AI NEWS · THU9 min read

Aug 20, 2026

10 stories from this update.

Businessfortune.com

Anthropic is reportedly preparing an October stock market debut at a $2 trillion valuation, which would be the largest ever

The Financial Times reported that Anthropic, the company behind the Claude assistant, is planning to go public in October 2026 at a valuation of around $2 trillion. If that figure holds, it would be the biggest initial public offering in history, meaning the biggest first sale of a company's shares to ordinary investors, larger even than SpaceX. Investors behind the number expect Anthropic's annual revenue run rate, which is its current revenue projected out over a full year, to reach $100 billion to $120 billion by the end of 2026. None of this is settled: the report says "The IPO is still under discussion, and the valuation has not been formally fixed within the company, sources said." Anthropic has confirmed nothing publicly.

Read the source

Safetyopenai.com

OpenAI paused some of its own AI training after deciding a new model had crossed a critical cybersecurity threshold

On August 18, 2026, OpenAI published an unusual explanation of why it had deliberately slowed its own work. The company said that on August 7 it determined that its model Astra had met what its Preparedness Framework calls the "Critical cybersecurity capability threshold," meaning the model is skilled enough at finding and exploiting software flaws to be dangerous if misused or poorly supervised. OpenAI responded with "a two-week pause in reinforcement learning (RL) training," reinforcement learning being the method that improves a model by rewarding good attempts, and said "Our largest planned frontier RL run remains on hold" pending smaller evaluations. It also pointed to a recent security incident involving Hugging Face, which it called the "OpenAI-Hugging Face incident," as a reason to lock down its research environments, and paused frontier model inference in research clusters for runs that could execute code. New safeguards include stronger sandboxes for running code, network controls that block "unauthorized access to the internet or other internal networks," and monitoring that covers roughly "20% of the inference compute" with the aim to "issue an alert within 30 minutes after concerning activity" is surfaced.

Read the source

Businesscursor.com

SpaceX has acquired Cursor, the company behind the popular AI coding assistant

Cursor announced on August 14, 2026 that it has been acquired by SpaceX, completing a process that began in April, when the two companies announced a partnership to speed up model training. Cursor says the deal gives it "access to the largest fleet of GPUs in the world, giving us the compute to build stronger models that are also more economical to run." GPUs are the specialized chips used to train and run AI models. Cursor pointed to Grok 4.6, released the day before the announcement, as "an early look at what we can now build together." No purchase price was disclosed. The company said its goal has not changed: "We still want to help people with ambitious ideas spend less time writing code and more time solving harder problems."

Read the source

Toolsblog.google

Google Sheets can now turn a spreadsheet into an interactive mini app that you build just by describing it

Google launched Sheets canvas on August 13, 2026, a feature powered by its Gemini assistant that turns a Google Sheets spreadsheet into a custom, interactive "mini-app." You open a spreadsheet, select the "Create canvas" option in the Ask Gemini side panel, and describe in ordinary language what you want, with no formulas or programming involved. Google's own examples are a study tracker for school assignments and deadlines, a fantasy football command center for player stats, and a wedding seating chart you can rearrange by dragging. The canvas and the spreadsheet stay in step: "Whether you make changes to your canvas or to your sheet, updates are synced between the canvas and the original sheet in real time." Because the canvas lives as a tab inside the sheet, it can be shared with anyone the way a normal sheet is. It is available globally in English to Google AI Pro and Ultra subscribers, to Google Workspace Business and Enterprise Standard and Plus customers, and to Google AI Pro for Education subscribers.

Read the source

Businesstechcrunch.com

Stripe is reportedly buying OpenRouter, the service that routes app requests between AI models, for more than $7 billion

Bloomberg reported on August 16, 2026 that the payments company Stripe has agreed to acquire OpenRouter for more than $7 billion. OpenRouter is a single doorway that lets software send a request to any of more than 400 different AI models and pick one by capability, speed or budget rather than being locked into a single provider, and the company says it has 8 million users worldwide. The reported price is a steep step up: OpenRouter was valued at $1.3 billion after a $113 million Series B round in May 2026 backed by Sequoia, Andreessen Horowitz, Menlo Ventures and Alphabet's CapitalG. The Wall Street Journal had reported talks the previous month. OpenRouter chief executive Alex Atallah has described the company as "the equivalent of Stripe for AI, because it provides customers with a single access point for different systems and prevents lock-in." A Stripe spokesperson said the company "does not comment on rumors or speculation."

