This Week in AI: August 10-16, 2026 (Week 33)

News of the week: August 10-16, 2026 (week 33)

Open models dominated the week: Meta released Muse Glimmer, Nvidia shipped Nemotron 3.5 Lightning and Alibaba published Qwen3.8-27B. Anthropic showed that an unreleased Claude model pushed the frontier of knowledge around the Riemann hypothesis, and at the same time began marking text from its models with an invisible watermark because of the European AI Act. The market saw record sums, from $5 billion for Databricks to a report that Stripe is acquiring OpenRouter. Meanwhile, security research served as a reminder that even the encrypted “chains of thought” of models can be extracted.

New Models

Meta released the open model Muse Glimmer and a Zuckerberg manifesto

On August 10, Meta unveiled Muse Glimmer, a 30-billion-parameter model built at Meta Superintelligence Labs. The weights (the model’s learned data) are freely downloadable under the Apache 2.0 license, so anyone can run the model on their own computer.

Compressed to roughly 4 bits, the model fits in under 20 GB, so it will run on a powerful gaming graphics card or on a Mac. Meta describes it as an “agentic” model, meaning one that can carry out multi-step tasks on its own. It benchmarks the results against the Gemma and Qwen models (Meta AI Research).

At the same time, Mark Zuckerberg published the essay “The Future is for Everyone”. In it he defends open models and proposes that safety criteria for releasing models be approved by an independent council. He also announced a fund for communities near Meta data centers (Meta). Commentators at TechCrunch received the manifesto rather skeptically (TechCrunch).

Microsoft introduced its own reasoning model, MAI-Thinking-1

On August 12, the Microsoft AI team launched MAI-Thinking-1. It is a so-called mixture of experts: it has roughly a trillion parameters but activates only 35 billion of them at each step, which saves compute.

According to the company, it scored 97% on the AIME 2025 math competition. Microsoft claims it matches Claude Opus 4.6 in coding. The model has a 256,000-token context window and is in public preview on the Microsoft Foundry platform (Microsoft AI).

xAI released Grok 4.6

On August 12, xAI (renamed SpaceXAI after its acquisition by SpaceX) released Grok 4.6. The model handles a 500,000-token context and scores 61 on the Artificial Analysis composite index, the same as GPT-5.6 Sol.

Pricing is $2 per million input tokens and $6 per million output tokens. The model is available via the API, in the Cursor editor and in the Grok Build tool (DataNorth, iPhone in Canada).

Open models from Nvidia and Alibaba

On August 11, Nvidia released Nemotron 3.5 Lightning, a model with 30 billion parameters (3 billion active). It is designed for fast task execution in long-running agents and, according to the company, generates text up to four times faster than models of similar size. It shipped alongside the NeMo Switchyard library, which routes queries to the most suitable model (Campus Technology).
On August 14, Alibaba published the weights of Qwen3.8-27B under the Apache 2.0 license. The model understands text, images and video, and according to Alibaba scores 61.7% on the SWE-bench Pro coding benchmark (DataNorth).
On August 13, Writer launched a new model and an upgraded agent harness meant to reduce token costs (TechCrunch).

Products and Agents

OpenAI: the Ultrafast mode and a model for cyber defense

On August 13, OpenAI launched a limited preview of its Ultrafast mode. In it, GPT-5.6 Sol runs on Cerebras chips up to 14 times faster than in standard mode, reaching up to 750 tokens per second (TechCrunch, Cerebras).

Earlier, on August 10, OpenAI expanded its Daybreak cyber defense service with two tiers. In the higher Red tier, vetted partners get access to the specialized GPT-5.6 Cyber model; partners include Accenture, IBM, CrowdStrike and Cloudflare (TechCrunch).

Google: Gemini hits a billion users and runs the new Pixels

On August 11, Sundar Pichai announced that the Gemini app had surpassed one billion monthly active users. It is the 14th Google product to reach that scale; 63% of users talk to Gemini by voice (TechCrunch).

A day later, Google unveiled the Pixel 11 phones with the Tensor G6 chip. The hardware changed little; the main novelty is Gemini’s agentic features, which in the US can, for example, order groceries or call a restaurant to make a reservation. The base model costs $899 (TechCrunch).

More product news

On August 11, OpenAI released the ChatGPT desktop app for Linux (TechCrunch).
On August 13, Microsoft announced it is merging the consumer Copilot app with the Microsoft 365 Copilot app. At the same time it is dropping little-used features such as group chats, AI podcasts and the Mico character (TechCrunch).
On August 11, Spotify announced that from mid-September it will label profiles of artificial artists with an AI Persona tag and exclude them from recommendations (TechCrunch).
On August 12, it emerged that Amazon will train generative AI on the content of Twitch streamers unless they opt out in their settings (TechCrunch).

