Suleyman is CEO of Microsoft AI, and his post summarizes a Science paper titled Sequential diagnosis with language models, which Science lists as accepted August 10, 2026 and published in print October 22, 2026. The paper calls the system DxO for Diagnostic Orchestrator, while Microsoft's blog calls it MAI DxO. According to the paper, all authors are Microsoft employees who hold company equity, and no outside funds were used.

Suleyman's post says the MAI Diagnostic system is now 1.7x more accurate than physicians in the emergency department, even when they have access to the web, that it costs less and orders fewer medical tests, and that the work is published in Science. The attached image is the cover of the October 22 issue, themed AI in Medicine. Microsoft's blog post of the same date describes MAI DxO as working like a panel of virtual physicians that propose diagnoses, choose questions and tests, and challenge the emerging view over several rounds.

The paper reports results on SDBench, a benchmark built from 304 New England Journal of Medicine clinicopathological conference cases and 325 emergency department to admission encounters from MIMIC IV ED. On the emergency cases, DxO with GPT 5 reached 80.3% accuracy against 46.8% for participating generalist physicians, which is the roughly 1.7x ratio. The paper's figure caption labels those emergency physician results as open book, and elsewhere describes open book as having search and reference tools. On the NEJM cases it reported 83.2% against 36.1% for physicians with search tools and 22.3% without.

On cost, the paper is more nuanced than the post. The unconstrained configuration that scored 80.3% had a mean testing cost of $4979 per emergency case, against $3400 for physicians. Budget constrained configurations scored 73.2% at $2634 and 69.5% at $2037, both below the physician figure. Costs are estimates from a Kaiser Permanente Washington price file.

The authors say these are controlled, text only comparisons with individual generalist physicians and do not establish superiority over specialists, multidisciplinary teams or real world practice. They also say the benchmark has no imaging or waveforms, uses only pathological cases, and measures diagnosis alone, not treatment, safety or outcomes. All authors are Microsoft employees.

Microsoft says it will test the system in a prospective study with Beth Israel Deaconess Medical Center in Boston, where patients interact with the system directly and performance is measured against physicians' independent assessments. It also says it is open sourcing SDBench for select NEJM cases so other researchers can reproduce and challenge the evaluation, and the paper says the emergency case subset will be released on PhysioNet subject to credentialing. Watch for independent replications, comparisons against specialist teams, and whether any clinical deployment follows. Microsoft's own page says the true test is performance in real clinical environments.

Formal verification means mathematically proving that code follows a set of rules, and it has been costly because proofs take expert effort and are often far larger than the code. Anthropic's Frontier Red Team worked with Theorem, a company that supplied formal models of the processor and operating system used in the proofs.

On October 22, Anthropic said it used Claude to prove security properties in more than one hundred real world open source libraries over a few weeks, in collaboration with Theorem. All figures below come from Anthropic's own write up, and the write up says to Anthropic's knowledge this is the first large scale effort to apply AI formal verification to production software at the assembly code level.

Anthropic says Claude proved that fuzz test entry points never crash for any input under 64 MB, across 239 entry points in 123 libraries. The proofs cover compiled RISC V binaries, total about 180 million lines in the Rocq proof language, and a typical proof runs to about 600,000 lines. Anthropic says a Rocq checker validates the proofs, so the people who rely on them need to trust the formal models of the processor and Linux, the exact theorem statements, and the checker.

Anthropic says the work also found bugs, including a rediscovery of a known libexpat vulnerability, which the write up identifies by its CVE number and an out of bounds write that appeared only on rare inputs of around 20 MB. It says Claude patched the bugs it found and that it audited the patches for shortcuts.

The write up lists limits. Formally verified does not mean bug free, the proofs assume a single threaded static binary on a RISC V processor under Linux, proofs can take hundreds of millions of tokens, and Anthropic has not published every proof. It says it is releasing a subset of the proofs and the models so others can check them.

Watch for independent checks of the published proofs and for security researchers to test whether the stated assumptions and trusted components hold up. Anthropic says the main gap is extending the models beyond single threaded static RISC V binaries to the x86 and ARM systems most people run. Cost is the other open question, since proofs took very large amounts of compute. Theorem says it will run the same pipeline for other open source projects, which would show whether results carry over.

MiniMax posted at 9:06 PM Eastern on October 22 that M3.1 Flash weights are public, with links to Hugging Face and ModelScope. The Hugging Face model card describes a native multimodal mixture of experts model with a context window of up to 1M tokens. The post is a company announcement, and the benchmark results in it are MiniMax's own claims. The post quotes MiniMax's earlier announcement of the model.

MiniMax said on X that the weights for MiniMax M3.1 Flash are now public, with 428B parameters in total and 23B activated per token. The post lists scores of 76.2 on BankerToolBench, 73.7% on DeepSWE v1.1, 61.7 on OSWorld V2, 29.6 on Agents' Last Exam and 82.2 on CyberGym. These are MiniMax's own figures. It also links to its API, its M Plan and MCode for desktop.

The Hugging Face model card describes the model as built for coding, agentic workflows, computer use and cybersecurity. It accepts text, image and video input, produces text, and supports up to 1M tokens of context. According to the card, M3.1 Flash builds on MiniMax M3 with sparse attention throughout the model, quantization aware training and DSpark speculative decoding, and the release includes the DSpark draft head.

The card's comparison table sets MiniMax's own results next to DeepSeek V4.1 Flash, Kimi K3, GLM 5.3, GPT 5.6 Sol and Opus 5. On several rows the table shows at least one rival scoring higher than M3.1 Flash, for example DeepSWE v1.1 (77.0 for DeepSeek V4.1 Flash 1017 against 73.7) and OSWorld 2.0 (65.7 for GPT 5.6 Sol and 70.6 for Opus 5 against 61.7). The card notes that results are for the official weights and that third party quantizations may differ. Its evaluation notes list the agent harness versions MiniMax used.

