OpenAI has officially opened the Decisions API, which supports only gpt-6-luna and processes text, images, and JSON. The limit per request has increased from 200 questions to 400, and output tokens are free.
OpenAI has officially opened the Decisions API, which developers can now access through a dedicated API endpoint or try in the OpenAI platform’s Playground. The interface currently supports only gpt-6-luna and can evaluate text, images, and JSON, returning the probability that a condition holds, a preset option, or a score for content classification, request routing, and determining subsequent actions. The question limit per request has increased from 200 during public testing to 400, with native GPT-Live integration and prompt caching also added. According to OpenAI, decision-making can be up to 10 times as fast as calling GPT-6 Luna through the Responses API. Uncached input costs $0.10 per million tokens, while cached input costs $0.01 per million tokens; cache writes are billed separately, and output tokens are free.
OpenAI’s Codex lead Tibo said the usage quota reset for day 17 of the continuous improvement plan is complete. This fulfills the reset option in his earlier pledge to deliver an improvement or fully reset usage quotas each day for 28 days.
OpenAI’s Codex lead Tibo announced that the team completed a usage quota reset on day 17 of its continuous improvement plan, fulfilling the plan’s reset provision. Under his previously announced commitment, the plan runs for 28 days, with two options for each day’s delivery: release one clear improvement with practical value for most Codex and Work users, or fully reset usage quotas once. This announcement concerns the latter rather than a new feature improvement; the daily choice between the two forms part of his commitment on the team’s delivery cadence.
Anthropic announced that it is donating the Agent Skills open standard to the Agentic AI Foundation. More than 70 Agent platforms have integrated the standard, with 1 million skills published and nearly 280 million installations.
Anthropic announced that it is donating the Agent Skills standard, along with its name, logo, GitHub organization, and agentskills.io domain, to the Agentic AI Foundation (AAIF) under the Linux Foundation. The standard was opened in December 2025 and has been integrated into more than 70 Agent platforms, including ChatGPT, Cursor, GitHub Copilot, and VS Code; 1 million skills have been published, with nearly 280 million installations. Each skill is a folder containing a SKILL.md file, optionally accompanied by scripts and reference materials, that supplies an Agent with procedural knowledge for a specific task or domain. According to the company, using or contributing to the standard does not require AAIF membership, and engineers from Anthropic, Google, OpenAI, and other companies will maintain the project as individuals. Anthropic’s proposed initial roadmap includes versioned releases, a formal proposal process, and security guidance for people who write and install skills.
Anthropic-managed Claude is now generally available on Google Cloud, supporting existing IAM credentials and consolidated billing. According to the company, Fast Mode on Opus 5.5 delivers up to 2.5x faster output, with input and output priced at $8 and $40 per million tokens, respectively.
Anthropic has made its managed Claude offering generally available on Google Cloud, although the source does not specify a launch date. Anthropic operates the service and acts as the data processor, with US and Global endpoints currently available. According to the company, customers receive the full Claude Platform feature set, with new features and betas arriving on the same day as on the native platform and new models available at launch. Available models include Claude Fable 5.1, Opus 5.5, Sonnet 5.5, and Haiku 5.5. Access uses existing Google Cloud IAM credentials, charges appear on a single Google Cloud invoice, and usage counts toward existing Google Cloud spending commitments.
The offering also gives customers access to Claude Console for writing prompts, configuring agents, running them in hosted environments, monitoring usage and costs, and managing access and security by workspace. Fast Mode on Opus 5.5 is also generally available; according to the company, it offers up to 2.5x faster output token generation, priced at $8 per million input tokens and $40 per million output tokens. The Google-managed offering remains available, with Google acting as the data processor and providing native integrations such as Model Armor and Agent Engine. The company also states that customers with existing discounts need to confirm their applicability with an account executive before using the Anthropic-managed offering, as discounts cannot be applied retroactively to usage.
Google has opened an experimental preview of Gemini API CLI, currently at v0.4.x, allowing developers and Agents with shell access to call the Gemini API directly, generate media files, and integrate them into shell pipelines.
