What Google, OpenAI, and Anthropic aren't saying out loud
Every URL the pipeline pulled into ranking for this issue — primary sources plus the supporting and contradicting findings each Researcher returned. Inline citations in the issue point back here.
Sources
We’re launching two specialized TPUs for the agentic era. blog.google
The eighth generation of Google’s TPU includes two specialized chips that will power the future of AI.
Google unveils two new TPUs designed for the “agentic era” arstechnica.com
Google’s new generation of Tensor AI chips is actually two chips, one for inference and one for training.
Introducing OpenAI Privacy Filter openai.com
OpenAI Privacy Filter is an open-weight model for detecting and redacting personally identifiable information (PII) in text with state-of-the-art accuracy
Is Claude Code going to cost $100/month? Probably not - it’s all very confusing simonwillison.net
Anthropic today quietly (as in silently , no announcement anywhere at all) updated their claude.com/pricing page (but not their Choosing a Claude plan page , which shows up first for me on Google) to add this tiny but significant detail (arrow is mine, and it’s already reverted ): The Internet Archive copy from yesterday shows a checkbox there. Claude Code used to be a feature of the $20/month Pro plan, but according to the new pricing page it is now exclusive to the $100/month or $200/month Ma…
Anthropic tested removing Claude Code from the Pro plan arstechnica.com
Untenable demand has Anthropic exploring new approaches to rationing its service.
GPT-5.5 Bio Bug Bounty openai.com
OpenAI opens a red-teaming challenge against GPT-5.5’s bio-risk safeguards, paying up to $25,000 for universal jailbreaks that bypass biosecurity refusals. The bounty narrows scope to bio-specific harms rather than general jailbreaks, marking OpenAI’s first domain-targeted bounty tied to a frontier model release.
Introducing ChatGPT Images 2.0 openai.com
OpenAI ships its second-generation image model under two names — ChatGPT Images 2.0 in the consumer app and GPT-Image-2 via API — touting improved text rendering, multilingual prompts, and visual reasoning. Latent Space’s coverage frames it as OpenAI’s response to recent Google and Black Forest Labs image releases.
(AINews) OpenAI launches GPT-Image-2 latent.space
with Cursor getting a $10B contract with xAI and a right to acquire for $60B.
Introducing workspace agents in ChatGPT openai.com
OpenAI launches workspace agents in ChatGPT, Codex-powered cloud agents that execute multi-step workflows across connected enterprise tools. Pitched at teams rather than individual developers, the agents run asynchronously in sandboxed cloud environments and inherit organization-level permissions for secure access to company data.
Changes to GitHub Copilot Individual plans simonwillison.net
GitHub tightens Copilot Individual plans, pausing new signups, restricting Claude Opus 4.7 to the $39/month Pro+ tier, and dropping older Opus models, citing agentic workflows that consume far more compute than the original per-request pricing assumed. New token-based session and weekly limits replace the request-count model.
Shopify CTO Mikhail Parakhin details the company’s 2026 AI adoption surge in a Latent Space interview, including an unlimited Claude Opus 4.6 token budget for engineers and internal tools Tangle, Tangent, and SimGym used to scale AI across the merchant platform.
Mozilla: Anthropic’s Mythos found 271 security vulnerabilities in Firefox 150 arstechnica.com
Mozilla used Anthropic’s new Mythos security model to audit Firefox 150, surfacing 271 zero-day vulnerabilities ahead of release. Mozilla’s CTO calls the model “every bit as capable” as top human security researchers, marking one of the largest disclosed AI-driven vulnerability hauls in a shipping browser.
Making ChatGPT better for clinicians openai.com
OpenAI offers ChatGPT for Clinicians free to verified U.S. physicians, nurse practitioners, and pharmacists, with HIPAA-aligned handling for clinical documentation, decision support, and research workflows. Verification gates access, distinguishing the tier from consumer ChatGPT and from OpenAI’s enterprise health pilots with hospital systems.
10 Things That Matter in AI Right Now technologyreview.com
LLMs+ technologyreview.com
When ChatGPT launched as an experimental prototype in late 2022, OpenAI’s chatbot became an everyday everything app for hundreds of millions of people. LLMs like ChatGPT were the new future: The entire tech industry was consumed by the inferno, with companies racing to spin up rival products. The ashes of the old tech world still…
Supercharged scams technologyreview.com
When ChatGPT was released to the public in late 2022, it opened people’s eyes to how easily generative AI could churn out vast amounts of human-seeming text from simple prompts. This quickly caught the attention of criminals, who soon began using large language models to produce malicious emails—both the untargeted spam kind and more sophisticated,…
World models technologyreview.com
AI systems have already gained impressive mastery over the digital world, but the physical world is still humanity’s domain. As it turns out, building an AI system that can compose a novel or code an app is far easier than developing one that can fold laundry or navigate a city street. To get there, many…
Weaponized deepfakes technologyreview.com
For years, experts have warned that deepfakes—AI-generated videos, images, or audio recordings of people doing or saying things they haven’t actually done in real life—could be deployed in malicious ways. These dangers are now here. Improvements in deepfake technology, and the widespread availability of easy-to-use and cheap (or free) generative models, have made it easier…
Indian med student rakes in thousands with AI-generated MAGA hottie arstechnica.com
“Emily Hart” is a young, AI-created conservative woman who likes to take off her clothes.
(AINews) Tasteful Tokenmaxxing latent.space
a quiet day lets us reflect on the top conversation that AI leaders are having everywhere.
Pentagon wants $54B for drones, more than most nations’ military budgets arstechnica.com
The proposed Pentagon drone investment rivals Ukraine’s entire military budget.
