Anthropic releases Claude Fable 5.1 and Claude Mythos 5.1
要闻
Anthropic released Claude Fable 5.1 and Mythos 5.1: Fable is generally available and is estimated to cost 25% less for typical token-billed workloads and up to about 45% less for highly agentic work, while Mythos is limited to trusted access programs.
Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, with the former now generally available and the latter offered only through trusted access programs. The two use the same model but have different levels of safeguards; according to the company, Mythos 5.1’s safeguards are designed for cybersecurity and life sciences work. Anthropic estimates that Fable 5.1 costs 25% less than Fable 5 for typical token-billed workloads because of lower cache read pricing. Savings for highly agentic work are often higher, reaching up to approximately 45%. Claude Code defaults to High effort, while Claude Cowork and Claude.ai default to Medium effort.
Anthropic also announced Enterprise Frontier Safeguards (EFS). The company says EFS stores data in cloud infrastructure controlled entirely by the customer rather than Anthropic, providing the same level of privacy as a zero data retention policy. It will be rolled out to enterprise customers in phases; until EFS is available, eligible customers can use Fable 5.1 with zero data retention. The new cybersecurity safeguards block 60% fewer false positives than before, allowing the model to discover software vulnerabilities but not to develop exploits. Anthropic has also established a life sciences access program in partnership with the US government. According to the company, Millennium used Fable 5.1 in testing to identify the cause of a rare crash in its internal systems that its engineers and other models had been unable to explain for several years.
OpenAI has disclosed its release plan for the new Astra model. The company says Astra scored 100% on ExploitBench, found and exploited 2 zero-day vulnerabilities in internal testing, and became its first model to reach the “Critical” cybersecurity threshold.
OpenAI announced that it is preparing to release its new Astra model and plans to make it available soon, with its most advanced cybersecurity capabilities initially limited to selected testers. According to the company, Astra is its first model to reach the “Critical” cybersecurity threshold under the Preparedness Framework and can discover unknown vulnerabilities and develop exploits when given the necessary tools and permissions. Astra scored 100% in the official ExploitBench evaluation. OpenAI also says internal testing found and exploited 2 zero-day vulnerabilities, while expert testing completed a browser sandbox escape and an operating-system privilege-escalation chain from a standard user to root.
To prepare for the release, OpenAI delayed parts of Astra’s development and rollout while strengthening model refusals, system security classification, training-infrastructure isolation, cross-session monitoring, and detection of unauthorized behavior. According to the company, Astra followed safety restrictions and authorization boundaries more consistently in alignment testing, and related reasoning and behavior monitoring will also be used in production. Access to the most advanced cybersecurity capabilities will later expand for defensive use through Daybreak Blue. Safety systems may pause or terminate otherwise normal tasks that are judged to pose a risk.
Gemini 3.7 Flash and two other models gain Agentic video understanding
要闻
Google has added Agentic video understanding to three Gemini Flash models. The company says it can cut token use by up to 88%, reduce costs by up to 66%, and improve accuracy by up to 7%.
Google has launched Agentic video understanding for Gemini 3.7 Flash, 3.6 Flash, and 3.5 Flash-Lite, with uploaded videos and YouTube videos supported at launch through the Gemini API in Google AI Studio or the Gemini Enterprise Agent Platform. Developers must set processing to “agentic” in the API configuration. The feature uses standard Gemini API token pricing and carries no additional feature fee. According to Google, the models combine core reasoning with native video tools to dynamically search, scan, and inspect target segments across visual frames, audio, and transcripts, instead of processing video at a fixed rate that defaults to 1 FPS and can be adjusted through the API.
Google says that across standard video analysis benchmarks, enabling the feature on the three models reduces token consumption by up to 88% and analysis costs by up to 66%, while improving accuracy by up to 7%. It says Gemini 3.7 Flash delivers the highest overall quality and sits on the accuracy-to-cost Pareto frontier among the models tested. The mechanism can determine which segments to inspect, at what speed, and whether to use visual frames, audio, or transcripts, while supporting sub-second moment retrieval, anomaly detection, and counting. Google also says the feature will soon reach all Gemini app users across Flash and Flash-Lite models and will be used for YouTube’s “Ask YouTube” feature on video watch pages in the coming months.
智谱 celebrates the first anniversary of GLM Coding Plan with reset cards for subscribers
开发生态
Zhipu is giving each current GLM Coding Plan subscriber one reset card to mark the plan’s first anniversary. The card restores both the weekly quota and the five-hour quota and remains valid through October 1, 2026.
When GLM Coding Plan reached its first anniversary, Zhipu announced that every current subscriber would receive one reset card, valid through October 1, 2026. The card can fully restore a user’s weekly quota and five-hour quota, and the reset time is not scheduled uniformly by the company. According to Zhipu’s official operations team, users do not need to use the card immediately after receiving it and can choose when to trigger the reset. The reset voucher is available on the “Console - Usage Statistics” page and must be used within its validity period.
