AI Weekly Brief
Anthropic starts the IPO clock
Anthropic's confidential draft S-1 starts the IPO clock earlier than expected, before frontier-AI economics are publicly disclosed.
TL;DR
- Anthropic confidentially filed a draft S-1 with the SEC, starting the IPO-review clock about two weeks after its $65B Series H and ahead of the Q3 timeline its trajectory implied.
- OpenAI made its frontier models and Codex generally available on Amazon Bedrock in AWS Commercial and GovCloud regions, placing its models inside its main distribution rival's cloud.
- One day brought three open-weight releases: MiniMax M3 (1M context, multimodal, agentic), NVIDIA's Cosmos 3 physical-AI world model, and JetBrains' Mellum2 (12B MoE).
- Anthropic expanded Project Glasswing to ~150 more critical-infrastructure organizations and mapped a year of AI-enabled cyberattacks to MITRE ATT&CK, finding attackers shifting AI use to post-compromise stages.
- Healthcare AI clearances clustered: the FDA cleared AI tools from GE HealthCare (with a Predetermined Change Control Plan), Philips, and Clarius within three days.
Editor’s Pick
The timing is the signal. The prior brief projected a Q3 2026 filing tied to an October IPO target, yet the draft arrived roughly two weeks after the $65B Series H, suggesting revenue quality and internal controls are already at registration readiness. Because the submission is confidential and made under Rule 135, the financials that will ultimately reprice the sector (ARR composition, customer concentration, gross margins, and compute cost structure) stay private until a public filing, so the immediate effect is to start the clock, not to disclose the numbers. The cadence around it is notable: in the same week Anthropic expanded a critical-infrastructure security program, published an empirical cyber-threat study carried in Verizon's DBIR, and formalized a tiered services-partner network, assembling the regulatory, security, and go-to-market posture expected of a public company. When the public S-1 lands, its audited disclosures will become the first hard reference points for frontier-AI unit economics and will directly shape OpenAI's own listing calculus.
Model Releases
MiniMax M3. An open-weight model combining frontier coding (vendor-reported 59.0% on SWE-Bench Pro, 66.0% on Terminal-Bench 2.1), native multimodality, and a 1M-token context via the new MiniMax Sparse Attention, which reports ~9x prefill and ~15x decode speedups at 1M tokens; weights are pledged within ~10 days.
NVIDIA Cosmos 3. An open world foundation model for physical AI whose mixture-of-transformers architecture unifies vision reasoning, world generation, and action prediction, released on Hugging Face as Cosmos 3 Super (64B) and Cosmos 3 Nano (16B).
JetBrains' Mellum2. A 12B Mixture-of-Experts model (2.5B active) trained from scratch on natural language and code under Apache 2.0, built for latency-sensitive routing, RAG, and sub-agent work with more than 2x faster inference than similar-sized models.
H Company's Holo3.1. An open computer-use model family spanning 0.8B to 35B-A3B, shipping its first quantized checkpoints (FP8, Q4 GGUF, NVFP4) for local inference and targeting robustness across web, desktop, and mobile agent environments.
AI Tools & Products
OpenAI extends Codex to non-developers. Six role-specific plugins bundling 62 apps and 110 skills for analysts, marketers, sales, designers, and investors, plus a preview of shareable interactive Sites and in-place annotations, aimed at the non-developers who now make up ~20% of Codex's 5M+ weekly users.
xAI's Composer 2.5 in Grok Build. A fast model now available in Grok Build, tuned for long-running tasks and complex instruction following.
xAI's Grok Imagine Video 1.5 Preview. A new image-to-video model on the xAI API that animates a single still into short 720p clips with text-directed camera movement and multi-shot continuity, positioning xAI against Veo and Seedance after OpenAI pulled Sora.
Grok becomes Vapi's default voice. Grok becomes the default text-to-speech engine for Vapi's 12 core voices across the platform's 2.5M+ voice agents, with Grok speech-to-text and custom voice cloning exposed through the Grok Voice API.
Research Breakthroughs
Executable games expose weak belief revision. A benchmark of 474 executable games frames reasoning as evidence gathering under partial observability; frontier models handle clean tasks (Gemini-3.1-pro 74.2%) but drop sharply under counterfactual revision (to 57.6%) and necessity-judgment pruning.