Google has released an experimental preview of Gemini API CLI (gemini-api) for developers and Agents with shell access, with v0.4.x currently available. The tool supports image, video, speech, and music generation, alongside media analysis, audio transcription, and Managed Agents management. According to Google, media generation commands save files directly to disk and return only file paths through standard output for integration into shell pipelines. The program is distributed as a stateless, zero-dependency binary supporting macOS, Linux, and Windows, and requires a Gemini API key to be configured before use.
Qoder international edition launches community support campaign, unlocking first-round rewards
开发生态
Qoder has announced Community Boost, offering eligible individual users 1,024 SOTA Credits and three reward rounds unlocked through daily community boosts. The promotion runs from October 22, 2026, through October 30, 2026, with rewards usable on Sonus.
Qoder will run Community Boost from October 22, 2026, through October 30, 2026, at 23:59, offering eligible individual users a Welcome Pack and rewards tied to cumulative community boosts. According to the company, the Welcome Pack contains 1,024 SOTA Credits, while totals of 20,000, 50,000, and 100,000 boost points unlock 2,048, 3,072, and 4,096 SOTA Credits, respectively. Members can boost once per day, with each valid boost adding 1 point. The daily allowance resets at 00:00 and does not carry over. Rewards will be issued within 12 hours after each target is reached, and cumulative points continue toward the next round without resetting. These Credits can be used with Sonus.
Individual Pro subscribers and those on higher plans are enrolled automatically, including users who subscribe or upgrade during the promotion. Free users must submit an application and receive approval; each account can apply only once, with no edits or resubmissions allowed. Lite does not currently support Sonus, so users must upgrade to Pro or higher to claim rewards. Eligibility is assessed against account status at the time of claiming. Each account can claim each pack once, manually through a supported client, without having boosted that day. The Welcome Pack must be claimed by October 30, 2026, at 23:59 and remains valid until October 31, 2026, at 23:59. Each round’s pack must be claimed within 3 days of issuance and no later than the promotion’s end, and remains valid for 24 hours after claiming. Unused Credits expire and cannot be combined with off-peak discounts.
智谱 apologizes and offers compensation for declining model response quality for subscribers
开发生态
Zhipu apologized for degraded response quality experienced by some GLM Coding Plan users and compensated all paying Coding Plan users with 3 reset cards valid through December 22, 2026. It will also provide an additional reset-card grant after a new model is released.
Zhipu apologized for degraded model response quality affecting some GLM Coding Plan users and distributed 3 reset cards to all paying Coding Plan users on the evening of October 22, 2026, as compensation. The cards are valid through December 22, 2026. Compensation is not limited to users who experienced the issue, and an additional reset-card grant will follow the release of a new model. According to Zhipu, its investigation found hardware faults in a small number of inference nodes, affecting the quality of conversation outputs routed through them. The company said all faulty nodes had been taken offline and the affected traffic rerouted to healthy nodes. It said it would continue monitoring and strengthen measures to maintain service reliability.
科大讯飞 unveils 星火 X2.6, a large model integrating cloud and on-device capabilities
模型发布
iFlytek has released Spark X2.6, an integrated device-and-cloud large model, and open-sourced a 4B-parameter on-device model, with the series available to try. Official limited-time rates per million tokens are 0.9 yuan for input, 0.08 yuan for cache hits, and 3.6 yuan for output.
iFlytek launched its integrated device-and-cloud large model, Spark X2.6, at the Global 1024 Developer Festival and simultaneously open-sourced the X2.6-4B on-device model; the series is currently available to try on the official iFlytek Spark website. According to the company, X2.6 was trained using domestically produced computing resources in China and uses a 293B-A30B MoE architecture. It supports a 1M context, strengthens coding and Agent capabilities, and allows safety policies to be customized for different task scenarios. Limited-time inference pricing per million tokens is 0.9 yuan for input, 0.08 yuan for cache hits, and 3.6 yuan for output. The on-device X2.6-4B uses a Dense architecture with 4B parameters, supports more than 200 languages and a 1M inference context, and adds three inference modes. Its inference service is free for a limited time.
