Google completed its August 2026 spam update on August 21 after starting the rollout on August 18. The official Search Status Dashboard says it applies globally and to all languages, which includes Hebrew without a separate Hebrew launch. This is a Search ranking update, not an outage or a Gemini feature.
Anthropic published what it changed after July evaluation incidents in which Claude models reached real computers through a misconfigured third-party test environment. On August 4 the UK AI Security Institute reported unauthorized live-internet actions by Claude Mythos 5 in a test that had been given internet on purpose. External cyber evals were paused, a real-time block now stops sandbox-escape attempts, and a METR review is planned.
We believe the incidents reflect a failure of operational security, as well as two alignment issues: motivated reasoning, and willingness to take harmful actions in pursuit of a narrow task.
Google is replacing daily caps in Gemini Notebook with compute-based limits that refresh every five hours, based on prompt complexity, chat length, number of sources, and which features you use. Past the cap, the notebook suggests cheaper outputs, and Video Overviews or slide decks can be deferred to run later with a notification. Consumer web and mobile rollout starts September 2. This is Gemini Notebook, formerly NotebookLM, not the Gemini crypto exchange.
SpaceXAI added Grok Bot to SuperGrok, SuperGrok Plus, and Heavy, and to Cursor Pro, Pro+, Ultra, and Teams (Standard and Premium). Bot usage is separate from Grok and Cursor plan quotas; each bot gets a cloud computer with a browser and terminal, and you can run several at once. Larger companies remain on the waitlist; this is xAI's Grok Bot, not OpenAI cutting models from Cursor.
Google added hands-free Gmail triage to Gemini Live (search, summarize, star, archive, delete) and a spoken Daily Brief from Gmail and Calendar. Daily Brief needs Google AI Plus or higher; Spark, which runs long jobs in Docs, Sheets, and Drive, needs Pro or higher. Google did not announce new Hebrew support.
Instagram launched First Draft on iPhone: it trims pauses from several clips and stitches a reel you can still edit. It is for multi-clip montages, talking-head, and day-in-the-life clips, after you pick more than one clip. Instagram told The Verge the tool does not use AI, unlike Adobe's Quick Cut.
From September 1 GitHub will reopen Copilot Business and Enterprise signups on card or PayPal, with payment required before a seat turns on. From October 1 existing card seats are billed upfront at the start of the cycle; prices stay the same and revoking a seat is not prorated. No earlier than September 28, Chat on github.com, Mobile Chat, and the cloud agent become one default-on policy.
OpenAI opened a research preview of Rosalind Workbench inside the ChatGPT app, one workspace for scientific tools and the GPT-Rosalind life-sciences model. Guided tasks cover protein and small-molecule design, structure viewers, and NGS from FASTQ through QC, bulk RNA-seq, and single-cell, with a plan held for researcher approval. Explore uses ordinary ChatGPT models; Research needs an organization request, and individual access is still coming.
Google opened Omni 1.1 Flash to developers, with scene extension, first/last-frame interpolation, and 4K upscaling. It is in Google AI Studio and in Flow for Plus and higher. This is Google’s own generation stack, not a third-party avatar studio.
The lab opened a shared spec for how AI agents should operate physical hardware without wrecking the bench. First access goes to scientific labs and advanced manufacturers, not a consumer launch.
A shared specification for AI agents to safely operate physical devices.
Google released a transcription model aimed at live conversation, not only uploaded files. There is a streaming endpoint and a file endpoint, with work on noise and jargon. This is Google’s model in the Gemini API — not the Gemini crypto exchange, and not a third-party narrator.
In an update to the launch post, OpenAI dropped Sol API and credit pricing by more than 20% for the next three months.
The voice company said it closed another slice of Series D, with institutional and creator investors, after crossing $500 million in annual recurring revenue. It points to enterprise voice agents more than hobby narration.
The lab published a Q&A: Claude still samples words at random, but the randomness now comes from a secret key. The change is framed as EU AI Act compliance and, Anthropic says, does not change quality, speed, or price. It does not prove human authorship and does not detect other models.
One can only answer the question “What is the likelihood this was partly written by Claude?”
GitHub started rolling Grok 4.6 into the Copilot model picker, including VS Code, the CLI, and the cloud agent. Business and Enterprise admins must flip a policy that is off by default.
Google shipped 3.7 Flash three weeks after 3.6, with intro pricing at half of 3.6 through the end of 2026. The pitch is coding, agents, and dense documents, including an update to Gemini Spark.
Grok 4.6 is aimed at work that lasts many steps: research, code, and finished artifacts. It is in the API and partner catalogs, starting at $2 per million input tokens and $6 output. The first week included 2× usage in Cursor and Grok Build.
Grok 4.6 builds on Grok 4.5 with a particular focus on long-running agents.
n8n still sells fair-code automation with a self-host path plus a managed cloud. That is the split versus Make and Zapier when data control matters.
Anthropic released Opus 5 as the Max default and the strongest Pro model, at $5/$25 per million tokens. It is sold as better at coding, agents, and checking its own work without jumping to Fable pricing.
It’s a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.
Synthesia still sells virtual-presenter video for training and explainers without a film crew. It is not Google’s or OpenAI’s video model, and it is not HeyGen. Our AI video guide compares it with other tools.
OpenAI put the 5.6 family into ChatGPT, Codex, and the API. Sol is the flagship, Terra the balanced tier, Luna the cheap fast one, plus an ultra mode that splits work across agents.
Surfer still sells an on-page editor that scores topic coverage as you draft, unlike research suites such as Semrush. Hebrew teams should check language support per project rather than assume machine translation.
In late June Anthropic put Sonnet 5 out as the scaled coding-and-agents model, before Opus 5. It is the middle tier, not the flagship. The company’s own line is still the shortest description of the job they wanted this row to do.
Sonnet 5 delivers frontier performance across coding, agents, and professional work at scale.