GAIS turns MCP servers into training environments. Grounded Interaction Synthesis converts real Model Context Protocol servers into executable Python environments and plans tasks around deep inter-tool dependencies, letting base Qwen3 models trained on the data match or beat their official instruction-tuned versions on BFCL, τ²-Bench, and ACEBench.
Anthropic maps a year of AI-enabled cyberattacks. Mapping 832 banned malicious accounts to MITRE ATT&CK, Anthropic's Frontier Red Team found medium-or-higher-risk actors rising from 33% to 56% over the year and AI use shifting from initial access toward post-compromise techniques; agentic orchestration has no ATT&CK technique ID.
Workflows & Techniques
ALAR escalates from latent to explicit reasoning. Adaptive Latent Agentic Reasoning defaults to compact latent reasoning on routine agent turns and escalates to explicit chain-of-thought only for harder decisions, preserving accuracy while cutting generated tokens by up to 43.6% on agentic search and 84.6% on tool use.
Industry & Ecosystem
Anthropic confidentially files a draft S-1. Anthropic submitted a confidential draft registration statement to the SEC, starting review and giving it the option to go public, roughly two weeks after its $65B Series H and earlier than the expected Q3 filing; share count and price are not yet set.
OpenAI models and Codex arrive on AWS. OpenAI's frontier models and Codex are generally available on Amazon Bedrock in AWS Commercial and GovCloud regions, with Daybreak flagged as coming, letting AWS enterprises adopt OpenAI through existing security, procurement, and governance workflows.
Anthropic expands Project Glasswing. After ~50 initial partners using Claude Mythos Preview surfaced more than 10,000 high- or critical-severity flaws, the program extends to ~150 new organizations across 15+ countries in power, water, healthcare, communications, and hardware.
Claude Partner Network adds a Services Track. A three-tier services structure scored on certified practitioners, production deployments, and references, plus a partner hub; more than 40,000 firms have applied and over 10,000 consultants are Claude-certified since the network's March launch.
AI Startups
Town raises $55M for a personal AI assistant. The Series A funds a personalized assistant that learns how individuals work across their existing tools rather than requiring users to become prompt experts.
Mecka AI raises $60M for robot data. The Series A funds a robotics training-data engine that captures human hand gestures, gaits, and physical motion from body sensors and iPhones, supplying the scarce interaction data general-purpose robots need.
Apoha raises $36M to design new materials. Emerging from stealth, Apoha builds AI models that design new substances such as proteins, food products, and paints from a novel class of data describing how materials physically behave.
NP Company raises €6M for engineering simulation. The AI-native simulation startup for aerospace, defense, energy, and automotive raised a Partech pre-seed, with Mistral AI co-founders among its angels, to build foundation models that replace multi-day computations with real-time inference.
Healthcare AI
FDA clears GE HealthCare's MIM Contour ProtégéAI+ 2.0. The 510(k) clearance for radiation-therapy auto-contouring adds an MR Brain model and an updated CT Male Pelvis model, and notably includes a Predetermined Change Control Plan pre-authorizing future model updates across more anatomies.
FDA clears Philips Elevate Plus ultrasound AI. A 510(k)-cleared AI and automation suite for the EPIQ Elite and Affiniti ultrasound systems, including Auto Measure Abdomen, aimed at standardizing routine exams and reducing repeat scans amid staffing pressure.
FDA clears Clarius Ejection Fraction AI. Integrated into the Clarius handheld ultrasound app, the tool delivers automated, objective left-ventricular ejection-fraction measurements at the point of care.
Adaptive Innovations raises $50M for AI homecare. Emerging from stealth with Felicis, Bain Capital Ventures, and Optum Ventures (plus a previously undisclosed $10M seed), the company is building an AI-native homecare agency and expanding its clinical workforce across states.
Waypoint Bio raises $20M for AI-designed CAR T. The Series A advances a spatial-biology and AI platform for designing CAR T cell therapies for solid tumors, funding its WAY-103 program toward an investigator-initiated trial.