The card recommends SGLang, vLLM and TokenSpeed for serving. The Hugging Face page's automatic size panel shows 220B parameters, which differs from the 428B in the card and post; the page does not explain why.

The card lists integrations for Transformers, KTransformers, unsloth and ATOM as still in progress. Independent evaluations of MiniMax's benchmark claims had not been published in the sources reviewed. The card says third party quantizations may behave differently from the official weights, so early community reports on smaller builds may not match the figures above. The licence terms for commercial use were not reviewed here and should be read on the model page before deployment. MiniMax lists model@minimax.io as its contact for the release.

Bloomberg first reported the round citing unnamed sources, and Reuters and The Wall Street Journal later reported similar figures from their own sources. DeepSeek, Tencent and CATL had not commented, so the deal is reported, not officially announced.

Shortly after midnight on October 22 Eastern time, Bloomberg posted that DeepSeek has raised about $15 billion in its latest round, one of China's largest ever and a sign of its ambition to compete globally with Anthropic and OpenAI. zerohedge had posted the headline earlier.

Reuters then reported, citing two people with knowledge of the matter, that DeepSeek is wrapping up a round of about 100 billion yuan from existing and new investors, with CATL and Tencent among the biggest. One person said the round should close in the coming days but that details could still change. Reuters said that would double the original target of 50 billion yuan. The Wall Street Journal reported the $15 billion figure as a second private round and said DeepSeek aims for a Shanghai listing early next year.

TechCrunch noted the reported round came two days after the US ordered federal civilian agencies to remove DeepSeek technology, and that it follows a June round of more than $7 billion at a $50 billion valuation. No valuation for the new round was given in the coverage found, and none of the companies had commented.

Watch for DeepSeek, Tencent or CATL to confirm the round and its valuation, and for the round to formally close in the coming days as Reuters's source expected. Also watch the planned Shanghai listing timeline, any new US restrictions on DeepSeek, and how the money changes its compute spending and model release pace against Chinese and US rivals. A full list of investors would show how much state backed money took part.

Madison Mills is a senior AI reporter at Axios, and this is her own scoop based on unnamed industry and bank executives. Banks have not been named, and the scale of the activity is not given.

On October 22, Axios reporter Madison Mills posted a scoop that major US banks are detecting unauthorized third party AI agents inside customer accounts, including some attempting to take actions that appear to be independent of user instructions, according to sources who spoke to Axios.

She wrote that the findings have alarmed banking executives, who are racing to restrict access to their systems as AI moves faster than financial institutions can respond, according to multiple industry and bank executives. Several banks are preparing for worst case and even day after scenarios, including how to quickly secure or restore customer accounts if things go wrong.

Several bank executives told Axios they struggle to tell human users from AI agents. One said banks also struggle to determine whether agents are authorized by customers or may be operating without their permission, which the executive likened to rogue agent behavior.

The report does not name the banks, the agents or the companies behind them, and does not say whether any money was lost. unusual_whales relayed the scoop, credited to Axios. The Axios article could not be loaded for this summary, so details come from Mills's post.

Watch whether banks name the agents involved, block agent access, or publish rules for customers who want to use AI agents with their accounts. Regulators may weigh in on who is liable when an agent moves money. Also watch for agent makers to add ways for banks to verify that an agent is acting with a customer's consent. A follow up from Axios naming specific banks or agents would show how widespread the problem is.

Max Zeff is a Wall Street Journal reporter, and the post shares a WSJ story he wrote with Keach Hagey and Berber Jin. Jasmine Wang, Tomek Korbak and Mikita Balesni were fired by OpenAI in early October, and on October 9 Reuters reported OpenAI saying an investigation found they violated policies on handling sensitive information. The new WSJ reporting adds the phone checks and the Signal chats. The post was published at 9:13 PM Eastern on October 22.

Max Zeff posted at 9:13 PM Eastern on October 22 that the three researchers OpenAI fired spent the summer fighting to keep an AI system's chain of thought visible, that OpenAI then asked to see their Signal chats, and that a week later they were let go. The attached image is an excerpt of the WSJ story.

Per the excerpt, OpenAI's security team called the three researchers into meetings last month and asked them to unlock their personal phones. The WSJ says the company suspected them of leaking information about its models' internal designs to an external AI safety researcher. Security officers saw that some of the researchers had turned on disappearing messages for Signal chats with that person. A week later, all three were summoned to simultaneous 2 p.m. meetings and fired. Korbak is quoted saying he was told there was a breach of trust because he dealt with a third party through unsanctioned channels.

The Verge, summarizing the WSJ, says the chat was with Ryan Greenblatt, a former chief scientist at Redwood Research, and that OpenAI reportedly accused Greenblatt of leaking details about its Astra model to The Information. The Verge says Greenblatt has denied that. India Today, also citing the WSJ, says the leak probe concerned Astra's opaque recurrence, which could make the model harder to monitor.

OpenAI told the WSJ the researchers were fired for violating clear policies for handling sensitive information and not for raising safety concerns or speaking out. On October 9 OpenAI said an investigation found a significant breach of trust beyond what the researchers' own letter described, and the researchers said they were not the source of The Information's article. OpenAI has not published details of the alleged violations, and the researchers' motives and the underlying leak are unverified.

Watch for any OpenAI response to the new details, and for whether the company publishes specifics of the policy violations it alleges. Casey Newton posted on October 23 that the three researchers gave their first podcast interview, which may add their account. Also watch for reactions from OpenAI's board and safety committees, which the researchers addressed in their October letter, and for whether outside safety evaluators such as Redwood Research and METR comment on the access disputes. None of these steps has been confirmed.