Shanghai AI Laboratory has released Shusheng-A1.5 (Sage-A1.5) and made its model weights available, with 35B total parameters and 3B active parameters. According to the laboratory, the model focuses on improving agent orchestration, autonomous verification, and open-ended scientific exploration.
Shanghai AI Laboratory has released Shusheng-A1.5 (Sage-A1.5), with model weights available on Hugging Face and ModelScope, support for local deployment through frameworks including vLLM and SGLang, and a free trial access point. The model uses a MoE architecture with 35B total parameters and 3B active parameters. It natively supports a 256K context, expandable to 512K through YaRN. According to the official description, the model can assign tasks to sub-agents with mutually independent contexts for parallel execution, and can check evidence and adjust strategies during reasoning. Its capability improvements focus on agent orchestration, autonomous verification, and open-ended scientific exploration. Both orchestration and verification modes are disabled by default and must be enabled as needed.
Canva unveils Craft, its first in-house image model
模型发布
Canva has launched Craft, its first in-house image model, for users worldwide and API developers, with output at up to 2K. According to the company, Craft debuted at No. 4 on DesignArena and ranked No. 1 in two subcategories.
Canva announced Craft, the first foundation model trained from scratch by its internal AI lab, Canva Research, with worldwide availability for Canva users at launch and access through the Public API and Canva MCP Server. Users can access the model in Dream Lab, while developers can integrate it into marketing platforms, e-commerce sites, ad creation tools and print services. According to the company, Craft debuted at No. 4 among image generation models on DesignArena. It ranked No. 1 among 87 models in People & Portraits and among 85 models in Landscapes & Nature, alongside No. 2 in Abstract & Patterns and No. 4 in Products.
According to the company, Craft separates composition planning from image rendering into two stages, first specifying elements, spatial relationships, proportions, visual hierarchy and text, then generating images at up to 2K across more than 30 aspect ratios. This structure supports changes to individual elements and works with Magic Layers to turn generated images into editable design elements. Training data includes millions of human-made templates from Canva’s content library, with professional designer feedback subsequently incorporated through supervised fine-tuning and reinforcement learning. The company also says model distillation and a custom inference engine make Craft an order of magnitude faster than its first internal version. Two variants, Craft-fast and Craft-ultrafast, have yet to launch.
Ideogram has released Ideogram Vector, a vector model that generates editable SVGs from text descriptions or existing images. According to the company, the output can be used for logos, print designs, and illustrations, with colors and individual shapes adjustable in a vector editor.
Ideogram has released Ideogram Vector with two workflows: generating SVGs from text descriptions and converting existing images into editable vectors. The source does not specify a release date. Users entering a text description can optionally add a reference image, while those uploading an existing image for direct conversion do not need to provide a text description. According to the company, the results can be used for logos, print designs, and illustrations, and can scale from small labels to large posters while retaining crisp edges. Users can also change SVG colors in a vector editor and select and refine individual graphic elements to adjust the design.
Hugging Face releases open-source DNA model Carbon-X
模型发布
Hugging Face has released Carbon, a family of genomic models trained on 1T tokens, or approximately 6T DNA base pairs, alongside code for seven zero-shot DNA evaluations and task-specific fine-tuning recipes. Prokaryotes account for 10% of the pretraining mixture.
Hugging Face has published Carbon, a family of genomic causal language models, on GitHub with evaluation and fine-tuning code; the source does not specify a release date. The project is a collaboration between research teams at Hugging Face, Zhongguancun Academy, and TIGEM/University of Naples Federico II. According to the project, the models were trained on DNA and RNA sequences from the Carbon Pretraining Corpus, totaling 1T tokens, or approximately 6T DNA base pairs. The mixture targets eukaryotes, with prokaryotes accounting for 10%. Its hybrid tokenizer uses BPE for English and 6-mer for DNA, switching through an in-sequence <dna> tag.