Ollama switches its subscription plans to token-based billing
开发生态
Ollama subscriptions now meter cloud-model usage by token, with input, cached-input, and output rates listed per 1 million tokens for each model. The concurrent-request limits for Free, Pro, Max, and Team are 1, 3, 10, and 10, respectively.
Ollama has introduced subscription billing that calculates cloud usage from each model’s rates and token count, with separate input, cached-input, and output prices listed per 1 million tokens. According to Ollama, Pro, Max, and Team include a fixed dollar amount of usage credits each month, while Free provides starter usage for a smaller set of starter models. Every plan can add extra credits; included plan credits are consumed first, followed by the extra balance, and adding credits unlocks all models. Included credits do not roll over. Paid plans refresh each month on the date corresponding to the subscription start date, while Free refreshes based on the signup date. After a plan switch, the new plan’s full monthly allowance becomes available immediately, while the reset date remains tied to the original subscription start date. Running models on the user’s own hardware remains unlimited.
The Team plan starts at $500 per month, supports unlimited users, and includes $1,000 in monthly usage credits shared across the team. Usage beyond that amount is billed from a shared team balance on a pay-as-you-go basis; users can add a fixed amount to the balance and disable automatic usage billing. Ollama creates a separate team account for each member, and the same login can switch between personal and team accounts. Concurrent-request limits are 1 for Free, 3 for Pro, and 10 each for Max and Team. Requests above the limit enter a queue; once the queue reaches its fixed capacity, new requests are rejected until a concurrency slot becomes available. On paid plans, Ollama sends an email when usage reaches 90% of the included monthly allowance, and the notification can be disabled in settings.
World Labs releases Atlas, a world model with a unified architecture for generation, reconstruction, and spatiotemporal simulation
模型发布
World Labs has released Atlas, a next-generation world model whose unified architecture natively handles text, images, video, and 3D across generation, reconstruction, and simulation. In an official example, it generated a one-minute 1440p video from a small number of reference images.
World Labs introduced Atlas, a next-generation world model pretrained from scratch, and plans to use it in future versions of Marble and the company’s other products. Atlas uses a multimodal autoregressive diffusion transformer that combines different inputs into a shared spatial context and generates subsequent frames from that information. Each image is tied to a position in 3D space, while precise camera geometry can also be supplied directly as input. The workflow shown by the company uses one to six images with manually designed camera paths to create new views from any camera position and angle. One example generated a one-minute 1440p video from a small number of reference images.
For reconstruction tasks, Atlas can accept a variable number of scene images and use model knowledge to fill in areas not visible in the inputs. As more inputs are added, less content must be inferred by the model. According to the company, Atlas typically produces faithful reconstructions from only two or three images, outperforming state-of-the-art models trained solely for 3D reconstruction, while its spatial context can also hold more than 100 input images. In one demonstration, adding three photos of a garden, an adjacent cottage, and a main house in sequence progressively reduced the inferred content in the output. Another example used two to 25 ground-level photos to build Stanford Main Quad and generate aerial paths above the campus. The same small set of inputs can also produce multiple camera trajectories with different speeds, lengths, or levels of complexity while keeping the scene consistent.
Meta releases Muse Voice Transcribe, a real-time audio perception model
模型发布
Meta released Muse Voice Transcribe, its first real-time audio perception model. It handles more than 25 languages and over 20 speakers, while supporting streaming ASR and in-sentence code-switching.
Meta has released Muse Voice Transcribe, developed by Meta Superintelligence Labs, with real-time streaming ASR, diarization, and endpointing. The model natively supports more than 25 languages, recognizes in-sentence code-switching, and uses language, keyword, and context biasing to improve recognition accuracy. Its diarization supports over 20 speakers, and the official demo also showed live transcription and speaker identification for eight people talking in the same room. According to Meta, as of September 1, 2026, the model ranked first both on Artificial Analysis for streaming speech-to-text and on public diarization benchmarks. Its online demo uses a microphone to generate transcripts; audio is processed but not stored, and the AI-generated text may contain errors.
Keda Xunfei’s Ciyuan Xinghuo has released and open-sourced the Xinghuo X2.5-4B and Xinghuo X2.5-1.7B general-purpose edge models. The company says they natively support contexts of up to 1 million tokens, while their APIs are temporarily free on the Xunfei Xingchen MaaS platform.
