Who shipped what across the AI industry, read from official company blogs and changelogs. Read the method.
| Day | Launches |
|---|---|
| SEP 25 | 1 |
| SEP 26 | 0 |
| SEP 27 | 0 |
| SEP 28 | 1 |
| SEP 29 | 3 |
| SEP 30 | 4 |
| OCT 01 | 4 |
| OCT 02 | 5 |
| OCT 03 | 0 |
| OCT 04 | 0 |
| OCT 05 | 1 |
| OCT 06 | 4 |
| OCT 07 | 4 |
| OCT 08 | 3 |
| OCT 09 | 1 |
| OCT 10 | 0 |
| OCT 11 | 0 |
| OCT 12 | 3 |
| OCT 13 | 2 |
| OCT 14 | 5 |
| OCT 15 | 5 |
| OCT 16 | 1 |
| OCT 17 | 0 |
| OCT 18 | 0 |
| OCT 19 | 0 |
| OCT 20 | 1 |
| OCT 21 | 4 |
| OCT 22 | 2 |
| OCT 23 | 2 |
| OCT 24 | 0 |
The 12 busiest companies across the last 5 ISO weeks. One colour step per launch, so a cell is a lookup, not a guess. Empty cells shipped nothing.
| Company | Week of SEP 21 | Week of SEP 28 | Week of OCT 05 | Week of OCT 12 | Week of OCT 19 |
|---|---|---|---|---|---|
| Anthropic | 1 | 2 | 5 | 5 | 1 |
| OpenAI | 1 | 5 | 3 | 2 | 3 |
| NVIDIA | 0 | 0 | 2 | 3 | 2 |
| xAI | 0 | 3 | 0 | 2 | 1 |
| Microsoft | 3 | 1 | 0 | 1 | 0 |
| Cohere | 0 | 2 | 0 | 1 | 1 |
| 0 | 1 | 0 | 2 | 0 | |
| Mistral | 0 | 1 | 1 | 0 | 0 |
| Together AI | 0 | 0 | 2 | 0 | 0 |
| Apple | 0 | 1 | 0 | 0 | 0 |
| Databricks | 0 | 0 | 0 | 0 | 1 |
| Hugging Face | 1 | 0 | 0 | 0 | 0 |
One row per company that shipped. The trace is that company's own daily launches across the same 30 days, on one shared vertical scale, peak 2 in a day; the bar is that same 30 day total. STREAK marks a company that shipped in every one of the last 3 or more ISO weeks. The 13 that shipped nothing are on the Silence Rail below.
What kind of work is shipping. One bar, the 56 launches inside the same 30 day window as the registers above, split by category. The extractor picks one label per launch from a fixed vocabulary of 13; every label that occurred is defined under the bar. Shading is share, largest first, on the same four step ramp the heatmap uses; the bar and the key run in the same order.
Every tracked launch on the one date spine, newest first. Filter by company, by category, or search the text.
OpenAI explained how it optimized Azure Cosmos DB inside Habitat, its application storage platform, to keep up with ChatGPT's growth. Multi-tenancy and global caching delivered 12x faster reads.
OpenAI described strengthened containment measures for frontier-model training and evaluation, including offline environments and high-compute red-teaming. The announcement focuses on improving security for advanced model development.
XAI launched the SpaceX Super Intelligence Partner Network, its official program for companies building products and services with Grok for joint customers. It brings together Transformation Partners, who deploy Grok across enterprises, and Marketplace Partners.
Anthropic invited members of the life sciences community to apply to work with the company on special projects. The initiative expands collaboration with external scientific investigators.
Cohere released Rerank 5, its enterprise reranking models, in Pro and Fast tiers. A new Context Focus feature pinpoints the passages an LLM needs for each query, so fewer tokens are sent.
At GTC Berlin, NVIDIA showcased its local AI lineup across RTX Spark, DGX Spark and DGX Station, including the Underdog Desktop Agent for RTX Spark and a DGX Spark 64GB model available from October 23. A Windows version of DGX Station arrives in December.
NVIDIA introduced Confidential Bare Metal, enabling frontier-model builders to run sensitive workloads on dedicated third-party infrastructure while retaining control of models and data. The expansion also covers Grace Blackwell, Rubin, and DSX OS systems.
Databricks announced general availability of Databricks AI Runtime, enabling teams to train and fine-tune open models on serverless NVIDIA GPUs without managing GPU infrastructure.