Trends & Predictions
Frontier labs head toward public listings
Frontier lab financingStrengthenedContinues fromMay 29
Anthropic confidentially submitted a draft S-1 to the SEC roughly two weeks after its $65B Series H, earlier than the expected Q3 filing, giving it the option to go public.
- Prediction (Dec 2026): When Anthropic converts the draft into a public S-1 with audited financials, it sets the first public benchmarks for frontier-AI revenue quality, customer concentration, and compute margins, repricing peer private valuations and pressuring OpenAI's listing timeline.
- Resolves: Yes if by December 5, 2026 Anthropic has made a Form S-1 with audited financial statements public on SEC EDGAR.
Open-weight models diversify in architecture, modality, and size
Open-weight modelsStrengthenedRevivesMay 1
MiniMax M3 paired a 1M-token context and native multimodality with 59.0% on SWE-Bench Pro (weights pledged), NVIDIA released Cosmos 3 world models at 64B and 16B, JetBrains shipped Apache-2.0 Mellum2 (12B, 2.5B active), and H Company added quantized Holo3.1 computer-use checkpoints.
- Prediction (Dec 2026): The open-weight tier is judged mainly on sustained agentic and multimodal task completion rather than single-turn benchmarks, and at least one Western enterprise standardizes on a self-hosted open-weight model over a closed frontier API for cost and data residency.
- Resolves: Yes if by December 5, 2026 a Fortune 500 or FTSE 100 company publicly states (press release, earnings call, or engineering blog) that it has replaced a closed frontier API with a self-hosted open-weight model, citing cost or data residency.
Based on MiniMax M3NVIDIA Cosmos 3JetBrains' Mellum2H Company's Holo3.1
Frontier cyber capability ships to defenders behind gated access
AI cyber offense and defenseStrengthenedRevivesMay 15
Project Glasswing, whose ~50 initial partners surfaced more than 10,000 high- or critical-severity flaws with Claude Mythos Preview, extended to ~150 organizations in power, water, healthcare, and communications; Anthropic's threat study of 832 banned accounts found medium-or-higher-risk actors rising from 33% to 56%.
- Prediction (Dec 2026): A standards body such as MITRE begins formally cataloguing agentic-orchestration attacker techniques, and at least one government moves toward registration or licensing requirements for general-access frontier cyber models.
- Resolves: Yes if by December 5, 2026 MITRE ATT&CK has added a technique or sub-technique for AI-agent orchestration of attacks (per the ATT&CK release notes) and at least one national government has published a draft rule requiring registration or licensing of frontier models with offensive cyber capability.
Based on Anthropic expands Project GlasswingAnthropic maps a year of AI-enabled cyberattacks
Agent benchmarks expose reliability gaps and lose trust
Agent evaluation and reliabilityStrengthenedContinues fromMay 29
A benchmark of 474 executable games found frontier models solid on clean tasks (Gemini-3.1-pro 74.2%) but falling to 57.6% under counterfactual revision and pruning evidence they needed.
- Prediction (Dec 2026): Agent training pipelines adopt grounded, dependency-aware data synthesis and belief-revision-specific evaluation as static single-turn benchmarks fail to predict reliability in interactive deployment.
- Resolves: Yes if by December 5, 2026 at least two of OpenAI, Anthropic, Google DeepMind, and xAI report scores on an interactive belief-revision or partial-observability agent benchmark in a model or system card.
Frontier labs build services capacity to deploy into enterprises
Enterprise adoptionStrengthenedRevivesMay 15
OpenAI made its frontier models and Codex generally available on Amazon Bedrock, including GovCloud, and extended Codex to non-developers, now ~20% of its 5M+ weekly users; Anthropic's partner network added a Services Track after more than 40,000 firms applied.
- Prediction (Dec 2026): Multi-cloud availability of the leading closed frontier models becomes a baseline enterprise procurement expectation, eroding the model exclusivity that anchored hyperscaler AI differentiation.
- Resolves: Yes if by December 5, 2026 OpenAI's frontier models are generally available on at least three of AWS, Microsoft Azure, Google Cloud, and Oracle Cloud, per each provider's model catalog.
Based on OpenAI models and Codex arrive on AWSOpenAI extends Codex to non-developersClaude Partner Network adds a Services Track
In this brief
- Organizations