Agent Skills lets developers package instructions, scripts and reference files in a folder with a SKILL.md file so an agent can learn a task. The Agentic AI Foundation sits under the Linux Foundation and was cofounded by Anthropic, Block and OpenAI, according to Anthropic.

On October 22, ClaudeDevs posted that Anthropic is donating Agent Skills to the Agentic AI Foundation, the same foundation it donated MCP to, and that more than 70 agent platforms have integrated the standard since December 2025, including ChatGPT, Cursor, GitHub Copilot and VS Code.

Anthropic's announcement says it is handing over the Agent Skills name and logo, the GitHub organization and the agentskills.io domain, and that people can keep using or contributing to the standard without joining the foundation. It says 1 million skills have been published and nearly 280 million installed. Engineers from Anthropic, Google, OpenAI and other companies will maintain it as individuals, with David Soria Parra and Mahesh Murag as lead maintainers. Anthropic proposed a first roadmap covering versioned releases, a formal proposal process and security guidance for skill authors.

The Linux Foundation said the foundation added four projects at once: Agent Skills from Anthropic, Open Responses from OpenAI, OpenEnv from Hugging Face and OpenSandbox from Alibaba, bringing it to 10 projects in its first year. Boris Cherny of Anthropic posted that Agent Skills started as an experiment in Claude Code.

Watch for the first versioned release of the standard and the security guidance for people who write and install skills, since installed skills can run scripts. Also watch whether Google, OpenAI and others take active maintainer roles, and how the foundation coordinates Agent Skills with MCP and its other new projects. Adoption numbers beyond the 70 platforms will show whether neutral governance speeds uptake.

The post shown is a short relay of a CNBC report, and OpenAI has not confirmed the round in the sources reviewed. CNBC says it is based on people familiar with the deal, who asked not to be named. OpenAI has been reported to have raised $122 billion in March at an $852 billion valuation.

This is a report, not an announcement. The lead post is from zerohedge and relays a CNBC headline. It is secondhand, and the facts below come from CNBC's article, published at 4:15 PM Eastern on October 22 and updated at 4:41 PM, and from other outlets summarizing it.

CNBC reports that OpenAI is close to finalizing a round of up to $30 billion at a $1.25 trillion pre money valuation, meaning the figure does not include the new capital. It says investors from the United Arab Emirates are leading alongside Coatue, and that the financing is being positioned as an opportunistic bridge ahead of a possible IPO. According to CNBC, OpenAI told investors on Thursday it is aiming to secure signatures by the end of next week, and the final size is still being determined. A person familiar also told CNBC that OpenAI is negotiating a commercial deal to secure more computing capacity from the UAE.

OpenAI has not confirmed the figures. TechCrunch says OpenAI declined to comment, and notes that terms are still being negotiated and could change. Bloomberg reports that the valuation discussed had shifted by $150 billion from $1.4 trillion, citing people involved in the talks, and that details could still change.

For context, CNBC says OpenAI confidentially filed its prospectus with the Securities and Exchange Commission in June without saying when it plans to debut, and that its chief financial officer told employees in August that OpenAI will be a public company in 2027. TechCrunch says the March raise valued the company at $852 billion post money.

Watch whether OpenAI or its investors confirm the round, its final size and valuation, and whether the signatures arrive by the end of next week as CNBC reports. The reported UAE compute deal is another item to watch, since neither company has described it. Also watch for a filing date for the IPO, which CNBC says OpenAI has not disclosed. Until OpenAI confirms, the figures remain unnamed source reporting from CNBC, Bloomberg and others.

Reuters is relaying a report from Maeil Business Newspaper that cites industry sources, and the companies had not confirmed the amounts in the coverage found. FC BGA substrates are the base layers that high end processors such as AI accelerators sit on.

On October 22, Reuters posted that Samsung Electro Mechanics gets a $2.8 billion chip investment from Tesla and Nvidia, according to a report. Its story says the South Korean company will receive investment valued at 3.7 trillion won to build facilities making FC BGA substrates for AI chips, citing Maeil Business Newspaper, which cited industry sources.

Per Maeil, Tesla will put in about 1.1 trillion won toward a dedicated line at the Sejong plant in exchange for guaranteed supply from 2029 through 2033, for chips used in its self driving cars and Optimus robots. Nvidia will put in about 2.6 trillion won for expansions at Sejong and Busan. Maeil said Nvidia presented the partnership at GTC in Germany on October 21 without naming an amount, and that the paper confirmed the figure through its own reporting.

Maeil also reported that mass production is planned for the second half of 2028, and that the company's chief executive said in March that server substrate demand is at least 50% above capacity. No first hand Tesla, Nvidia or Samsung Electro Mechanics post confirming the dollar amounts was found.

Watch for Samsung Electro Mechanics, Tesla or Nvidia to confirm the amounts in a filing or statement. Also watch whether Nvidia's reported advance payments to Ibiden and capacity deal with Unimicron are confirmed, and whether other chip buyers follow with prepaid supply deals. Construction progress at Sejong toward 2028 mass production will show whether the substrate shortage eases. Rival suppliers such as Ibiden and Unimicron may announce their own expansions.

Claude was already sold on Google Cloud through a Google managed offering, where Google is the data processor. The new Anthropic managed option runs the Claude API on Anthropic's infrastructure while billing and access go through Google Cloud.

On October 22, the Claude account announced that Claude Models on Google Cloud (Anthropic managed) is generally available. It said customers get the full set of Claude Platform features with Google Cloud IAM credentials, billing and commitment retirement.

Anthropic's announcement says new features and betas ship the same day they go live on the native Claude Platform, and lists Claude Managed Agents in beta, code execution, computer use, agent skills, the Files API, citations, web search, web fetch, the MCP connector, prompt caching and batch processing. It says the offering is available with US and Global endpoints, billing is through a single Google Cloud invoice, and usage draws down an existing Google Cloud commitment. Anthropic advises customers with an existing discount to contact their account executive first, because discounts cannot be applied retroactively.