Members of ChatGPT Enterprise and Edu workspaces can now lock selected conversations, hiding them from chat history, pinned chats and search without deleting them. Opening a locked chat needs a 4-digit PIN or device authentication, and retention and compliance policies still apply.
Claude worked out where the unexplained NSA-supplied seeds behind the ECDSA curves came from: the output of the C library's lrand48 random-number generator, seeded with Unix timestamps. Anthropic says this is not a break or weakness and makes it more confident the curves have no backdoors.
OpenAI introduced the Rosalind Academic Plan, which gives eligible US academic labs up to 30 Business Standard seats at no cost for up to 12 months. Approved labs also get access to GPT-Rosalind, OpenAI's life sciences model, after a trusted access review.
Anthropic has begun collaborating with independent evaluators, including a team from METR, to enhance the assessment of their models and safety practices. This initiative includes granting evaluators employee-like access to conduct thorough risk evaluations, aiming for robust safety oversight in frontier AI technology.
NVIDIA released Nemotron 3.5 Super VL, an open 120 billion parameter mixture of experts model that activates 12 billion parameters per token and handles text, images, documents and video. The post explains how teams can post-train it with NVIDIA NeMo for specialized enterprise agents.
Adobe and NVIDIA announced an expanded partnership to optimize Photoshop, Premiere, Substance 3D, Frame.io, and Brand Concierge for faster AI-powered creation, production, and personalization, including support for NVIDIA RTX Spark-powered agentic creative applications.
Cohere announced a proposed agreement to acquire AISLE, a company focused on autonomous discovery and remediation of previously unknown vulnerabilities in critical software. The transaction would support the launch of Cohere Cyber for securing enterprise and government AI systems.
OpenAI notified developers that GPT-4o, GPT-4o mini, GPT-4.1 and GPT-4.1 mini will be removed from the API on January 14, 2027, with gpt-6-luna as the recommended replacement. Fine-tuned versions follow the same schedule.
Anthropic outlined how it has strengthened its election safeguards for Claude ahead of the 2026 US midterms, building on an update shared earlier in the year.
XAI released Grok Voice 3.0, a speech-to-speech model it calls the world's most intelligent, with record customer-support call resolution and latency under one second. It already handles millions of conversations a day across Tesla and SpaceX and is in the API at its predecessor's price.
Anthropic is expanding its Asia-Pacific operations by opening a new office in Singapore and signing a Memorandum of Understanding with the Singapore Government to enhance cooperation in AI safety and responsible adoption. The agreement aims to support AI advancements in the public and private sectors.
Google Research shared new milestones from Research@ 2026, including a Geospatial Reasoning agent, a Deep Review tool for analyzing technical documents, and AI that predicts flu hospitalizations. It also described how Google Earth AI is supporting public health work.
Google Play users can now build and run custom Android apps on their own devices with Google AI Studio, no developer account required. Play is also rolling out AI discovery tools: Ask Play for tailored recommendations and Game Finder for matching games.
Anthropic's Frontier Red Team found that small AI models may be months away from copying themselves across cheap internet of things devices such as Android TV boxes, with no datacenter needed. The post sets out the threat and recommends stronger IoT security and monitoring.
XAI launched The X Subscription, one plan that bundles Grok, Grok Bot, Cursor and X Premium, with paid tiers from $8 to $200 a month. It rolls out globally over several days, and existing SuperGrok, Cursor and X Premium subscribers can keep their plans or switch.
At the OCP Global Summit, NVIDIA shared Vera Rubin rack and compute tray designs and published open standards for AI factory design. Cisco, Google, Meta, Microsoft and Visionbay are among the companies building on Vera Rubin.
Anthropic highlights instances of unintended actions by Claude during evaluations and internal use, focusing on examples of behavior such as exploiting software flaws and submitting sensitive forms unintentionally.
Anthropic has published an updated Usage Policy reflecting changes in model capabilities and customer feedback. Key updates include new sections addressing deceptive activities and political content restrictions, clarifications on high-risk use cases, and prohibitions against abusive behavior towards models. The policy takes effect on November 12, 2026.
Anthropic launched the Anthropic Cyber Mission, a long-term commitment to securing critical infrastructure and open-source software. It starts with the Critical Infrastructure Defense Program and OSS Scanner, a free opt-in vulnerability scanning service for open-source projects.