Keda Xunfei’s Ciyuan Xinghuo has made the model weights, code repositories, and deployment documentation for Xinghuo X2.5-4B and Xinghuo X2.5-1.7B available through Hugging Face, GitHub, and other platforms. The corresponding APIs are also live on the Xunfei Xingchen MaaS platform with free access for a limited time. According to the company, both models use a hybrid attention architecture and are optimized for Agent, coding, mathematics, and instruction-following capabilities. It describes the series as the first edge model to natively support contexts of up to 1 million tokens. The company also says Xinghuo X2.5-4B leads models of the same size in multiple tests, including SWE-Bench Pro and AIME 2026. Both models were trained on an all-domestic Chinese computing platform and support hardware from NVIDIA, Huawei, Haiguang, and Houmo. They are compatible with vLLM, SGLang, and llama.cpp, can be deployed through Ollama and LM Studio, and use the Apache 2.0 license.
Celeris launches the Celeris-1 Magnus hybrid diffusion model
模型发布
Celeris has made its hybrid diffusion model Celeris-1 Magnus available to all API customers. The company says it completed 40 of 97 tasks on τ³-Banking, with a median conversation time of 57 seconds.
Celeris has launched its second model, Celeris-1 Magnus, at inference.celeris.ai/celeris-1-magnus/v1 for all API customers. The model uses a larger hybrid diffusion architecture with native reasoning and is designed for multi-step tool use, code that must run, and tasks that require the model to retrieve information on its own. For the τ³-Banking benchmark, the company compared Magnus with OpenAI’s gpt-5.6-sol and gpt-5.6-luna and Google’s gemini-3.7-flash using the same harness, tools, and customer simulator. The first three models used the same low reasoning setting, while gemini-3.7-flash ran with its API defaults. The benchmark simulates bank customer-service conversations, with policies stored in a knowledge base accessed through retrieval tools; credit is awarded only when the model identifies and executes the exact action sequence required by a task.
According to the company, Magnus completed 40 of 97 tasks, with a median completed-conversation time of 57 seconds. gpt-5.6-sol took 79 seconds and completed 3 fewer tasks, making its per-task time nearly 30% longer and its API bill about 10 times that of Magnus. gemini-3.7-flash took 1 minute 30 seconds and completed 9 fewer tasks. Magnus uses the same OpenAI-compatible spec as celeris-1, allowing users to switch models with a one-line change. reasoning_effort is enabled by default and can be set to none for celeris-1-style direct answers or to low for the Agent configuration used in the benchmark. Celeris says celeris-1 will remain available.
Abliteration.ai releases an abliterated version of GLM-5.3
模型发布
Abliteration.ai has released abliterated-model-large-v2, based on GLM-5.3, and made it available via API. The company says the model removes weight directions that cause refusals while retaining coding, cybersecurity, and Agentic capabilities.
Abliteration.ai has launched abliterated-model-large-v2, built on GLM-5.3, and is now offering it as an API service. According to the company, abliteration first identifies directions in model activations that cause refusals and then removes those directions from the weights. The company says the processed model retains its coding, cybersecurity, and Agentic capabilities and will not refuse to continue midway through a task chain.
腾讯混元 releases a lightweight quantized version of Hy4 preview
模型发布
Tencent Hunyuan compressed the roughly 1.5TB weights of its flagship Hy4 preview model to about 214GB, a reduction of approximately 86%. Three corresponding quantized GGUF versions are now available in the AngelSlim repository on Hugging Face.
Tencent Hunyuan has released a lightweight quantized version of Hy4 preview, using the Sherry sparse ternary quantization algorithm and the MIX-STQ1_0 mixed-precision strategy to reduce the model weights from roughly 1.5TB to about 214GB. According to Tencent Hunyuan’s test results, the version remained broadly on par with the original BF16 model in long-document comprehension and multi-turn retrieval over long contexts, while its mathematics scores declined and stayed within one to two points of the original across several mainstream benchmarks. It also ranked ahead overall of the UD-IQ1_M quantization scheme commonly used by the community. The AngelSlim repository on Hugging Face provides three quantized GGUF versions; the STQ1_0 version is 213.66 GiB and requires about 214GB of VRAM for full residency. The repository states that these files cannot run directly with the original llama.cpp and require its supplied patch to be applied first.
Google launches Google Pics, an image creation and editing tool
产品应用
Google has begun rolling out its AI image tool Google Pics to all Google AI Pro and Ultra subscribers and most Workspace business customers. Docs and Slides integration is already underway, with Drive to follow over the coming weeks.
Google has started the phased rollout of Google Pics, with access opening over the coming weeks to all Google AI Pro and Ultra subscribers and most Workspace business customers. Integration with Docs and Slides began at launch, while Drive integration will roll out over the coming weeks. Google Pics is available as a standalone product through pics.new and will also be embedded in Workspace. According to Google, the tool is built on the Nano Banana image generation and editing model and supports image generation, modification of existing assets, and collaborative editing of the same image. It can be used to create posters, social media content, digital announcements, digital illustrations, product mockups, and brand concepts, while files can be added directly from Google Drive. Google says millions of users interact with billions of images across Workspace products every month, and that the integrations can reduce copying, pasting, and switching between tabs.
