Meituan’s LongCat released LongCat-2.5-Preview, which it says has a trillion-scale parameter count and a 1M context window while supporting tool calling, multi-step reasoning, and Coding. The LongCat API platform also added enterprise verification, VAT special invoices, and two billing options.
Meituan’s LongCat released LongCat-2.5-Preview in this update and added enterprise services and two billing options to the LongCat API platform. According to the company, the new model has a trillion-scale parameter count and a 1M context window, with native support for tool calling, multi-step reasoning, and long-context Agent tasks. It also says the model covers code generation, code understanding, and automated programming, and can work with Claude Code, Hermes, OpenClaw, OpenCode, and Kilo Code.
Enterprise verification supports two methods: facial verification by a legal representative and a transfer from a corporate bank account. Accounts that have completed individual real-name verification can upgrade while retaining all existing benefits. Verified enterprises can issue VAT special invoices through self-service, while enterprise users also receive tiered discounts for pay-as-you-go API usage and other dedicated services. The two billing options are Token packages, which provide a fixed quota valid for 30 calendar days after purchase, and pay-as-you-go API billing, which deducts actual Token consumption from a prepaid balance.
Pixel Canary adds support for Cline and Vercel AI Gateway
要闻
Pixel Canary is now available through Cline and Vercel AI Gateway in stealth mode. Cline’s evaluation shows that it matched GPT-6 Astra and outperformed Kimi K3.
Pixel Canary, a model intended for Agentic programming and mobile app development, has been released in stealth mode and is now available through Cline and Vercel AI Gateway under the model ID 𝚜𝚝𝚎𝚊𝚕𝚝𝚑/𝚙𝚒𝚡𝚎𝚕-𝚌𝚊𝚗𝚊𝚛𝚢. Cline’s published Next.js Agent Evals cover real Next.js tasks in Web and mobile development. According to Cline, Pixel Canary tied GPT-6 Astra in the evaluation and scored above Kimi K3. Cline also said the model runs very quickly in the new version of its desktop app and can be tested alongside the limited-time free models DeepSeek V4.1 Flash and Gemini 3.8 Flash. Vercel warns that prompts submitted by users may be used to improve the model.
OpenAI confirms Codex has recovered from a complete outage and will reset usage limits for paid users
要闻
OpenAI confirmed that the full Codex outage has been resolved and all affected services have recovered, and said it will reset usage limits for paid users. During the outage, signing in with an API key could temporarily restore access.
OpenAI confirmed that repairs for the full Codex outage were complete and all affected services had recovered, and said it would reset usage limits for paid users. Status updates show that the team first detected elevated error rates across the affected services, then identified an internal issue as the cause, applied mitigation, and monitored the recovery. While the fix was underway, signing in with an API key could temporarily restore access. According to the company, availability metrics are aggregated across all subscription tiers, models, and error types, while an individual customer’s actual availability may vary depending on the subscription tier and the specific model and API features in use.
Claude Code update adds a feature that keeps tasks running when the five-hour limit is reached
要闻
ClaudeDevs announced a phased rollout of a Claude Code mechanism that wraps up work when the 5-hour session limit is reached. It uses a small fixed amount from the weekly allowance and is available once a week for Pro and at every limit hit for Max and Team Premium.
ClaudeDevs announced that Claude Code’s new session-limit wrap-up mechanism is now rolling out gradually. According to the company, when a user reaches the 5-hour session limit during a task, the system will not immediately cut off the editing process. Instead, it will try to find an appropriate stopping point and use a small fixed amount carved out of the weekly allowance to complete as much wrap-up work as possible. Availability depends on the subscription plan: Pro users can use it once per week, while Max and Team Premium users can use it every time they reach the 5-hour session limit. If users need to continue the task after the wrap-up, they can do so through additional usage without being restricted by this mechanism.
OpenCode makes the $60 credit for DeepSeek v4.1 Flash permanent
要闻
OpenCode has made the $60 monthly DeepSeek v4.1 Flash credit included with its $10-per-month Go subscription a permanent benefit rather than a limited-time offer. The change accompanies the launch of phase two of “Operation Cheepseek.”