Google Cloud's blog says both options deliver the same models on day zero at the same pricing, and the difference is who acts as data processor and when first party tools arrive. Google's documentation adds that data storage and inference happen on Anthropic managed infrastructure, which may route outside Google Cloud, and that a per request inference_geo setting can pin inference to a geography. Anthropic says the Google managed option remains the choice for customers who need Google Cloud as processor, strict regional data residency, or native tools such as Model Armor or Agent Engine.

Watch which large Google Cloud customers move to the Anthropic managed option, and how many keep the Google managed one for data residency reasons. The routing language in Google's documentation, which says inference may leave Google Cloud, is likely to matter to regulated buyers. Also watch whether the same day feature promise holds as Anthropic ships new models and betas, and whether Amazon and Microsoft respond with similar arrangements.

The lead post is from Romain Huet, who describes himself as head of developer experience at OpenAI, so it is an announcement from a company employee. OpenAI's own page, dated October 22, says the Decisions API is generally available with GPT 6 Luna through a dedicated endpoint. The API returns a probability, a choice or a score for questions a developer defines. OpenAI's page says input costs $0.10 per million tokens and there are no output token charges.

Romain Huet posted at 4:37 PM Eastern on October 22 that the Decisions API is generally available, and that to celebrate he had Codex reimagine a very important app, with a hot dog emoji. The post does not name the app, and the attached two minute video is not described in the text.

OpenAI's announcement says applications and agents make many small decisions, such as classifying an input or routing a request, and that latency and cost add up at scale. It says the API focuses GPT 6 Luna on questions with bounded answers and works on text, images and structured inputs up to 10x faster than GPT 6 Luna through the Responses API. That speed figure is OpenAI's own.

Since the public beta, OpenAI says it added a native GPT Live integration, prompt caching, inference residency in the US and EU, native JSON inputs, images in question instructions, and a limit of 400 questions per request, twice the previous limit. The three answer types are predicate, choice and score.

OpenAI's page quotes customers. Box reports about 5x faster route selection on 30 certificate of insurance cases, Datadog reports 40x faster p95 latency than available frontier models with comparable accuracy, Ramp reports 4x faster and 30% cheaper than GPT 6 Luna, and Sierra reports under 280 ms p95, and Speak reports a drop from 1.4 s to 0.13 s in time to next action with GPT Live. These are customer statements published by OpenAI, not independent measurements. OpenAI lists input at $0.10 per million tokens and cached input at $0.01.

OpenAI's page says the API is available with GPT 6 Luna through the dedicated Decisions endpoint, and lists no other model. Independent measurements of the speed and cost claims were not found in the sources reviewed, and the customer figures come from OpenAI's own page. OpenAI says it continues to work on latency, including added capacity in several regions in Asia. Details on caching for 30 minute and 12 hour writes are in OpenAI's pricing guide. Whether more models are added to the endpoint is not stated.

Craft was built and trained from scratch by Canva Research, the company's in house AI lab. DesignArena ranks image models through blind crowdsourced comparisons, and its live leaderboard is linked.

On October 22, Canva introduced Canva Craft, saying image generation is now available to everyone, powered by its first frontier model built specifically for design. It said Craft makes Canva the world's second most advanced AI lab for image generation on DesignArena's global leaderboard.

Canva's newsroom post says Craft debuted as the fourth most advanced image model on DesignArena and ranks first for People and Portraits and for Landscapes and Nature, second for Abstract and Patterns and fourth for Products. The lab ranking counts the best model per company, which is how a fourth place model makes Canva the second lab. Capital Brief independently reported the fourth place debut and said Canva Research ranks second among labs, behind OpenAI. The second lab framing is Canva's own, and the ranking was not checked live against the leaderboard here.

Canva says Craft uses a planning model that writes a composition plan, then a rendering model that produces images up to 2K across more than 30 aspect ratios. It is available to everyone in Canva, according to the company.

Watch whether Craft holds its DesignArena position as more votes come in and rivals ship new models. Also watch how Canva prices Craft through AI credits and its API, and whether it builds further models such as video or layout generation. Competition with Adobe Firefly and OpenAI's image tools for designers is the larger test. Independent leaderboard snapshots over the coming weeks will show whether the debut rank was durable.

Ant Ling is the large model team at Ant Group, which publishes models under the inclusionAI name. Artificial Analysis independently lists the model at 44 on its index, but the agentic gains in the post are Ant's own figures.

On October 22, Ant Ling introduced Ling 3.1 Flash 1019, saying it scores 44 on the Artificial Analysis Intelligence Index, up from 41 for Ling 3.1 Flash. It claimed gains in agentic work, with AA Briefcase rising from about 1400 to 1524 Elo and AutomationBench AA from 62% to 68%, and said API access and open weights are coming soon.

Artificial Analysis confirms the headline number. Its model page lists Ling 3.1 Flash 1019 at 44, and its changelog lists the earlier Ling 3.1 Flash at 41. A separate model routing blog, Orca Router, noted that Artificial Analysis lists no public API and a proprietary license for now.

The Briefcase and AutomationBench numbers come from Ant's post and were not checked against Artificial Analysis's detailed results here. Because no weights or API are public yet, outside developers cannot test the model themselves. The chart in the post places it near other models scoring in the mid 40s.

Watch for Ant to release the open weights and API it says are coming soon, which would let developers check the results. Also watch the license terms when weights arrive, and how it compares with other Chinese models scoring in the mid 40s once pricing is known. A detailed Artificial Analysis breakdown would show where the gains come from. Rival Chinese labs may answer with their own updates.

Kevin Weil previously led product and science efforts at OpenAI and is now Warp's chief executive. According to Bloomberg, the round values the company at $865 million after the money.

