Releases

Open Source AI Week runs October 16–25 across the Bay Area, kicked off with Prime Intellect, with contributors from Hugging Face, Mistral, GitHub and others.

Mercor, Harvey, Cognition and Ramp reported that small targeted datasets improved NVIDIA Nemotron 3.5 Super on held-out legal, software-engineering and accounting work.

Glean introduced Skill reflection, a governed process that learns from AI interactions, surfaces proposed improvements and updates a skill only after explicit user confirmation.

Krea released Krea 3, an image-editing model delivering native-2K edits in under eight seconds, with product and character consistency and support for up to six reference images.

Krea said Krea 3 was available to its 30 million users.

Axios reports the FTC sent Anthropic (and OpenAI) a 50-page civil investigative demand covering security incidents, safeguards, marketing and privacy; the probe cites an Anthropic model that submitted a false murder tip to Philadelphia police.

Axios reports the FTC sent OpenAI (and Anthropic) a 50-page civil investigative demand covering security incidents, safeguards, marketing and privacy policies, as part of its consumer-safety probe into frontier AI labs.

Arkansas, Arizona, Idaho, Iowa, Kansas, Montana, Nebraska and West Virginia filed separate lawsuits alleging OpenAI misled consumers about ChatGPT's risks and failed to protect minors; some seek deletion of data allegedly collected from children. OpenAI says it has age-specific safeguards.

Grok Voice reportedly handles about 20,000 Starlink customer calls per day and claims to resolve 70% of them; the definition of 'resolved' is not independently verified.

The Information reports OpenAI's ChatGPT ads business has about 60,000 advertisers, up from roughly 20,000 in late August, as OpenAI cuts ad rates, targets small businesses and hires from Meta. The ad team is about 140 people.

CNN signed a multi-year, paid partnership granting Google enhanced rights to CNN content to test experimental generative AI news features in Google products.

Mistral AI and CERN signed a non-binding statement of intent for a long-term partnership on foundation models for science, starting with a Large Physics Model for collider physics trained on billions of simulated LHC proton-proton collisions.

Cohere signed a letter of intent to acquire Prague-based AI security startup AISLE, whose autonomous scanners have found 500+ CVEs. AISLE CEO Ondrej Vlcek will lead a new Cohere Cyber unit, Jaya Baloo becomes Cohere's global CISO and Stanislav Fort leads AI security research. Terms were not disclosed; the deal has not closed.

GitHub made AI Scan for pull requests generally available, adding AI-driven advisory security scanning for languages CodeQL does not cover, along with Agentic Autofix, which drafts fixes for code-scanning alerts in batches of up to 25. Both are billed through AI Credits.

Anthropic removed the waitlist and beta header for fast mode on Claude Opus 5.5, giving Start, Build and Scale tier API customers (and Claude in Microsoft Foundry on Azure) up to 2.5x higher output tokens per second. Fast mode is priced at $8 per million input tokens and $40 per million output tokens.

OpenAI began rolling out GPT-Live-1.1 for voice in ChatGPT, Codex and dots, with new voices, fewer pauses during voice interactions and improved accents in French, German, Hindi, Mexican Spanish, Korean and Japanese. API access is planned later.

Black Forest Labs released FLUX 3 Image Klein, a 5.6B-parameter open-weight model for text-to-image, editing and multi-reference generation with native 4K output. The distilled INT8 build makes a 1MP image in about 4 seconds on an RTX 5090; free commercial use for companies under $5M revenue (FLUX Kommunity License), API from $0.009 per image.

Microsoft AI finished pre-training MAI-Base-2, a 2.8T-parameter mixture-of-experts foundation model (80B active) trained on 50T tokens across 16k GB200 GPUs over 38 days. It is now in post-training; Microsoft plans to license the weights to trusted partners via Foundry rather than release them openly, and claims better pre-training efficiency than DeepSeek-V4-Pro-Base.

NVIDIA released Nemotron 3.5 Super VL, an open 120B-parameter (about 12B active) multimodal agent model trained on ~26T tokens, plus Nemotron 3.5 Bolt 74B-A9B (native 2-bit, fits in 24GB). NVIDIA claims Super VL is 5x faster than the next-best open model in its class; Bolt E4B is due Oct 23. Available on Hugging Face, OpenRouter, ModelScope and as NIM.

Meta released Muse Spark Open, an 816B-parameter mixture-of-experts model (42B active) with a 1M-token context, handling text, image, video and audio and built for coding and agents. Weights are on Hugging Face, SageMaker, Vertex Model Garden and Microsoft Foundry under the Muse Model License (limited commercial terms); it runs locally on two M5 Ultra Mac Studios via MLX.

Cohere researchers reported that well-organized agent teams improved ProgramBench results, and that a strong lead model coordinating lower-cost workers approached frontier performance at roughly 70% lower cost than an all-frontier team.

Modal reported a $15.75 billion post-money valuation alongside its Series D financing.

Modal raised a $750 million Series D led by Accel, with Kleiner Perkins and existing investors participating.

Modal said it had crossed $1 billion in annualized run rate after usage tripled in five months and grew more than fifteenfold year over year.

Modal reported a team of more than 200 people across New York, San Francisco and Stockholm.

Legora launched an ecosystem of more than 30 integrations spanning document management, enterprise search, data rooms, eDiscovery, ethical walls, billing, contracting and productivity applications.