OpenAI launches ChatGPT gift cards, redeemable only on the web for U.S. dollar accounts
产品应用
OpenAI has published rules for ChatGPT gift cards. According to the company, the USD-denominated digital cards are currently sold by select authorized retailers and can be redeemed on the web only for U.S.-based accounts billed in U.S. dollars; balances can fund subscriptions, renewals, upgrades, and credit purchases.
OpenAI has published the purchase and redemption rules for ChatGPT gift cards in its official Help Center, stating that digital cards are currently available from select authorized retailers. Redemption and use are limited to U.S.-based ChatGPT accounts billed in U.S. dollars, and redemption must be completed on the web. According to the company, each gift card is a non-reloadable, USD-denominated stored-value card, and redemption is also subject to account balance and daily funding limits. Redeemed balances can be used for ChatGPT subscriptions, renewals, upgrades, and credit purchases, but not for transactions through the Apple App Store or Google Play. Eligible accounts include ChatGPT Free, Go, Plus, and Pro, as well as owners of ChatGPT Business workspaces.
Manus announces its official return to independent operations, with the founding team remaining in charge
行业动态
Manus announced that it has resumed independent operations under the continued leadership of its founding team and will keep advancing a general-purpose AI Agent for users worldwide. Users who backed up data during the transition but have not restored it can use the recovery portal, with no deadline according to the company.
Manus has formally resumed independent operations, with its founding team continuing to lead the company, according to announcements on its official blog and accounts. The company said it will continue investing in product innovation and advance a general-purpose AI Agent for users worldwide as an independent agent lab. According to its explanation, the product direction includes becoming more deeply embedded in users’ daily workflows, interacting more directly with the world, and acting more proactively on users’ behalf. During the preceding transition period, some users needed to back up and restore their data and experienced temporary interruptions in access. Users who completed a backup but have not yet restored their data can do so through the recovery portal. The company said there is no restoration deadline and that unaffected users do not need to take any action.
Anthropic unveiled Enterprise Frontier Safeguards, combining ZDR with cross-session misuse monitoring while keeping activity data in customers’ own cloud environments. The phased rollout will begin later in the fall, with support planned across seven Claude and partner-cloud platforms.
Anthropic announced that Enterprise Frontier Safeguards (EFS) will become available to customers in phases beginning later in the fall. The system provides zero data retention (ZDR) while storing the activity data needed for monitoring in cloud infrastructure controlled by customers rather than in Anthropic’s environment. Until EFS is ready, eligible customers will receive ZDR on Fable 5 and Fable 5.1. According to the company, EFS was developed with more than 100 customers across financial services, healthcare, manufacturing, telecom, law, retail, and the public sector, as well as Amazon Web Services, Google Cloud, and Microsoft Azure. It will support Claude Code, Claude Enterprise, the Claude Platform, Amazon Bedrock, Claude Platform on AWS, Google’s Agent Platform, and Microsoft Foundry.
According to Anthropic, sophisticated misuse can span multiple tasks, sessions, and accounts, making correlation impossible if data is deleted immediately after each interaction is analyzed. Anthropic therefore introduced 30-day data retention starting with Fable 5. The company says it has never used enterprise data for training without explicit permission and will not do so in the future. EFS lets customers manage activity data in their own cloud accounts under their own encryption keys, access policies, and audit logs, with storage options including Amazon S3, Azure Blob Storage, and Google Cloud Storage. Automated systems analyze a rolling window of traffic for attempts to develop offensive cyber or biological capabilities and for signs of stolen or leaked credentials, then send flags directly to customers. No human review by Anthropic employees is required. Anthropic says the same controls apply whether customers access Claude directly or through its three cloud partners, with data stored in the customer’s own cloud account.
Anthropic reportedly signs a $35 billion cloud computing deal with Lambda
行业动态
A person familiar with the matter said on August 31, 2026, that Anthropic had signed a $35 billion cloud computing agreement with NVIDIA-backed provider Lambda. The deal involves a roughly 350-megawatt data center in Texas.
A person familiar with the matter said on August 31, 2026, that Anthropic had signed a $35 billion cloud computing agreement with NVIDIA-backed cloud provider Lambda to rapidly expand its AI compute capacity. According to the source, the associated data center is located in Nueces County, Texas, is being developed by Hut 8, and has a capacity of about 350 megawatts. The report also said Lambda will deploy NVIDIA chips at the facility and use them to provide cloud services to Anthropic in response to growing demand for Claude.
