safety evaluations and best practices before making the standard open source. Development of MHS began as a collaboration between Anthropic and HHMI Janelia Research Campus. It works with any device that has a programmable interface, is model-agnostic, and allows an Agent harness to connect through standard protocols such as the Model Context Protocol. Anthropic says Agents can use it to control microscopes, liquid handlers, robotic arms, and other equipment in parallel for tasks including drug discovery experiments and laser calibration on a quantum computer, while reducing hardware integration from weeks or months to hours or minutes.
MHS uses a standardized driver to unify interfaces across different devices, applying primitives such as read and write to retrieve or set parameters. It also publishes device information in a common format so devices and Agents can discover one another and communicate across networks without a custom translator. Users can add information such as weight, measurable properties, adjustable parameters, and safety limits through natural-language tags, which the driver uses to generate a reference file automatically. Equipment can be controlled through three mechanisms: MCP, the command line interface, and code files (APIs). An Agent can orchestrate steps, monitor results, adjust parameters in real time, and place long-running or high-speed tasks in code files. Anthropic says Claude repeatedly adjusted a laser and observed the results through a camera during testing, then generated a deterministic script that completed laser alignment with a single command.
蚂蚁百灵 releases the finance-enhanced model Ling-3.0-flash-Fin
模型发布
Ant Ling has released the finance-focused Ling-3.0-flash-Fin model with 124B total parameters and 5.1B activated parameters. It is temporarily available for free on two platforms, while its weights are scheduled to be open-sourced in the week following release.
Ant Ling has released the finance-focused Ling-3.0-flash-Fin model, although the source text does not specify an exact release date. Built on Ling-3.0-flash, the model has 124B total parameters and activates 5.1B parameters per run. According to the company, it performed strongly across multiple financial benchmarks and also improved its general-capability score. The model is now available through Vercel AI Gateway and OpenRouter. Free access on Vercel AI Gateway runs until September 25 as stated in the source, which does not specify the year, while OpenRouter offers a one-month free trial. The company also said the model weights are scheduled to be open-sourced in the week following release.
fal launches H3 Max, a post-trained model based on MiniMax H3
模型发布
fal has released H3 Max, a post-trained video model based on MiniMax H3. According to the company, it generates a 5-second video in under 3 seconds, delivers roughly 35 times the throughput of the official H3 endpoint, and ranks first in all three human-preference evaluation categories.
fal now offers H3 Max through its platform. The post-trained video model was developed by fal Research from the open-weights MiniMax H3 and optimized by fal’s inference team. According to fal, the team added new data during post-training, focusing on prompt adherence and visual quality while iterating on the model and inference stack together. The model was trained and served on NVIDIA GB200 NVL72 systems, which the company says deliver up to 2 times the per-chip performance of the previous-generation accelerators it had used. H3 Max was also trained on fal Serverless.
fal’s quality evaluation primarily used head-to-head human-preference comparisons across overall quality, prompt understanding, and aesthetics, with results aggregated through Bayesian Elo ratings with 95% confidence intervals. The 12 video models evaluated included the official MiniMax H3 endpoint, Gemini Omni Flash, Wan 3.0, Seedance 2.5, Kling 3, and Veo 3.1. According to the company, H3 Max ranked first in all three categories, won a majority of head-to-head comparisons against every tested model, and also placed first in independent rankings from Artificial Analysis and Design Arena. fal says the model generates a 5-second video in under 3 seconds, provides roughly 35 times the throughput of the official H3 endpoint, and is 15 times faster on average than models with comparable quality. It is available through Playground, fal Agent, and the API, with a 50% discount during the first week.
Midjourney V8.2 image editing model enters testing
模型发布
Midjourney has opened testing of its first V8.2 image edit model to all users and updated the UIs on its main and alpha sites. The company is collecting edge cases and interface feedback and says rapid follow-up updates will continue.
Midjourney’s first V8.2 image edit model has entered testing for all users, and the UIs on midjourney.com and alpha.midjourney.com have also been updated. According to the company, the alpha site will continue running experiments around redesigning the interface for the model, and the test is expected to surface edge cases where the model does not work as intended. If an image produces poor results, users can click the sad face emoji on the right side of the lightbox or submit the corresponding jobid or URL while the image remains open in the lightbox. Interface suggestions and model edge cases can be posted in #ideas-and-features, while results made with the model can be discussed or shared in #edit-showcase. The company says rapid follow-up updates will continue, with this round of testing involving a broader community.
BreezeBlue open-sources the Breeze TTS 2 speech model
模型发布
BreezeBlue has released Breeze TTS 2, an open-weight speech model for real-time interaction. The company says it ranks No. 1 among open-weight models on the Artificial Analysis TTS leaderboard and supports streaming 24 kHz PCM output.