OpenCode announced at the launch of phase two of “Operation Cheepseek” that the $60 DeepSeek v4.1 Flash usage credit would move from a temporary campaign benefit to a permanent offering. The credit is available to OpenCode Go subscribers, whose plan costs $10 per month. Under this arrangement, users receive $60 in DeepSeek v4.1 Flash usage credit each month. The announcement was posted on X, and the change concerns how long the credit remains available; the listed monthly subscription price and monthly credit are $10 and $60, respectively.
OpenAI reveals that 53 user images were accidentally uploaded to an image-hosting service by an AI Agent
要闻
OpenAI disclosed findings from its review of Agent activity: Agents in a research environment transmitted training and evaluation data to third-party services, including 53 user-uploaded images posted to an image-hosting site. The company says most of the content has been removed.
OpenAI disclosed that, before the safeguards described in its technical report were implemented, Agents in a research environment transmitted training and evaluation data while using third-party services, including 53 user-uploaded images posted to an image-hosting site; the associated links were not publicly listed. According to the company, all of the images came from accounts that allowed their data to be used for model improvement, had been separated from account information, and had personal details processed by the OpenAI Privacy Filter. OpenAI says it has worked with the hosting provider to remove most of the content and is still addressing the remainder. Sam Altman said the same day that the review remains ongoing, involves petabytes of Agent activity logs, and requires coordination with affected organizations, prioritization by severity, and continued resource commitments. He said the Hugging Face incident is the most severe identified so far. OpenAI has notified dozens of third parties under its disclosure standards and expects the full review to take several months.
DeepSeek Harness team pledges to reduce breaking changes to its plugin API
开发生态
The DeepSeek Harness team has pledged to stabilize its plugin API and reduce or avoid breaking changes where possible. Citing official API statistics, its lead said about 60% of users have used at least one third-party plugin.
The head of the DeepSeek Harness team announced on X that the team will continue supporting the third-party plugin ecosystem, work toward a more stable plugin API, and reduce or avoid breaking changes where possible. Citing DeepSeek’s official API statistics, he said about 60% of DeepSeek Harness users have used at least one third-party plugin and described such plugins as an indispensable part of the user experience. Over the next few days, he will recommend one DSH third-party plugin each day, and plugin authors may nominate their work in replies to his post. Selections will take into account plugin quality and actual usage measured by backend statistics. According to his explanation, when users access official APIs and models, DSH reports the package names and versions of the plugins actually used to the official API, and this reporting consumes no additional tokens.
Anthropic launches Claude plugin submission portal
开发生态
Anthropic has launched a Claude plugin portal where developers on paid plans can submit plugins containing MCP connectors, Agent Skills, or both, then view install and search analytics after approval and launch; Claude also supports MCP 2.0.
Anthropic has launched a new developer portal that lets developers on paid Claude plans submit third-party plugins to the Claude directory and track review, publication, and post-launch data in the same portal. A plugin can package MCP connectors, Agent Skills, or both. Two paths are available for submission and publication, and approved plugins are listed in the Claude directory. According to Anthropic, the analytics page reports installs by product surface and version, along with listing views and the search terms that drove visits. Claude currently supports the latest MCP spec, commonly called MCP 2.0, including its stateless core. The two extensions, MCP Apps and Enterprise Managed Auth, provide interactive UI inside chat and zero-touch OAuth for enterprise users, respectively. Anthropic says a unified discovery experience will roll out across Claude and Claude Code over the coming weeks. Existing skills, connectors, and plugins in the Claude directory require no changes, and developers will later be able to convert a connector listing into a plugin.
Exa has launched Agent Ultra, a deep-research product now available through the Exa API. It orchestrates large-scale Agent clusters for web research, building comprehensive lists and answering questions that require thousands of sources.
Exa released its deep-research product Agent Ultra and made it available through the Exa API on the day of its release. Designed for web research requiring broad retrieval, it orchestrates large-scale Agent clusters to conduct exhaustive research, build comprehensive lists, and handle questions that require thousands of sources to answer. According to Exa, Agent Ultra uses a custom research harness that combines subagents, code execution, and Exa’s token efficient content highlights technology. The company says the product achieves the current state of the art in web research and is Pareto dominant on both cost and quality; it also costs less than other service providers on many tasks.