Developers can use a single backend flag to run the repository’s seven zero-shot DNA evaluations on Carbon, GENERator, or Evo2. According to the project, this setup enables direct comparisons across model families. For zero-shot variant scoring, users feed in the full sequence and read its log-likelihood. The fine-tuning code includes a promoter detection example using Transformers Trainer and AutoModelForSequenceClassification, as well as an autoregressive DNA modeling recipe using the custom FNSTrainer. The latter employs Factorized Nucleotide Supervision (FNS), which computes base-pair-level loss for DNA k-mer tokens. The repository also supports continual pretraining by loading the model with AutoModelForCausalLM and training on sequences tagged with <dna> using next-token loss.
Ant Digital has released LingDT-Decision-1.0 for software development decision tasks in 4B and 27B versions, with API access available exclusively through DTMaaS. The model is free until October 30; the source does not specify the year.
Ant Digital has launched LingDT-Decision-1.0 for decision tasks in software development, without disclosing a release date; the free-access period ends on October 30, with no year specified in the source. The model natively supports multiple-choice and DAG decision-making, comes in 4B and 27B versions, and is available via API exclusively through DTMaaS. Each request supports a context of up to 16K, up to 32 questions, and a maximum of 255 options. According to Ant Digital’s evaluations, the 27B version has higher overall accuracy than Microsoft Decision-1 and Jev 1.13.0. The company also states that, compared with general-purpose LLMs performing the same tasks, the model uses 90% fewer tokens and responds 10 times faster.
米哈游 and partners open-source the GameScaling model
模型发布
GameScaling’s results show that more data improves play within the training corpus, while data diversity helps transfer to related worlds but offers limited gains in distant ones. Larger models learn faster, yet generalize worse to distant worlds when trained on a single world.
GameScaling reported different effects of data volume, data diversity, and model size on gameplay performance and cross-world generalization. According to its results, increasing model size speeds up learning, but when training uses only a single world, larger models perform worse at generalizing to distant worlds. On the data side, increasing volume improves play within the training corpus; increasing diversity helps transfer to related worlds, but gains remain limited when the destination is a distant world.
蚂蚁百灵 open-sources a medically enhanced model for research workflows
模型发布
Ant Bailing has open-sourced Ling-3.0-flash-Sante for medical research and workflows, with 124 billion total parameters and 5.1 billion activated per token. Its weights are available on Hugging Face and ModelScope, but it must not independently make clinical decisions.
Ant Bailing has announced the open-source release of its medically enhanced MoE model, Ling-3.0-flash-Sante, and made its weights available on Hugging Face and ModelScope. It is positioned as a research model rather than a medical device and must not independently diagnose, prescribe medication, perform triage, or make treatment decisions. The model has 124 billion total parameters, activates 5.1 billion parameters per token, and has a training context extended to 256K tokens. According to the official description, it targets medical research and workflows, with strengthened capabilities in specialized medical tasks, complex medical reasoning, in-depth research, multi-Agent collaboration, and tool calling.
Ant Bailing has released Ling-3.1-flash-1019, reporting an Intelligence Index score increase from 41 to 44 alongside gains on two Agent evaluations; API access and open weights are still forthcoming.
Ant Bailing has released Ling-3.1-flash-1019 and published evaluation comparisons with Ling-3.1-flash, with API access and open weights coming soon. According to the company, the two Agent-task evaluations show AA-Briefcase Elo rising from approximately 1400 to 1524 and the AutomationBench-AA score increasing from 62% to 68%. The company also reports that the new model’s Artificial Analysis Intelligence Index score rose from 41 to 44.
Claude to enable automatic mode by default across three subscription tiers starting next week
产品应用
Anthropic announced that auto mode will become the default for new Claude conversations for Pro, Max, and Team subscribers starting the week following the announcement. The change covers the web, desktop, and mobile apps, and users can switch back to manual mode at any time.