BreezeBlue has launched Breeze TTS 2, an open-weight text-to-speech model for real-time interaction. According to the company, the model ranks No. 1 among open-weight models on the Artificial Analysis TTS leaderboard and outperforms frontier proprietary systems. Natural-language instructions can design a voice without reference audio or direct a reference speaker’s tone, emotion, pace, and delivery. Using --cfg-scale 4 strengthens instruction-following. Speaker cloning requires clear reference speech with as little background noise as possible and an exact transcript.
The Breeze TTS 2 CLI and single-concurrency streaming API use PyTorch eager streaming by default and skip graph warmup. The API returns mono 24 kHz signed 16-bit little-endian PCM, while --fast-all enables the best configuration for every inference stage at the cost of additional cold-start time. The default Docker image targets H100/Hopper（sm90）, with a separate build method for A100, and all required model components are included in the checkpoint. The source code uses Apache License 2.0, while model weights, checkpoints, adapters, derivative models, and self-hosted outputs are separately governed by the BreezeBlue Research and Non-Commercial License. Commercial use requires written authorization from RESONIA, INC.; the contact address is contact@breeze.blue.
Cohere releases the Parse 5 document parsing model
模型发布
Cohere has officially launched Cohere Parse 5, a document parsing model that handles text, tables, forms, and images. According to the company, it scores 79.2 on ParseBench and costs $1.50 per 1,000 pages.
Cohere has officially released and launched Cohere Parse 5 for building RAG systems and automating document processing. The model recognizes text, tables, forms, and images in documents, then produces machine-readable documents with bounding boxes. According to the company, it scored 79.2 on ParseBench, outperforming competitors including Mistral OCR 4. Cohere has priced the service at $1.50 per 1,000 pages; according to its explanation, this is up to 95% cheaper than frontier large models.
New Qoder launches with coding and general-purpose modes plus a desktop app
开发生态
Qoder has launched a new Agent workspace and desktop app with separate Coding and general-purpose modes. Its international and China editions are now available to download, while its global user base has exceeded 6 million and the system supports more than 40 connectors and over 70 plugins.
Qoder has released its next-generation Agent workspace and desktop client, with downloads now available for both the international and China editions. The product is built around Coding capabilities and includes programming and general-purpose modes. According to the company, the system includes leading global models and an Auto intelligent scheduling engine, while Harness technology supports complex, long-chain autonomous tasks. Qoder also says the platform can connect to work systems through more than 40 connectors and over 70 plugins, with voice activation and proactive reminders added as new features. Qoder’s cumulative global user count has exceeded 6 million, and the company is also offering limited-time benefits including dedicated Credits and model discounts.
Alibaba Cloud’s Bailian large-model service platform will cut Qwen3.8-Flash pricing at 12:00 Beijing time on August 27, 2026, lowering input prices from RMB 1.00 to RMB 0.80 and output prices from RMB 3.00 to RMB 2.70.
Alibaba Cloud’s Bailian large-model service platform announced that new pricing for Qwen3.8-Flash will take effect at 12:00 Beijing time on August 27, 2026, with input prices changing from RMB 1.00 to RMB 0.80 and output prices from RMB 3.00 to RMB 2.70. The notice lists input and output as two separate pricing changes: the input price will move from RMB 1.00 before the adjustment to RMB 0.80 afterward, while the output price will move from RMB 3.00 to RMB 2.70. According to the official statement, the actual timing of the change is subject to implementation on the Bailian platform.
AISI releases optstop, an open-source tool for dynamically optimizing LLM evaluation compute
开发生态
AISI open-sourced the Python package optstop to reduce unnecessary sampling in LLM evaluations through dynamic stopping. In its tests, the tool saved 57%–97% of planned trials under every configured condition.
AISI has open-sourced the Python package optstop and integrated it with the Inspect evaluation framework to stop LLM evaluation sampling dynamically once an estimate reaches a predefined precision requirement. According to AISI, a fixed sample size cannot adjust to the uncertainty remaining in an evaluation and may continue sampling model-task combinations that are already sufficiently precise. optstop tracks the credible interval at two levels—repetitions within a task and tasks within a grouping—and decides whether to stop under either a precision or stabilisation rule. At the grouping level, a hierarchical Bayesian model combines information from different tasks while handling the unequal sample sizes created when some tasks stop early. If neither rule is satisfied, the evaluation runs through its entire planned budget.