On October 22, Kevin Weil introduced Warp Sciences, a new company applying AI and robotics to help develop medicines faster. He said it raised a $215M Series A led by Lightspeed to bring frontier AI out of the browser and into the lab, and to build scientific capacity in America. He named his cofounders as DJ Patil, FAIR cofounder Manohar Paluri and drug hunter Sol Reisberg.

Bloomberg reported, as relayed by Investing.com, that IVP, Premji Invest and Lux Capital also joined, putting the post money valuation at $865 million. Per that report, Warp is targeting the labor intensive step where chemists make hundreds or thousands of variants of a promising molecule, work drug makers have largely outsourced to Chinese firms such as WuXi AppTec. Warp plans fully automated labs, starts with about 20 people, and its chemists will wear cameras and sensor gloves to collect data for its own AI models, according to the report.

Business Wire carried the company's announcement describing its focus on US chemistry services for drug research. The claims about speed are the company's own goals, and no results have been published yet.

Watch for Warp to open its first automated lab, sign drug makers as customers, and publish any data showing faster or cheaper molecule work. Also watch how US policy on Chinese contract research firms shapes demand. Rival AI for science startups may raise similar rounds as investors back robotic labs as a source of training data. Any partnership with a large drug maker would be an early sign of demand.

Auto mode first became the default in Claude Code for Pro, Max and Team plans in August. Instead of asking for approval at each step, an automated classifier reviews actions and blocks risky ones, according to Anthropic.

On October 22, the Claude account posted that starting next week, sessions across the Claude web, desktop and mobile apps will run in auto mode by default on Pro, Max and Team plans. Rather than asking users to approve each step, an automated check blocks risky actions, and users can switch back to manual at any time.

Anthropic's article, updated the same day, says auto mode is rolling out next week as the default for new sessions in most Claude apps on those plans. Safe actions go ahead and risky ones are blocked while Claude looks for another route. Enterprise admins will get an email before auto mode becomes the default for their organizations.

The same article describes the earlier change in Claude Code, where auto mode became the default for new sessions on August 14. That change applied only to Claude Code, while next week's update extends the same default to the main Claude apps. The post does not list which apps are excluded from the word most.

Watch for the rollout next week and whether users report blocked actions or unwanted ones getting through. Enterprise admins should look for Anthropic's email before the default reaches their organizations. Also watch whether Anthropic publishes data on how often the classifier blocks actions in the apps, and whether rivals make similar automatic approval the default. The list of apps excluded from the rollout is another detail to watch.

Grok Bot runs tasks on its own cloud computer, and the mobile app uses the same Bots and conversations as desktop. The new feature asks the phone's operating system for permission to reach built in apps.

On October 22, the Grok Bot account posted that Grok Bot can now use Calendar, Reminders, Contacts and more apps on a phone, rolling out now on iOS and Android. The attached image shows a permission list with an Allow button beside Messages, Location, Calendar, Home, Reminders and Contacts.

TeslaNorth reports the announcement came at 3:14 PM Pacific and says SpaceXSI did not say whether every app on the list is available on both platforms, or which app version is needed. It also describes recent phone features for the Bot earlier this week, including an iPhone Action Button shortcut and handling X direct messages. A separate story in this deck covers the report of Grok Bot arriving in X Chat.

AlphaSignal says the feature requires updating to the latest app, may reach accounts gradually, and relies on operating system permissions that users can revoke in phone settings or the app. It describes likely uses as finding free time and creating calendar events, listing and completing reminders, and looking up contacts. Those uses are AlphaSignal's description, not details published in the post.

SpaceXSI's mobile documentation says the app needs iOS 18 or later, or Android 9 or later, and an eligible plan, and that work continues in the cloud when the app is closed. This is a first party announcement, and the post does not cover how phone data is processed or retained.

Watch whether the rollout reaches every account and whether the full permission list shown in the image, including Messages and Home, becomes available on both iOS and Android. Details on how phone data is processed and stored would help users judge the privacy tradeoff. Also watch how reliably the Bot edits calendars and contacts in real use, and whether rival assistants from Apple, Google and OpenAI respond with deeper phone access.

Grok Bot is an AI agent that works on its own cloud computer and often reaches sign in pages. 1Password describes the feature as agentic autofill, where the password manager fills a login on the agent's behalf after the user approves it.

On October 22, the Grok Bot account posted that Grok Bot can now use a user's 1Password account. Users approve the logins they want to share, and the Bot signs in for them. The post says the integration is rolling out to users over the next few days, so not everyone has it yet.

1Password's page for the integration says 1Password fills a saved login so Grok Bot can sign in and continue a task, and that Grok Bot does not see the password or the one time code. It says Grok Bot can request specific logins that the user approves, and that connecting does not give it access to the whole vault. It also warns that signing in can give the Bot access to information and actions inside that account, so users should review the account and task before approving.

1Password's general agents page says the approval happens on the person's own device, and that when several saved logins match, the person can choose which one is used. It lists Grok Bot and Muse as agents rolling out over the coming days across all platforms.

On eligibility, 1Password says the integration is included in Individual and Families plans on 1password.com. Accounts on 1password.eu and 1password.ca, and Business and Team Starter Pack accounts, are not currently eligible. The security claims here are 1Password's and the Grok Bot account's own descriptions and were not independently tested.

Watch whether the rollout reaches all eligible users over the next few days, and whether 1Password extends it to business and regional accounts, which it says are not eligible now. Independent security reviews of how approvals and fills work would help test the claim that the agent never sees credentials. Also watch how other agent makers and password managers respond, since 1Password lists other agents on the same page.

SemiAnalysis is a semiconductor and AI research firm, and its post is a third party claim about Anthropic, not an Anthropic announcement. Anthropic's Max plan help page describes five hour and weekly usage limits and says Anthropic may apply other limits at its discretion. The post is marked as part one of two, and the second post was not retrieved for this entry. No Anthropic statement confirming or denying a test was found as of this scan.