Read the source

Modelsblog.google

Google released Gemini 3.7 Flash and cut its price for developers in half through the end of 2026

Google released Gemini 3.7 Flash on August 13, 2026, calling it "our most intelligent workhorse model yet for coding and agents." Through December 31, 2026 it costs $0.75 per million input tokens and $3.75 per million output tokens, half of what the model will cost from January 1, 2027, when the price rises to $1.50 and $7.50. Tokens are the word fragments AI companies use to meter usage, so those figures are essentially the price of the text going in and the text coming out. Google reported sizeable gains over Gemini 3.6 Flash on its own benchmarks, including 43.6% against 34.4% on FrontierCode 1.1 Main, 65.3% against 49.0% on DeepSWE v1.1, and 30.4% against 17.0% on AutomationBench. Developers reach it through Google AI Studio, Google Antigravity and Android Studio, businesses through the Gemini Enterprise platform, and Google AI Pro and Ultra subscribers get it in Gemini Spark.

Read the source

Hardwarecerebras.ai

Cerebras unveiled the CS-4, an AI computer it says answers up to 30 times faster than systems built on GPUs

On August 18, 2026, the chipmaker Cerebras announced the CS-4, a rack sized AI system built from three of its Wafer Scale Engine 3 Turbo processors. Cerebras says it is "up to 30 times faster than GPU solutions," measured on the GPT-OSS-120B model at more than 4,400 tokens per second for a single user, up to twice as fast as the previous CS-3, and delivers "up to 10x more throughput per watt than the CS-3." First shipments began in the quarter of the announcement, and the company did not disclose pricing. Speed matters here because it is the difference between waiting for an AI answer and watching it appear as you read. Chief executive and co-founder Andrew Feldman said, "In AI, speed is productivity. Historically, fast inference meant using smaller and less capable models. Cerebras CS-4 delivers industry-leading speeds on the largest frontier models, fundamentally changing the paradigm." Chief technology officer and co-founder Sean Lie put the benefit in terms of AI agents: "Being 30 times faster doesn't just make a response feel fast. It gives an agentic system room for more than an order of magnitude as much reasoning, verification, or tool use in the same wall-clock time."

Read the source

Modelsdecrypt.co

Chinese lab Z.ai released GLM-5.3, a coding model it plans to publish as open weights within weeks

Z.ai released GLM-5.3 on August 14, 2026, a 743 billion parameter model built for software coding work. The company says the only thing it changed was more post-training, the extra practice stage that comes after a model's initial training: "Scaling post-training is all we did for GLM-5.3." The result solves more while using fewer words. It scored 34.5% on Z.ai's own Code Bench at maximum effort using about 75,000 output tokens per task, against 23.4% and 96,000 tokens for GLM-5.2. It still trails closed rivals on Terminal Bench 3.0, scoring 28.3 against 33.7 for Claude Fable 5 and 34.6 for GPT-5.6 Sol. Z.ai says the weights are "set for public release in about two weeks" after safety evaluations, meaning anyone will then be able to download and run the model on their own machines, and its API pricing runs at roughly a tenth of US frontier rates. The company also reported a sharp jump in cybersecurity ability: GLM-5.3 tops the CyberGym benchmark at 84.5% and "flagged 2,436 vulnerabilities across 269 open-source projects, 1,097 of them medium-to-high severity."

Read the source

Safetyopenai.com

OpenAI previewed a way to catch AI misuse across many conversations without its staff ever seeing customer data

On August 19, 2026, OpenAI described a new system called Private Safety Processing, aimed at business customers who use its models under a zero data retention agreement, meaning OpenAI does not keep their prompts and responses after processing them. The problem it addresses is that safety checks usually look at one conversation at a time, while the most serious misuse, such as someone repeatedly probing safeguards or coordinating across several accounts, only becomes visible when you can compare many interactions. Under the new approach the content stays either on the customer's own infrastructure or on OpenAI systems encrypted with keys the customer controls, and automated checks look for patterns there. When something concerning turns up, OpenAI receives only "a narrowly defined signal indicating the type of activity involved" rather than the underlying text. Images flagged as possible child sexual abuse material are still retained, because reporting them is a legal requirement. OpenAI said the rollout and a technical white paper are planned for September 2026.

Read the source

Conceptshuggingface.co

Hugging Face's mid-year review finds Chinese labs releasing the biggest open AI models while tiny models dominate actual downloads

Hugging Face, the main public library where AI models are shared, published its review of open models from January through August 2026, and two findings are worth knowing even if you never download a model. The first is geography. Chinese labs released the largest open models this year, ranging from 754 billion to 2.78 trillion parameters in a given month, while US labs stayed under 130 billion in five of the seven months. The report puts it plainly: "In almost every month of 2026, the largest and most performant open model from a Chinese lab was larger than any model an American lab released." Chinese releases above 20 billion parameters also came with permissive licenses, 59% Apache 2.0 and 22% MIT, which means few strings attached for anyone reusing them. The second finding is a gap between attention and use. Models under 1 billion parameters accounted for 83% of all downloads and models above 100 billion for just 1%, and exactly one repository appears in both the 25 most downloaded and the 25 most liked lists. In other words, the models people talk about are not the small ones quietly doing the work. Alibaba's Qwen family has become the community's default starting point, with 151,448 models built on top of it, 2.6 times Meta's total.

Read the source

All AI news & updates