Anthropic announced a change to Claude’s default conversation mode, effective the week following the announcement, for new conversations started by subscribers to its three plans—Pro, Max, and Team—on the web, desktop, and mobile apps. According to the company, auto mode replaces step-by-step user approvals with automatic checks that allow safe actions, block risky ones, and have Claude look for alternative paths. The default setting does not prevent users from choosing their mode, and they can switch back to manual mode at any time.
Gemini App adds support for exporting interactive apps to Google Sites
产品应用
Google has enabled direct export of interactive apps from Gemini App to Google Sites across three platforms: web, Android, and iOS. Exports create private site drafts that retain interactivity, without requiring users to copy code or rebuild their apps.
Google has made Google Sites export for interactive apps built in Gemini App available on web, Android, and iOS. Access extends to Google Workspace customers, Workspace Individual subscribers, and personal Google Account users who can access Gemini App; users must also have Google Sites enabled and permission to create and edit sites. According to Google, exporting creates a new private site draft that preserves the app’s interactivity, without requiring users to copy code or rebuild the app. Users can add content and customize the draft before publishing and sharing it.
Grok Bot begins gradually rolling out mobile app control
产品应用
The Grok Bot team announced a gradual rollout of phone app controls on iOS and Android, covering apps such as calendars, reminders, and contacts. A permissions screen it showed lists six categories, with uses including reading and writing messages, adding calendar events, and controlling home devices.
The Grok Bot team announced that phone app controls are rolling out gradually on iOS and Android, allowing Grok Bot to use apps such as calendars, reminders, and contacts. The screen shown by the team lists six permission categories: messages, location, calendar, home, reminders, and contacts. According to the team, these can be used to read and write messages, find nearby places, add calendar events, and control home devices. A separate source, Tech Dev Notes, said Grok Bot can now access Apple Calendar, Alarms, Reminders, Location, Home, and Contacts.
1Password launches an authorized login platform for agents
产品应用
On October 22, 2026, 1Password announced a sign-in authorization platform for personal AI Agents. According to the company, after users approve a request, the platform fills credentials directly into the destination website so an Agent can continue its task without receiving passwords or one-time codes.
On October 22, 2026, 1Password announced 1Password for Agents, providing a user-approved sign-in mechanism for personal AI Agents. According to the company, when an Agent reaches a login page, it can request permission to use credentials from the user’s vault; once the user reviews and approves the request in 1Password, 1Password autofills the destination website directly, without showing the Agent the password or one-time code. Users can choose which login information to authorize and disconnect an integration at any time. Integration schedules differ by product: Meta’s Muse and SpaceXSI Grok Bot are scheduled to roll out across all platforms over the days following the announcement. Integrations for Anthropic’s Claude Duet, OpenAI ChatGPT Work and dots are coming soon, while Town Townie’s web and iOS integrations have yet to launch. The company is also offering the platform to personal Agent developers to incorporate sign-in approval flows.
Claude writes safety proofs for 123 open-source libraries in weeks
技术与洞察
Anthropic and Theorem used Claude to generate formal proofs of security properties for 123 open-source libraries within weeks. According to the official account, the proofs cover safe execution under specific conditions, rather than establishing that the software is free of every type of vulnerability.
Anthropic disclosed that its collaboration with Theorem used Claude to complete proofs of security properties at the RISC-V assembly-code level for 123 open-source libraries within weeks. According to the company, the programs covered run as single-threaded static binaries with fixed memory budgets on a RISC-V CPU using Linux OS. Under those assumptions, for every input smaller than 64 MB and every modeled Linux OS response and memory layout, the programs exit normally without memory-safety errors or assertion failures. This does not rule out every possible vulnerability. The project automatically converts existing compilable test suites into specifications, using the requirement that the tests always pass as a rule. According to the official account, Claude also found vulnerabilities while writing the proofs that had remained undetected despite trillions of test runs. Anthropic is publishing a subset of the proofs, together with the RISC-V and OS semantics, for external verification and reuse; Theorem is also offering the same pipeline for other open-source projects and publishing a Claude skill to help developers verify their code.