For evaluations in which a model succeeds only with very low probability, optstop uses a conservative mechanism that requires more data when the estimated success rate falls below 1%, reducing the risk that early stopping misses a single rare success. According to AISI, every stopping decision and its rationale are recorded and reported, allowing evaluators to verify the number of runs saved. AISI tested binary, ordinal, and continuous scoring across low, medium, and high expected model performance. The public benchmarks used included MATH, GPQA Diamond, and WritingBench; its results showed that optstop saved 57%–97% of planned trials under every configured condition.
Hugging Face unveils the open-source bipedal robot Microduck
产品应用
Hugging Face has unveiled Microduck, a 25 cm open-source biped that works out of the box and supports reinforcement-learning retraining of its behavior policies. It comes in 4 colors and is scheduled to ship before Christmas 2026.
Microduck will ship before Christmas 2026. The 25 cm open-source biped released by Hugging Face can be operated out of the box or retrained by users through reinforcement learning, and its introductory price excludes taxes and shipping. According to the company, each behavior is implemented as a retrainable policy that can be learned in a physics simulation on the user’s machine or through Hugging Face Jobs, then deployed to the physical robot. Demonstrated behaviors include a velocity-tracking gait, sitting down and standing up autonomously, touching its beak to the ground and returning upright, skating with rollers attached, and standing up from a flat position on its back.
Microduck comes in 4 colorways, and the standard package includes the robot, a battery, a USB-C cable, and a game controller. Additional configurations listed on the page include a dual charger with 2 batteries; a bundle with 3 spare motors, 5 motor cables, 2 batteries, a dual charger, 10 NFC tags, Hugging Face credit, a screwdriver, and a screw pack; and another with a laser pointer, an NFC polaroid, 2 rollers, a ball, and 10 NFC tags. According to its description, the SDK, simulation, and full RL training stack are available on GitHub under a permissive license. The physics simulation, every shipped behavior, and its policy can be read, forked, and retrained, while community support is hosted on Discord.
Google has expanded travel planning in Search’s AI Mode: flight price tracking now covers more than 180 countries and territories, points and miles rates are available globally, and hotel booking is rolling out in English in the U.S.
Google has integrated Google Flights price tracking and points and miles rates into Search’s AI Mode. Price tracking is now available in more than 180 countries and territories, while points and miles rates are available globally. Users can enter their itinerary and dates in a conversation, view the latest prices from more than 300 partner airlines and travel sites, and set alerts. They receive an email when prices change, while ticket purchases are completed on an airline or booking platform. Initial support for points and miles rates covers Alaska Airlines / Hawaiian Airlines, American Airlines, Choice Hotels International, Hilton, and Wyndham Hotels & Resorts. Accor, Flying Blue, Hyatt, LATAM Airlines, and Lufthansa Group will be added in the coming weeks.
Hotel booking through AI Mode has started rolling out in English in the U.S. and will integrate with Booking.com, Choice Hotels International, Expedia, Hilton, Hotels.com, IHG Hotels & Resorts, Marriott International, Priceline, Trip.com, and Wyndham Hotels & Resorts in the coming weeks. According to Google, users can receive a hotel list containing guest reviews and comparison factors through a conversation. After selecting a hotel, they can choose “Continue on Google,” select a room type, review the cancellation policy, and pay with Google Pay. The hotel or booking platform serves as the merchant of record and handles customer service.
Gemini Notebook adds support for importing purchased Google Play ebooks
产品应用
Gemini Notebook now supports selected ebooks purchased by users through Google Play. Expert Intelligence initially covers more than 100,000 books and involves collaborations with more than 15 bestselling authors.
Google has launched Expert Intelligence in Gemini Notebook, allowing users to add eligible ebooks purchased through Google Play Books to a Notebook. According to the company, the system answers questions using the book as its source, can generate Infographics, Audio Overviews and Quizzes, and can combine books with personal materials and other sources. The initial catalog covers more than 100,000 books from publishers including Bloomsbury, O’Reilly Media and Penguin Random House. Google has also worked with more than 15 bestselling authors to create Featured Notebooks containing supplemental sources and other supporting content.
Access controls limit use of a book in Gemini Notebook to users who purchased it through Google Play Books. When a Notebook is shared, collaborators need to buy their own copy; otherwise, the book remains unavailable in their Notebook. Users can check eligibility by opening the “Tools” badge on a book’s detail page and seeing whether Gemini Notebook is listed. Google is also providing one book at no charge to users in the U.S. while supplies last. According to the company, Expert Intelligence will later expand to the Gemini app and AI Mode in Search, with support planned for sources including third-party subscriptions, business research reports and textbooks.
谷歌DeepMind pilots double-blind evaluations of frontier AI models
技术与洞察
Google DeepMind is piloting a double-blind evaluation of Gemini Flash Lite with four organizations. According to the company, it is the world’s first use of confidential benchmarks to test a proprietary frontier AI model.