SemiAnalysis posted at 5:36 PM Eastern on October 22 that it detected an A/B test that lowers monthly API value by about 20% while raising five hour limits by about 20%, affecting only some accounts. It says many users on X had been reporting lower limits.

The attached chart is sourced to the SemiAnalysis Tokenomics Model. It compares two Max 20x accounts, using the same model, Opus 5.5, measured on the same day. In its figures, Account A has a five hour window of 1,740 and Account B has 2,076, a difference of 19%, measured in millions of agentic tokens. The weekly budgets are 6,579 for Account A and 5,266 for Account B, which is 20% lower. The chart converts these into a monthly API equivalent value of $11,726 for Account A and $9,385 for Account B.

Those numbers are SemiAnalysis's own measurements from two accounts, so the sample is small.

SemiAnalysis's October 5 newsletter, which is marked as paid, describes its general method of measuring how far each provider's usage meter moves per token type. In that piece it also says one unnamed provider confirmed an extremely tiny A/B test on limits; the article does not name the provider, and it predates this post.

Anthropic's Max plan help page confirms that plans have session limits that reset every five hours and weekly limits, and that Anthropic may limit usage in other ways. No Anthropic statement addressing the SemiAnalysis claim was found.

Anthropic has not publicly confirmed or denied a test of this kind in the sources reviewed, so the open questions are whether it responds and whether other accounts show the same split. SemiAnalysis marked its post as part one of two. Because the measurement covers two accounts, whether it applies to other Max tiers or to most accounts is not established. Anthropic's help page says users can see their next reset time under the Usage page in Claude settings.

Sante is a mixture of experts model, meaning only part of the network runs for each token. The model card on Hugging Face says it is meant for medical research and workflows, not clinical diagnosis or treatment.

On October 22, Ant Ling posted that it is officially open sourcing Ling 3.0 Flash Sante, a healthcare enhanced mixture of experts model with 124B total parameters and 5.1B active per token. It said the model delivers leading performance among similarly sized models on professional healthcare tasks and complex medical reasoning, and is built for research and workflows with stronger deep research, multi agent collaboration and tool use. It said the weights are available now.

The Hugging Face model card lists the model with 256K context, reasoning and tool calling, and says it is intended for medical research, evidence synthesis and tool assisted workflows rather than clinical diagnosis or treatment.

The chart in the post compares Sante with models including DeepSeek, GLM, Kimi, Claude, GPT and Gemini on healthcare tests such as HealthBench Professional and MedQA, plus agent tests such as BrowseComp. Some of the benchmarks are marked internal. These results are Ant's own and have not been independently verified. The post drew modest engagement compared with Ant's other release today.

Watch for independent medical benchmark results and for researchers to test Sante against larger general models on real workflows. Also watch whether hospitals or health research groups adopt it, given that the model card rules out clinical diagnosis. Regulators may also look at how open medical models are labeled and used. Ant's release cadence suggests further specialized Ling variants could follow.

Alexandr Wang, whose profile lists him as chief AI officer at Meta, posted the launch, so this is a company announcement. Meta's Japan newsroom post, dated October 23 Japan time, says Japan is the first launch region outside North America, where Muse launched in September. It says Muse is available on iOS, Android and muse.ai. Meta says many basic features are free, with subscription plans for heavier use.

Alexandr Wang posted at 8:41 PM Eastern on October 22, which is the morning of October 23 in Japan, that Muse is launching in Japan and is available in app stores, adding a joke about anime memes. The attached six second clip shows instant photo prints of a character in Japanese scenes.

Meta's Japan newsroom post says Muse is designed to be used like messaging another person. Meta says it plans and carries out tasks such as sending email and booking travel, can open a browser and fill in forms, keeps working after the app is closed, and notifies users when approval is needed, for example before sending an email or making a purchase. It also says Muse remembers things that matter to the user.

Meta lists connectors for Gmail, Google Calendar, Instagram, Stripe, Shopify, Trip.com, Agoda and Klook. For Japan it says it is advancing integrations with Mercari, Money Forward ME and Ticket Pia; the page does not say whether they are live.

On safety, Meta says Muse runs on an isolated virtual machine called Muse Secure VM per user, that a separate Sentinel agent manages internet access, that Muse cannot see passwords or payment methods, and that it asks before key actions. Meta also says users can opt out of having chats used for training and that data is not shared with Meta's ad systems. These are Meta's own descriptions.

Meta says newer Muse features shown at Meta Connect, including an advanced voice mode, Muse Realtime Avatar in English only, and a Muse email address, will have their Japan availability announced separately. The Mercari, Money Forward ME and Ticket Pia integrations are described as being advanced, with no dates or live status given. Pricing for the paid plans in Japan was not stated in the sources reviewed. Independent testing of the safety claims was not found.

Ben Lang's profile says he is building SpaceXAI via the Cursor team, so the post is an insider summary of company product updates, not independent reporting. The post lists six items without links. The x.ai Grok Bot changelog, a Grok Bot post on X and a Runway page confirm several of them. Other items had no first party source in the pages reviewed.

Ben Lang posted at 9:16 AM Eastern on October 22 a six item list of what the Grok Bot team shipped in recent days. Each item was checked against first party pages where one exists.

Reports, memos and dashboards in the chat: the x.ai Grok Bot changelog, version 0.75.2 dated October 22, says Bots can build Pages, such as reports, memos and dashboards, that you edit together or have the Bot keep up to date. Confirmed by the changelog.

3x faster computer tasks: a post from the Grok Bot account at 3:55 PM Eastern on October 16 says computer tasks finish 3x sooner on average and replies to everything else arrive in less than half the time. That is a company figure with no method given.