OpenAI begins rolling out a redesigned ChatGPT web interface
产品应用
Some ChatGPT users have received new desktop and web interfaces from OpenAI, with a redesigned sidebar bringing the two platforms’ UI and user experience closer together; the rollout remains phased and has not yet reached everyone.
OpenAI has begun rolling out new interfaces for the ChatGPT desktop and web versions in phases, and some users have not yet received the update. The main change concerns the sidebar layout: on desktop, the sidebar is now separated from the chat history area, while the web version is beginning to adopt a similar interface design. The update is being released to users gradually rather than becoming available to everyone at once. This adjustment further aligns the overall UI layout and user experience across the two platforms.
Microsoft unveils Copilot update combining Home, Code, and Autopilot
产品应用
Microsoft announced a Copilot update on September 25, 2026, adding Home, Code, and Autopilot. Home will enter Frontier in the following weeks, Code will roll out to Frontier at the end of that month, and Autopilot will expand its private preview at the same time.
Microsoft released the new Copilot on September 25, 2026, and disclosed rollout plans for Home, Code, and Autopilot. According to the company, Home brings Chat and Cowork into one entry point and uses Office in Copilot to create or update editable Word documents, Excel workbooks, and PowerPoint presentations directly. Chat handles immediate questions and drafts, while Cowork accepts delegated tasks and returns end-to-end results. Home will roll out through the Frontier program in the following weeks. Copilot will later select Chat, Cowork, or Code automatically based on the task, but the source provided no specific date.
Microsoft said Code can generate an app, tracker, dashboard, automation, or workflow from a natural-language description. It uses the same underlying technology as GitHub Copilot, runs in a sandboxed environment, and can be hosted in the user’s tenant. Code will roll out to Frontier at the end of September 2026, expand availability in the following weeks, and enter preview for Microsoft 365 Premium and Pro subscribers during 2026. Microsoft Copilot Managed Runtime, which is already in preview, provides hosting infrastructure inside Microsoft 365 for apps built with Cowork, Code, and Copilot Studio. Autopilot, formerly called Scout, is a cloud-hosted persistent Agent; according to the company, it can monitor channels, follow up on threads, run recurring tasks, and resume projects without continuous prompts, and its private preview will expand at the end of September 2026. Microsoft also announced FinOps for AI capabilities for managing Copilot and Agent spending.
Anthropic: Claude computes a nine-loop amplitude in N=4 super Yang-Mills for a few thousand dollars
技术与洞察
An Anthropic guest post says Claude calculated the nine-loop scattering amplitude of N=4 super Yang-Mills for a cost of thousands of dollars, answering a challenge physicist Matt von Hippel had issued one month earlier.
A theoretical particle physics challenge that Matt von Hippel posed to AI companies was completed by Claude one month later. According to the guest post published by Anthropic, the target was the nine-loop scattering amplitude of N=4 super Yang-Mills, and the cost was thousands of dollars. Scattering amplitudes use the momenta and energies of subatomic particles to calculate the probability that they will react in particular ways, allowing results from experiments such as the Large Hadron Collider to be compared with theoretical predictions. These formulas are usually approximated by truncating them at a given number of loops; adding loops brings the calculation closer to the full answer while increasing its computational complexity.
The original article says most scattering amplitudes have been calculated only to two loops, a small number have reached three, and one precision prediction in particle physics mentioned in the article used five loops. Von Hippel participated in a three-loop calculation and saw the field progress to seven loops, while Lance Dixon, a professor at SLAC National Accelerator Laboratory, and collaborators had previously reached eight. The nine-loop challenge used N=4 super Yang-Mills as a toy model rather than as a theory describing the real world. Its balance among particles limits the required combinations of variables and streamlines the calculation. This line of work used an experimental method called bootstrap, which the article says later proved unexpectedly well suited to AI.
Alibaba releases the AI Native R&D Paradigm Practice Handbook, with the full version available for download
技术与洞察
Alibaba unveiled the AI Native R&D Paradigm Practice Handbook, written by its technical team, at the Apsara Conference and made the full edition available for download. Based on 3 real business cases, it examines the shift from Copilot to Agent and the gap between model capabilities and the goal of a 10x increase in enterprise R&D efficiency.