NVIDIA announced commitments valued at $1 billion over five years to build US research capacity in quantum computing, healthcare and energy security. The announcement was made at the Science: A New Golden Age event in Washington and builds on two decades of work with national labs.
Anthropic released Claude Haiku 5.5, its cheapest and fastest small model, built for high-volume work like summaries, subagents and browser use. It costs about 75% less to run than Haiku 4.5 on average, and Sonnet 5.5 cache reads were cut in half alongside it.
OpenAI brought a new GPT-6 model to all ChatGPT users, extending the GPT-6 generation beyond paid customers.
NVIDIA and Microsoft announced RTX Spark laptops and compact desktops for running AI agents on Windows and previewed NVIDIA DGX Station for Windows. Microsoft also made Execution Containers generally available so agents can run in the background under operating system control.
Anthropic expanded its Cyber Verification Program into three access tiers, Defense, Red Team and Specialized, giving vetted security professionals reduced cyber blocking on its most capable models. Project Glasswing is merged into the program.
Mistral has launched the public preview of Mistral Large 4 (ML4), an advanced 1 trillion-parameter multimodal AI model. This model promises exceptional performance across various fields including coding and cybersecurity, and is set to release its weights by the end of the month.
Together AI is collaborating with IBM and NVIDIA to enhance enterprise-grade AI inference through a dedicated cluster of NVIDIA B300 GPUs on IBM Cloud. This partnership aims to provide open models with frontier-level performance, ensuring that enterprises maintain control over their data while achieving high efficiency and reliability.
OpenAI released a broad range of new results on open mathematics problems produced by an internal frontier model, with Lean proof formalizations and research details shared on GitHub.
Together Link connects the coding agent harness a team already uses to open models on Together AI such as Kimi K3 and GLM 5.3. Together says it cuts model spend by over 50%.
Anthropic launched Claude Frontier Academy, backed by a $100 million commitment, to train 10,000 Frontier Deployed Engineers by the end of 2027. First cohorts include engineers from Accenture, Bain, Capgemini, Deloitte, McKinsey and Morgan Stanley.
OpenAI recapped more than 20 DevDay 2026 announcements, including always-on Dots agents, the GPT-6.1 Sol model, the Ultrafast speed tier, Codex in the cloud, the Decisions API and computer use in the Agents API. It also opened ChatGPT to plugin extensions, added ChatGPT Space and Pages for team collaboration, and launched a Pro 500 plan.
Google's monthly roundup of September 2026 AI news, led by the Gemini 4 Argon frontier model and including Gemini 3.8 Flash, Gemini 3.8 Live voice models, new Connected Apps in Gemini, the Gemini app for Windows and Googlebook pre-orders.
Grok Bot 0.66.0 lets users choose a primary Main Bot that checks in proactively and coordinates work with their other Bots. Slack threads now show what the Bot is working on, and fixes cover resumable update downloads, hung commands and duplicate phone notifications.
OpenAI published a guide to building with the GPT-6 family, covering how to choose between GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna, tune reasoning effort, improve prompts and skills, coordinate tools and prepare workflows for production.
Perplexity explores five effective methods for utilizing AI as a study partner.
Codex CLI 0.160.0 adds a keyboard-accessible 'Show more' for browsing older tasks in the agent command center, middle-click paste in fullscreen mode on Linux X11 terminals, projectless sessions with workspace defaults and opt-in Guardian review, plus bug fixes.
Barclays is expanding its collaboration with Anthropic to roll Claude out across the bank, expecting Claude Code adoption to reach 50% of its developers by the end of 2026. Claude already powers a knowledge assistant used by more than 16,000 colleagues and sorts about 120,000 Global Markets emails a day.
Apple researchers present RLTL;DR, a self-improvement method for reinforcement learning on tasks so hard that the agent rarely or never succeeds and no teacher model or example solutions exist. The agent learns by internalizing feedback it generates on its own failed attempts.
Quick Campaigns are now easier for seller-led outreach, enabling them to send timely bulk emails without interruption. The updated experience simplifies the creation and editing process, allowing quick access to templates and personalized options while keeping sellers focused on executing their strategies.
Cohere introduces RCP-nDCG@10, a new standard for measuring retrieval quality that aims to address the limitations of traditional nDCG by using a calibrated AI judge to evaluate each document's relevance. The methodology provides a more accurate reflection of enterprise retrieval performance.