Aliases for the Bot email: the changelog entry for October 14 says Bots can send from a tagged form of their address, like name+shop@, and that a routine on that tag runs only for its mail.

Runway: Runway's own page describes a Runway plugin for Grok Bot that uses Runway credits. Robinhood and Coinbase are not in the changelog or any first party page read; iPhone in Canada lists them among new connectors in an October 17 article, and a HuggingNews article is headlined about them, both secondhand.

Notifications on the device in use: the October 14 changelog says that while you use Grok Bot on a Mac or Windows PC, replies notify you there instead of your phone. iPhone in Canada describes alerts going only to the device in use, so the post's wording is broader than the changelog text.

The iPhone action button is not in the changelog, and no first party source was found.

The changelog's latest entry is version 0.75.2 from October 22, which also lists video creation, pinned Library items and Shop Pay that the post does not mention. The action button, Robinhood and Coinbase items would need a Grok Bot announcement or help page to be confirmed first hand. The speed post gives an average, so individual tasks may differ. The speed claim is the company's own.

The Gemini API CLI is an open source command line tool for the Gemini API, built by the Google AI Studio team. It is separate from the existing Gemini CLI coding agent, and Google labels it experimental and not a supported Google product.

On October 22, Dmitry Lyalin, who lists himself as a staff member at Google DeepMind and Google AI Studio, posted that the Gemini API CLI is available in preview. He said it gives an agent the power to generate images, videos, text to speech or music, analyze media, transcribe audio, or manage agents. The post is a developer announcement from a Google employee.

The release itself is not new that day. Google AI Studio's introduction, written by Lyalin and Philipp Schmid, is dated October 19 and calls it an experimental preview at version 0.4.x. The post and the AICoder summary say the tool is a zero dependency binary, and the GitHub page lists an Apache 2.0 license.

According to the project's README, the commands cover text generation, image, speech, music and video creation, analysis of files and YouTube links, transcription, and managed agents, each using a default Gemini model. Google says it works with any agent that has shell access, naming Claude Code, Codex, Cursor and others. For agent use, commands can describe themselves offline, print request schemas, validate calls with a dry run, and return structured JSON.

The repository's documentation carries a disclaimer that this is not a supported Google product. AICoder notes that Google has published no performance benchmarks for it and that some developers find the name confusing next to the Gemini CLI.

Watch how the command surface changes as Google evolves it alongside the Gemini Interactions API, since the team says it wants feedback and the tool is experimental. Also watch whether Google moves it from a project labeled not supported to an official product, and how much coding agents such as Claude Code and Cursor actually use it. Adoption numbers and any pricing for the managed agents it launches would show whether it spreads beyond early testers.

Ideogram is an image generation company known for rendering text in images. SVG is a vector format that scales to any size and can be edited shape by shape.

On October 22, Ideogram introduced its first vector model. It said the model generates clean, editable SVGs with smooth gradients and sharp typography from a text prompt, or turns any image into a detailed, scalable vector, and that it is live in Ideogram and the API.

The launch video shows examples including a cartoon poster, a circular badge for a koi garden tea house, a mountain logo for a coffee brand applied to a coffee bag, and a landscape illustration, along with a slider that scales a vector shape up without losing sharpness. It also shows a person describing the product on a couch.

The post did not give pricing, benchmark results or limits on file complexity, and the quality claims are Ideogram's own. Vector generation competes with tools that trace images into vectors and with design suites that are adding their own AI features. The model page on Ideogram's site is linked for details.

Watch for designers to test whether the SVGs are truly clean and editable, with sensible layers and paths rather than many tiny shapes. Also watch pricing for the API and whether design tools such as Figma or Canva integrate it. Rival image labs may add vector output in response, and independent comparisons would show how it stacks up. Ideogram may also publish pricing and usage limits for the new model soon.

Muse Spark Open is the open weights version of Meta's Muse Spark, and SGLang is an open source engine for serving large models. The challenge asks outside engineers to beat SGLang's current speed on this model.

On October 22, Daniel Gross posted that a $100,000 challenge was launching to see who can make Muse Spark run fastest, with a link to the contest site. This is an announcement of a contest, not a model release. The post does not say who funds the prizes, and the contest page, read after the post, lists Meta, Paradigm and Andromeda in its sign in example.

According to the challenge page, the contest has two stages. The kernel stage runs from October 22 at 11 AM Pacific to November 4 at 11:59 PM Pacific. Entrants write a GPU kernel, which is tested on random inputs against a reference implementation and timed against SGLang on one H200, using the same inputs and a CUDA graph. The score is how much faster the entry is than SGLang. The page lists five kernel problems, including attention decode, a mixture of experts layer and full decode and prefill layers.

The serving stage runs from November 5 at noon Pacific to November 19 at 11:59 PM Pacific. Entrants submit a diff to SGLang that serves the model on a full eight H200 node, using hidden multi turn coding conversations at 1, 4, 16 and 32 concurrent users. The page says the top kernel competitors get priority for node time, and only the serving stage winners receive prize money.

The listed prizes are $50,000 and a signed H200 for first place, $30,000 for second, $10,000 for third, $3,000 for fourth, $2,000 for fifth and $1,000 each for sixth to tenth. The page says top contestants may be offered an interview for a role at Meta. When read, its leaderboard already showed a best entry at about 1.45 times SGLang's speed.

Watch the leaderboard as the kernel stage runs through November 4, and whether the best kernels hold up on the hidden serving workload in the second stage. Whether the winning changes are merged into SGLang or released openly is not stated on the page. Entrants and observers have also raised questions about GPU access and whether the benchmark can be gamed, and organizers have not published a response in the sources reviewed here.