Google DeepMind has launched a double-blind evaluation pilot for Gemini Flash Lite with the Singapore AI Safety Institute, OpenMined, AVERI, and MLCommons, testing the model against confidential benchmarks in a privacy-preserving environment. The setup uses Confidential Space from Google Cloud’s Confidential Computing portfolio to cryptographically verify that the external evaluation data and proprietary model remain private to their respective owners. Evaluators cannot view the Gemini model weights, while Google cannot read the test prompts.
High-stakes external evaluations previously required evaluators to hand over prompts or model providers to hand over model weights. The former risked exposing test questions in advance, while the latter risked exposing intellectual property. According to the company, zero-logging protocols and strict contractual safeguards were already used to protect external prompts, and the pilot adds technical and cryptographic protections to reduce the risk of benchmark contamination. It said these protections are particularly important for sensitive evaluations used in cybersecurity or by government bodies.
Anthropic launches a Claude subscription plan for scientists with 10,000 free seats
行业动态
Anthropic is opening 10,000 one-year seats in its Claude team plan for scientists to researchers worldwide. Standard is free, while Premium, with five times the usage limit, costs $15 per month.
Anthropic has launched the Claude team plan for scientists, initially offering scientists worldwide 10,000 free or discounted subscription seats for one year. Standard seats are free, while Premium seats provide 5x usage limits for $15 per month. According to the company, it plans to expand the program beyond the initial 10,000 seats over the coming months. Applicants must be a principal investigator or equivalent at an academic or nonprofit research institution, and can add researchers from their labs to the plan after verification.
The company is also expanding the range of fields and credits covered by AI for Science. Its support previously focused mainly on biological science researchers using Claude, but it now plans to include other fields and compute-heavy research. When a lab has used its allocated credits and Standard or Premium plans cannot meet its needs, any researcher may apply for up to $50,000 in credits per project. Biology and chemistry researchers remain limited to Opus-class models. According to the company, Claude Fable models will continue to block professional biology and drug development queries because of potential dual-use risks. Anthropic is working with the US government to establish an access program for life sciences professionals to use Mythos-class models, and the first participants have been enrolled.
英伟达 reportedly agrees to acquire Hugging Face for $12.9 billion
前瞻与传闻
According to The Information, citing people familiar with the matter, NVIDIA has agreed to acquire open-source AI model hosting platform Hugging Face for $12.9 billion. Hugging Face recently had about $150 million in annualized revenue, and neither company has officially commented.
The Information reported that NVIDIA and Hugging Face have agreed to a $12.9 billion acquisition deal. Neither company has officially commented. According to people familiar with the matter, the transaction would give NVIDIA a foothold in open-source AI, help it respond to the trend of closed AI companies developing their own chips, and support its return to the cloud services market. Hugging Face recently had about $150 million in annualized revenue, and the reported purchase price is approximately 80 times its forward revenue. The company previously rejected a $5 billion investment offer from NVIDIA, while its valuation in its last funding round was $4.5 billion.
英伟达 reportedly pauses revenue-sharing deals with AI cloud providers
前瞻与传闻
Nvidia has paused some transactions under a new financing program for AI cloud providers, according to people familiar with the matter. The program exchanges credit support for a share of revenue; employees feared antitrust scrutiny, while Nvidia said the business model remains in normal operation.
Nvidia has paused some transactions under its new financing program, according to people familiar with the matter, although the source article does not provide verifiable full dates for either the program’s launch or the pause. Under the program, Nvidia provides credit support to AI cloud providers in exchange for an agreed percentage of their revenue. The arrangement may later be revised or incorporated into another business. The people said Nvidia employees were concerned that the structure could trigger antitrust scrutiny and create the impression that the company was intervening in customers’ business operations. An Nvidia spokesperson said the business model remains in normal operation and is being continuously optimized and refined in response to market demand.
Google employee responds to controversy over Ox Alpha's insinuation
其他
Zhipu released GLM-5.3-Flash and confirmed that it was the previously previewed Ox Alpha. A Google employee said the Gemini team member’s post was poorly timed and had been responding to the growth and progress of Gemini 3.7 Flash.
After Zhipu officially released GLM-5.3-Flash and confirmed it as the previously previewed Ox Alpha, Google employee Logan Kilpatrick addressed questions surrounding an earlier post by a Gemini team member and attributed the controversy mainly to poor timing. The post had prompted speculation about a connection between the Gemini team and Ox Alpha, while Zhipu’s subsequent model release confirmed that Ox Alpha was GLM-5.3-Flash. Kilpatrick said the team had been responding to the growth and progress of Gemini 3.7 Flash.