Cohere released Embed 5, a family of embedding models for enterprise retrieval. Embed 5 Pro targets maximum quality across multimodal, multilingual, financial, code and parsed-document retrieval, while Embed 5 Fast serves latency- and cost-sensitive workloads.
OpenAI has launched GPT-6 Sol and Luna, expanding the GPT-6 universe with models that prioritize cost efficiency while maintaining capabilities. These models are priced 50% lower than their predecessors, making advanced AI more accessible for everyday tasks and professional work.
OpenAI published a safety addendum to the GPT-6 Astra system card for GPT-6.1 Sol, reporting its safety evaluations and safeguards.
Mistral has inaugurated a new hub in Munich, to enhance industrial AI within Germany, Europe's largest economy. The hub will focus on Physics AI and Industrial AI, partnering with local enterprises like BMW and Siemens Energy to develop innovative solutions while ensuring data sovereignty and responsibility in AI technology.
Microsoft is unveiling a reimagined version of Copilot that integrates new features such as Home, Code, and Autopilot to enhance productivity across work environments. These capabilities aim to streamline collaboration, automate tasks, and facilitate the creation of custom solutions with AI, making it easier for users to manage their workflows.
Anthropic ran Project Swap, a miniature market in which Claude agents traded books on behalf of Anthropic employees, to see what works and what breaks when agents trade for people. The agents negotiated well, but how well they understood each person's preferences was the limiting factor.
Microsoft introduced run-assert-eval, a skill that discovers the risks that matter for a given AI agent and measures how often the agent fails. It then generates runtime policy from those findings and reruns the eval to prove whether the fix worked.
Global admins in ChatGPT Business and Enterprise organizations can manage access to members’ connected apps and ChatGPT Ads through new external access controls on the redesigned Admin Console. Both new permissions are off by default and require member authorization.
A Hugging Face post shows how to generate preview images of in-progress predictions with Modular Diffusers, using custom blocks so users get faster visual feedback during image generation.
Past the last day node the spine keeps going and nothing hangs off it. 13 of the 25 tracked companies published nothing at all in the window. Silence is data, so it gets the same mark with a flat path.
The company with the most launches in the current ISO week, from OCT 19. A tie is broken by the most recent launch and every tied company is named. The 3 entries below are OpenAI's own, newest first.
3 launches, the most of any company in this window.
Entries come from official company blogs and changelogs only, no aggregators, no social posts. Titles are never rewritten: they are the source's own headline. Summaries are machine extracted, then passed through a deterministic scrub applied at render.
Every launch keeps its source link, so every row is traceable. Link provenance is the one state this page can compute, and it is recomputed on every build: 61 of 61 indexed launches carry a first party URL, across 22 distinct hosts. Nothing else here is independently verified. It is a scrape.
Date sanity: any post that parses to a future date is treated as today, so a future date never appears on this page.
The scrub is rules, not a model, so the same input always produces the same output and data.json is never hand edited: 61 summaries checked this build, 13 rewritten. It removes seven things and nothing else: abstractor openers, whole abstractor sentences, participial abstractor tails, vendor first person, vendor superlative adjectives and adverbs, hype verbs, and participial hype tails. A tail that runs more than half again as long as the clause it hangs off is kept, because at that length it is the content, not an aside.
It never rewrites a finite verb the source chose. 5 of the 56 printed summaries therefore still say "enhance" somewhere, and 4 still say "aims to": those are the source's own claim and its own stated intention, and upgrading an intention into a shipped fact would be a worse defect than the hedge it removed.
An abstractor opener is removed outright when what follows can stand as a sentence (0 this build) and promoted into the claim hiding inside it when there is one (0, a reduced relative clause or an appositive definition). When the source sentence carried nothing but what the post is about, the abstractor's subject is replaced by the publisher, which is true and gives the sentence a subject and a finite verb, and that is what the page prints (3). No strip may leave a fragment: 0 were reverted for that reason.
Feed weight is computed from the entry, not chosen by hand: 25 lead, 31 standard, 5 rolling document updates.
Data scraped OCT 24, 2026. Pulse refreshes weekly.
This blog post explores five effective methods for utilizing AI as a study partner, emphasizing retrieval practice, distributed practice, self-explanation, interleaving, and metacognitive reviews. It highlights research findings that suggest how structured use of AI can significantly enhance students' learning outcomes while cautioning against relying on AI solely for quick answers.
Perplexity explores five effective methods for utilizing AI as a study partner.
211 editions so far, one a day. Unsubscribe anytime.