The Kimi account is run by Moonshot AI, and the post contains only a link to a video, with no text. In the frames reviewed, the terminal text appears to read Model: Kimi K3.1, version 2.2.0. That is on screen text in a teaser, not a statement from Moonshot. The Kimi Code docs read for this entry list K3 as the flagship and do not list a K3.1 model ID.

The Kimi account posted at 9:41 PM Eastern on October 22 a 24 second video with no caption, only a link. It opens on a retro beige computer. The terminal in those frames appears to show Welcome to Kimi Code, a directory under a user named moonshot, a line reading Model: Kimi K3.1, version 2.2.0, a notice that Kimi Code Desktop is here, and a prompt to build a complex procedural 3D scene in Three.js, with a status line reading Kimi K3.1 thinking: high.

The rest of the clip cuts between glitching code and abstract images, including a white tree, a fish, colored ribbons, coral, metal discs, a school of fish and the moon, and ends on a KIMI wordmark. No specifications, benchmarks, pricing or availability are shown.

The Kimi Code docs page, read after the post, describes K3 as the current flagship, at 2.8 trillion parameters with a 1M token context window, and lists model IDs for K3, a 256K context version of K3, K2.8 Preview and K2.7 Code HighSpeed. K3.1 is not listed there.

The frame text was read from the video itself and is an observer reading, not a Moonshot statement. This entry does not treat a model name seen in a teaser as a release. No Moonshot model card, weights, pricing or launch post for K3.1 was found as of this scan.

Moonshot has not published a model card, pricing, weights or release date for K3.1 in the sources reviewed. The first confirmation would be a Moonshot launch post, a model entry in the Kimi docs, or a model ID in the Kimi API listing. Until specifications are published, the teaser shows a product name and a coding demo only. The Kimi Code docs currently list K3, a 256K context version of K3, K2.8 Preview and K2.7 Code HighSpeed as model IDs.

You.com is best known for web search APIs for AI. An inference router sends each request to a model hosted by one of several providers and can fail over to another provider if one is down.

On October 22, You.com announced Inference Router, which it describes as enterprise grade inference on demand. The post lists Kimi K3, GLM 5.3, DeepSeek V4.1 Flash, Qwen3.8 Flash, Muse Spark 1.3 and Mercury 3, with an uptime SLA of at least 99.9%, US hosting and built in web search. This is a company announcement, and the claims below come from You.com's own product page.

The page promises zero data retention, up to 75 percent lower inference costs and no platform fee. It says requests run first on You.com's own capacity and then fail over only to vetted US hosted providers, and it names Parasail, Meta Model API, Inception and Alibaba Cloud US, with Prime Intellect listed as coming soon. It says running web search on the same server as the model makes answers up to 15 percent faster in internal testing.

The page includes a chart from You.com's own run of Braintrust's open source search evaluation on the LiveNewsBench benchmark, using 1,329 questions at list prices. You.com says open weight models paired with its web search matched frontier models with built in search at lower cost. It also includes a comparison table against OpenRouter and dedicated providers, plus customer quotes. None of these figures were independently checked.

The page directs visitors to contact You.com for access, and it lists guided onboarding as part of the offer.

Watch for independent tests of the uptime SLA and the cost savings, since both are You.com's own claims. Customer results beyond the two quotes on its page would show whether enterprises adopt it. Also watch pricing details, since the page asks for contact information rather than listing rates, and how OpenRouter and other routers respond on reliability guarantees and web search.

Thibault Sottiaux, known as Tibo, leads Codex at OpenAI. On October 4 he promised that for 28 days the team would ship either one clear improvement for Codex users or a full usage reset each day.

Early on October 22, Tibo posted the day 17 entry of his 28 day series: bet you thought we forgot, but no, it is reset day today. He added that it is a good time to remind users to get Daybreak Blue access, which he said takes only 15 minutes, because you do not want to have to do it in the middle of an incident.

The series began on October 4, when he said that for the next 28 days the team would ship each day either a clear improvement relevant to most Codex and Work users or a full reset. The previous reset came on day 15, on October 20, and day 16 brought persistent caching in Codex and 12 hour prompt caching in the API.

According to the Open the Rank tracker, Tibo replied under the post that the reset lands tomorrow before 10am PST, meaning October 23. The post does not say which plans receive the reset or explain what Daybreak Blue is, and OpenAI has not published details about it in the sources found.

Watch for the reset to land before 10am Pacific time on October 23, as Tibo said in his reply, and for OpenAI to explain Daybreak Blue and who should sign up. With 11 days left in the series, watch whether the remaining days bring more resets or product improvements, and whether usage limits change after it ends. Users will also watch whether day 18 arrives on time.

Tech Dev Notes is an independent account that tracks app changes, so this is secondhand; no official launch post was found. Elon Musk reposted it, and the official Grok Bot account has replied to users that their Grok Bot will show up in their X Chat list once linked.

On October 22, Tech Dev Notes posted that Grok Bot is now on X Chat, with a screenshot of the X Chat inbox showing a pinned conversation named Grok Bot with a gold badge and the message Let me check your calendar and inbox for today. Elon Musk reposted it later that morning.

No official announcement of this change was found. The official @bot account has been replying to people who tag it on X, telling one user it was sending the request to their Grok Bot and that it will show up in their X Chat list once it is linked. That suggests Grok Bot can appear as a conversation inside X Chat for users who have linked accounts.

This builds on an October 16 update in which the official account said Grok Bot can now read X Chat, so users can connect X and ask it to work across their messages. The screenshot does not show who can use the feature, whether it needs a paid plan, or whether it is rolling out to everyone. Grok Bot is the agent product from SpaceXAI.

Watch for an official announcement from the Grok Bot or X accounts confirming who can chat with Grok Bot in X Chat and on which plans. Also watch how X handles privacy when an agent reads and replies to messages, and whether usage grows as people reach Grok Bot directly from their inbox. Rival assistants in other messaging apps may respond with similar features.