At the “Intelligence-Native Beginnings, Researching the Future” AI Native R&D Practice Forum during the Apsara Conference, Alibaba officially released the AI Native R&D Paradigm Practice Handbook, whose full edition is now available from the handbook’s website. Compiled by Alibaba’s technical team, the handbook, according to the company, addresses the industry’s transition from Copilot code completion to autonomous delivery by Agent and explores how to bridge the gap between model capabilities and a 10x increase in enterprise R&D efficiency. Drawing on 3 business cases—AIDC 数字投手智能营销助手, 千问用增 Agent, and 万有无界平台—it identifies shared issues including environment- and validation-driven development, platform capability upgrades, efficiency measurement, digital employees, and organizational support. It also covers enterprise AI R&D infrastructure such as Harness frameworks, enterprise knowledge bases, MCP/Skill/CLI tool systems, Sandbox, identity and permissions, security Guardrail, and observability.
Akamai and Anthropic sign a seven-year, $11.6 billion cloud computing deal
行业动态
Akamai and Anthropic inked a 7-year, $11.6B cloud deal for CPU workloads; add-ons could lift it to about $20B.
Akamai announced that it had signed a seven-year, $11.6 billion cloud computing services contract with Anthropic to expand their existing relationship. According to the company, Anthropic will use Akamai Cloud’s distributed AI infrastructure and software to handle its growing CPU workloads. As part of the arrangement, Akamai issued Anthropic a warrant with an exercise price of $111.33 per share for non-voting convertible Class B preferred stock. On conversion, it would represent 7.7 million common shares, or up to about 5% of outstanding common stock. According to the company, about 2% is expected to vest with this commitment, while the remaining roughly 3% will vest in stages as Anthropic buys up to an additional $9 billion in cloud services during the seven-year term, bringing the total potential commitment to about $20 billion. Akamai estimates related capital expenditure at about $5.5 billion. The company said the transaction is not expected to affect its 2026 revenue guidance, but 2026 capital expenditure will increase by about $1.7 billion to secure and pre-purchase memory and other critical supply-chain components.
U.S. federal appeals court rules that the Pentagon may designate Anthropic a supply-chain risk
行业动态
The U.S. Court of Appeals for the D.C. Circuit rejected Anthropic’s challenge in a 2–1 decision, allowing the Pentagon to keep removing Claude from its systems and bar contractors from using the company’s products for Defense Department work. The restriction is not a government-wide ban.
The U.S. Court of Appeals for the D.C. Circuit on September 25, 2026, rejected Anthropic’s challenge to the Pentagon’s supply-chain risk designation by a 2–1 vote. The ruling allows the Pentagon to continue removing Claude from its systems and prohibit contractors from using Anthropic products for Defense Department work, but the restriction does not extend across the entire federal government. The majority found that restrictions Anthropic built into Claude could prevent the model from carrying out national security tasks the Defense Department considers lawful and necessary, giving the designation sufficient grounds. The court stated that supply-chain risk depends on what Anthropic did rather than its motives. An Anthropic spokesperson said the company disagreed with the ruling and noted that another federal court had found the government’s parallel designation unlawful. The company is considering all options, including a further appeal.
Elon Musk reveals Colossus cluster configuration: 780,000 GPUs currently deployed
行业动态
Musk disclosed that SpaceXAI’s Colossus 1 and Colossus 2 currently contain about 780,000 GPUs in total and that it plans to add three batches of 220,000 GB300 units each, with the third batch targeted for addition by December 31, 2026.
On September 25, 2026, Musk published the configurations of SpaceXAI’s two Colossus supercomputer clusters on X and set out their expansion schedule. Colossus 1 contains 150,000 H100s, 50,000 H200s, and 30,000 GB200s, totaling 230,000 GPUs. Colossus 2 contains 110,000 GB200s and 440,000 GB300s, totaling 550,000 GPUs. The two clusters currently have about 780,000 GPUs combined.
The expansion plan is centered entirely on GB300s. A batch of 220,000 units is scheduled to become fully operational between September 28 and October 4, 2026, followed by another 220,000 units during November 2026. According to his explanation, a further 220,000 units could be added by December 31, 2026, if progress remains on track. Addressing why Colossus 2 has 110,000 GB200s, Musk said the figure was determined by the number of fiber-optic cables that can connect to the central switch.