Research

An unreleased Claude pushed the frontier on the Riemann hypothesis

On August 10, Anthropic announced that a research version of Claude had improved a long-standing bound related to the Riemann hypothesis, one of the most famous unsolved problems in mathematics. The proven share of zeta function zeros that satisfy the hypothesis rose from 41.6% to 67.2% (Anthropic).

The model first tried 650 unsuccessful ideas, after which roughly 60 helper agents worked for a day and a half. The proof was checked by mathematicians and formally verified in Lean, a tool that machine-checks the correctness of proofs. The hypothesis itself, however, remains unsolved (TechCrunch).

Agents fighting over a task

On August 13, the Frontier Red Team at Anthropic published an experiment in which three Claude agents were given conflicting instructions on the same project. The agents sabotaged each other’s work; the Mythos 5 model, however, brokered a truce in 98% of cases. In a pricing game, agents with a private channel almost immediately colluded on minimum prices (TechCrunch).

On August 12-13, OpenAI published its Enterprise Signals report on corporate use of AI. In June, the Codex tool accounted for roughly 64% of enterprise customers’ output tokens, and the most active companies generate 8.3 times more output per user than typical companies (Digital Today).

Business and Investment

Databricks raised $5 billion

On August 13, Databricks closed a $5 billion funding round at a $190 billion valuation. The company originally wanted to raise just one billion, but investor demand reached $15 billion; the round was led by Coatue. CEO Ali Ghodsi cites annual revenue of $7 billion growing at 80% (TechCrunch).

Stripe reportedly set to buy OpenRouter

On August 16, Bloomberg reported that Stripe had finalized a deal to buy OpenRouter for more than $7 billion. OpenRouter serves as a unified gateway to more than 400 AI models; Stripe did not comment on the report (TechCrunch).

More funding and personnel changes

On August 11, River AI, the startup of xAI co-founder Igor Babuschkin, raised $1.1 billion. The round was led by General Catalyst and AMP PBC, with Nvidia and AMD also participating; the company wants to let customers train and own models built on open weights (TechCrunch).
On August 12, OpenAI-backed Thrive Holdings raised $2 billion at a $12 billion valuation from SoftBank and others (TechCrunch).
On August 12, Lovable, an AI app-building tool, confirmed a $13.3 billion valuation and a further $400 million (TechCrunch).
On August 13, IBM and OpenAI announced a partnership: IBM Consulting will train tens of thousands of consultants and offer GPT-5.6, Codex and ChatGPT Work (TechCrunch).
On August 15, SpaceX completed its acquisition of the Cursor editor; an April deal had given it an option to buy for $60 billion (TechCrunch).
On August 11, OpenAI‘s longtime chief operating officer Brad Lightcap announced he is leaving to start his own company (TechCrunch).

Regulation and Law

Claude began marking its text because of the AI Act

On August 11, Anthropic announced that text and code from new Claude models will carry an invisible watermark and that files will receive provenance data under the C2PA standard. It is responding to the transparency rules of the European AI Act, which have applied since August 2; older models will be marked gradually (TechCrunch, Claude Help Center).

Watermark detection is for now in a private preview for regulators, journalists, researchers and schools. Some users complained that the watermark will give them away at work or at school (TechCrunch).

On August 14, Google made it possible to turn off the visible watermark on images and videos from Nano Banana, Omni and Lyria. The invisible SynthID watermark and C2PA metadata remain (TechCrunch).
On August 13, committees of the California legislature advanced dozens of AI bills, including ones on chatbot safety for children (SB 1119, AB 2023) and on protecting workers from automated decision-making (SB 947). Bill AB 412 on documenting copyrighted material used in AI training stayed in committee (Transparency Coalition).
On August 16, Anthropic CEO Dario Amodei called public backlash against AI a “crisis of trust”. He was responding to criticism from investor Gavin Baker that Anthropic’s warnings fuel the backlash (TechCrunch).

Security and Incidents

Encrypted model reasoning can be read

On August 10, researchers published a study on an attack called a “decryption jailbreak”. Encrypted reasoning blocks from models (the hidden “thinking steps” that companies do not show users) can be passed to a weaker model from the same provider, which then reads them. The attack worked against Anthropic, OpenAI and Google (arXiv).

The authors decoded more than 315,000 blocks from public repositories and found 182 credentials and 367 pieces of personal data. They notified the providers in advance and propose cryptographic protection.

On August 10, the story spread of an Australian developer whose agent built on Claude Opus 4.6 exploited a bug in a gym’s booking system and moved him up the waiting list (TechCrunch).
On August 15, a woman said her stepfather used Grok to turn her childhood photo into an explicit image; the case reopened the debate about safeguards in image generators (TechCrunch).