3 new ways Ads Advisor is making Google Ads safer and faster blog.google
The Road to Google Marketing Live graphic. May 20, 2026, 9am PT. White background with colorful bottom gradient.
Partnering with industry leaders to accelerate AI transformation deepmind.google
Google DeepMind partners with global consultancies to bring the power of frontier AI to organizations around the world.
That’s my designer - Claude bensbites.com
and it comes with a new model, Opus 4.7
AI needs a strong data fabric to deliver business value technologyreview.com
Artificial intelligence is moving quickly in the enterprise, from experimentation to everyday use. Organizations are deploying copilots, agents, and predictive systems across finance, supply chains, human resources, and customer operations. By the end of 2025, half of companies used AI in at least three business functions, according to a recent survey. But as AI becomes…
References
Nand Research analysis nand-research.com
TPU 8t increases memory capacity by 12.5% over Ironwood, [but] its memory bandwidth per chip is actually 11.5% lower, a deliberate trade-off likely aimed at improving silicon yield and reducing costs… the 8th generation marks the end of Broadcom’s exclusivity, with MediaTek joining as a co-designer for the inference-focused 8i.
Tom’s Hardware tomshardware.com
A single TPU 8t chip delivers approximately 12.6 PFlops (FP4)—trailing Nvidia’s R200 in pure compute… Google claims a 2.7x improvement in price-performance over Ironwood and an 80% boost in inference value-for-money on the 8i.
Pasquale Pillitteri (Jensen Huang on Dwarkesh Podcast) pasqualepillitteri.it
Huang dismissed the narrative of a ‘customer defection,’ specifically labeling Anthropic’s massive 3.5 GW commitment to Google silicon as a ‘unique instance, not a trend’… and challenged Google to submit the TPU 8 series to public benchmarks like MLPerf Inference Max, stating that ‘no rival has so far accepted the public challenge’.
Aragon Research aragonresearch.com
Unlike Nvidia’s general-purpose GPUs, which can be dynamically reallocated from training to inference based on demand, Google’s split forces providers to commit to a specific workload profile, potentially leading to stranded capacity if market needs shift.
Futurum Group futurumgroup.com
The TPU 8i abandons the longstanding 3D Torus interconnect for ‘Boardfly,’ a Dragonfly-inspired hierarchical topology that reduces network diameter by 56%, paired with a Collectives Acceleration Engine replacing SparseCore to speed autoregressive decoding… triples on-chip SRAM to 384 MB to hold KV cache locally.
Hacker News / Moomoo developer thread moomoo.com
Veteran developers recall ‘undocumented behavior and bugs’ in older PyTorch/XLA implementations, with some viewing the new native TorchTPU backend as a necessary admission that previous migration paths were too brittle for production… the real complaint is not the silicon but ‘GCP friction’—quota provisioning and generic error messages on large pods.
Ed Zitron, Where’s Your Ed At wheresyoured.at
Anthropic and OpenAI ‘conned’ users by offering flat-rate subscriptions that were mathematically impossible to sustain… Anthropic may be losing 200% to 3000% on its most active power users.
Business Insider businessinsider.com
CEO Sam Altman publicly mocked Anthropic’s move, emphasizing that OpenAI’s coding tools would remain accessible on lower-tier plans.
The Register theregister.com
the move was driven by ‘untenable demand,’ as Anthropic sought new methods to ration its services amid tightening compute capacity
GitHub: zacdcook/openclaw-billing-proxy github.com
the community developed ‘openclaw-billing-proxy’ on GitHub, a working tool that uses 7-layer bidirectional processing to inject Claude Code identifiers into requests, attempting to bypass Anthropic’s detection and maintain subscription-based billing
India Today (citing Anthropic’s revised docs) indiatoday.in
Anthropic more than doubled its estimated spending for active developers from $6 to $13 per day, with the 90th percentile of heavy users projected to spend $30 per day
VentureBeat venturebeat.com
Anthropic identified three specific errors… a system prompt instruction designed to curb verbosity—capping responses to 25 words—tanked coding quality across Opus 4.6 and 4.7
Tonic.ai benchmark report tonic.ai
OPF’s default recall plummeted to just 10% for web-crawl data and 38% for Electronic Health Record (EHR) notes, while Tonic’s own production redactor maintained F1 scores between 0.92 and 0.99 on the same data.
Hacker News discussion news.ycombinator.com
Because the model is probabilistic, it is unsuitable for deterministic secrets like passwords… a high-accuracy PII redactor can be trivially inverted to act as a PII extractor, allowing attackers to find sensitive data in unstructured text with minimal effort using the model’s own taxonomy.
AI Engineer Weekly (Substack) — ‘Privacy Filter is not an LLM’ aiengineerweekly.substack.com
Redacted text is often practically useless for complex tasks, as LLMs struggle to reason without the context provided by specific names or entities; de-sanitization mapping must happen client-side after the LLM responds.
Grepture comparison (Presidio/Piiranha/OPF) grepture.com
Microsoft Presidio still leads in structured PII (like SSNs and credit cards) due to its regex-based recognizers, whereas transformer-only models like Piiranha often fail to generalize beyond their synthetic training distributions.
StartupFortune — ‘Kills the last excuse against enterprise AI’ startupfortune.com
By providing the compliance infrastructure needed to bypass data-sharing objections in healthcare and legal sectors, OpenAI effectively ‘kills the last excuse’ against adopting frontier models like GPT-5.
UK AISI — Boundary Point Jailbreaking post aisi.gov.uk
BPJ was the first automated method to successfully defeat OpenAI’s input classifiers for frontier models without requiring human-led seed attacks, using single-bit blocked/unblocked feedback to iteratively bypass classifiers.