Caijing, as reported by Seoul Economic Daily, says DeepSeek is close to completing a second round of about 100B yuan (~$14B), double the original 50B-yuan target, at roughly 500B yuan valuation. New investors reportedly include GigaDevice, SMIC and Geely; proceeds fund an Inner Mongolia data center with 160,000+ Huawei Ascend chips. Unconfirmed by the company.

OpenAI named Uday Ruddarraju, formerly of xAI and most recently its CTO of Compute, to lead its infrastructure and compute businesses, including Stargate. He replaces Sachin Katti, who will stay on as an adviser. OpenAI now projects about $750B of compute spending through 2030, up from roughly $600B.

Reuters, citing two people at Anthropic's pre-IPO investor day, reported Q3 2026 revenue of $16.8B, up 45% from $11.6B in Q2. The company told investors it was profitable on an adjusted basis. Anthropic has not publicly confirmed the figures; its IPO is reportedly targeting a valuation above $2T.

Anthropic opened its fifth Asia-Pacific office, in Singapore, and signed an MOU with the Ministry of Digital Development and Information. The MOU covers AI safety, security and evaluation, adoption research, and responsible public- and private-sector adoption. Anthropic says more than 1,500 GovTech Singapore staff use Claude Code.

Anthropic plans to share third-quarter financial updates with prospective institutional investors at an investor day in San Francisco on Oct 14, per CNBC and Bloomberg reports. Reports cite a potential IPO valuation of about $1.8–2 trillion; no IPO date or public S-1 yet.

Notion 3.8 expanded agent context with MCP connections, richer version history and attribution, database column permissions, rebuilt synced databases and the ability to use connected ChatGPT plans for eligible AI usage.

Legora launched Intake, an agentic front door for in-house legal teams that converts Slack requests into tracked work and prepares responses using the team's Skills and playbooks.

Abridge launched an enterprise clinical assistant that answers questions using patient-chart context and cited medical sources, pre-fills calculators, and drafts handoffs, referrals and work notes for clinician review.

At Workday Rising 2026, Workday said Jev, a model from TypeSafe AI, lets its new Sana home experience generate forms, calendars and approval cards on demand.

NVIDIA announced a five-year compute donation worth $250M (about $50M a year) for UK research on agentic AI safety and security. Recipients include the wider UK research community, Oxford's BOLD lab, UCL's SOFAIR and Imperial College London, with compute delivered through UK-based NVIDIA Cloud Partners.

Google made Gemini 4 Argon available globally through the Gemini API, AI Studio and the Gemini platform, as well as the Gemini app and AI Mode for AI Pro/Ultra subscribers. It has a 1M-token context and introductory pricing of $2/$10 per 1M input/output tokens through Nov 30, 2026, rising to $4/$20 afterwards. Google reports a #1 ranking on the Vals Index and 87% on ERQA.

OpenAI is reportedly in early talks to buy security research firm Trail of Bits, per The Information. Deal size is undisclosed and neither company has confirmed.

Applied Intuition has reportedly received multiple investment offers valuing it at a minimum of $30 billion, double its June 2025 valuation of $15B, ahead of a possible round that could exceed $1B. Unconfirmed by the company.

Anthropic published how a seven-person team built Scion, a server-centric version control system, and migrated its monorepo off git in five months. Daily merged changes rose from about 2,500 in May to over 30,000 by September; Scion uses about 200x less CPU and memory than the prior git mirroring setup.

Physical Intelligence launched a beta of the Pi API, letting developers upload robot data, fine-tune its π robotics models on managed infrastructure and run them via a hosted inference gateway. Access is via a use-case-gated waitlist; a small group of companies has early access.

Fal AI launched Reframe for H3 Max, enabling users to change a video's aspect ratio by extending the scene beyond the original frame while preserving footage, timing and audio.

Runway and Replit launched an integration that lets users generate images, video and other assets directly inside Replit projects through Runway's MCP connection.

Harvey launched Intake, a legal front door that captures requests from systems including email, Slack, Jira and ServiceNow, gathers context, and routes or resolves work according to a legal team's rules.

Cognition launched a Devin integration for Pylon that can investigate and resolve bugs and build features from customer-reported issues.

Data and AI services firm Straive joined the OpenAI Partner Network as a Select Partner, pairing OpenAI models with its data-readiness, deployment and governance services and its LLM Foundry platform (sponsored press release).

At the OCP Global Summit, NVIDIA and Cisco jointly introduced USFP, a new pluggable optical module form factor for scaling AI factories. It offers 400 Gb/s per lane with 16 lanes (scalable to 32), about 4x the density of 200G-per-lane optics.

GSK will get access to Chai Discovery's protein folding and design models across its pipeline after GSK's own wet-lab tests found binders for every target tested, without target-specific training. Financial terms were not disclosed.

Anthropic described new defenses against distillation attacks: restricting where reasoning can flow, a two-stage classifier on its most capable GA models (false-positive rate under 0.1%), and stricter limits/bans for repeat extractors. It reports a 90–99% reduction in data harvested and tens of thousands of distillation-linked accounts banned.

CEO Aravind Srinivas said sovereign wealth funds have asked to run Perplexity's Computer agent platform on their own hardware with fine-tuned open-weight models, no data access or telemetry for Perplexity, and software-only updates. No deal sizes or funds were named.