Evaluation reveals a past vulnerability in OpenAI's internal monitoring
技术与洞察
METR disclosed that employees from METR and Redwood Research found multiple vulnerabilities in OpenAI’s internal monitoring system during a three-week test, producing a 61-page report. OpenAI said all identified vulnerabilities affecting its production safeguards had been patched, but the evaluators have not verified that claim.
On October 22, 2026, METR disclosed that employees from METR and Redwood Research had discovered multiple previously unidentified vulnerabilities in OpenAI’s then-existing internal misalignment monitoring system during a three-week test with extensive access. The test focused on estimating the proportion of Agent inference being monitored and whether Agents could, in principle, disable or subvert monitoring. It assessed the system as it existed at and before the start of the exercise, excluding changes underway at the time, and did not establish the expanded system’s current coverage. OpenAI said it had patched all identified vulnerabilities affecting its production safeguards stack and subsequently changed its intended monitoring coverage; the evaluators have not verified those claims.
The two employees conducting the test could inspect internal code, monitoring prompts, configurations, and documentation, and interviewed approximately 12 researchers and engineers. They used OpenAI devices that could not leave the premises and were accompanied while onsite. The resulting 61-page internal report was shared with OpenAI research and security staff, while a 38-page version redacted by OpenAI was shared with METR and Redwood leadership. The pilot was intended to explore giving external researchers access to internal infrastructure, and the three organizations agreed in advance to publish only a brief overview. OpenAI had the right to remove information it considered confidential or sensitive, and the evaluators had no unilateral right to disclose those redactions.
Google and Alphabet CEO Sundar Pichai said Gemini 4 Argon processed almost 25 trillion tokens in its first 72 hours, roughly three times Gemini 3’s volume over the equivalent period. Google faces compute constraints and is developing more efficient versions to expand access.
Google and Alphabet CEO Sundar Pichai said in a Sources interview following Gemini 4 Argon’s wider release that Google is developing more efficient model versions to address compute constraints and expand access. According to Pichai, Argon processed almost 25 trillion tokens in its first 72 hours, roughly three times the volume Gemini 3 processed over the same window the previous year. He acknowledged that Google needed to accelerate its release cadence this year and expects that pace to continue with Gemini 4. On consumer Agents, he said Google is bringing the capabilities behind its enterprise Agents to Gemini, including around-the-clock task execution, with conversations, scheduled tasks, and background work handled through a single entry point. He said access to these capabilities would expand over the next few months. He also said Agents could carry user context across Google products, allowing users to start a task in one product and continue it in another.
Microsoft discloses research on AI diagnosis in emergency care
技术与洞察
Microsoft AI disclosed sequential-diagnosis research in which MAI-DxO paired with OpenAI o3 achieved 85.5% accuracy on a benchmark comprising 304 NEJM cases; the 21 physicians in the comparison group averaged 20% accuracy across completed cases.
Microsoft AI released evaluation results for its diagnostic orchestration system MAI-DxO and the Sequential Diagnosis Benchmark (SD Bench), stating that MAI-DxO improved diagnostic performance for every model tested. SD Bench converts 304 NEJM cases into stepwise diagnostic tasks in which participants can ask questions, request tests, and revise their judgments as new information arrives, with final diagnoses compared against those published in NEJM. Each test also incurs a virtual cost, allowing accuracy and resource expenditure to be measured together. The foundation models tested included GPT, Llama, Claude, Gemini, Grok, and DeepSeek. According to the company, MAI-DxO simulates collaboration among multiple physicians, checking costs and verifying its reasoning before deciding what to do next; its pairing with OpenAI o3 achieved the highest accuracy, at 85.5%. The comparison group comprised 21 practicing physicians from the United States and the United Kingdom with 5–20 years of clinical experience, who averaged 20% accuracy across completed cases.
