Wednesday, 7 October 2026

Anthropic has put Claude inside Google Docs, Sheets and Slides. Claude for Google Workspace is in public beta on all paid Claude plans, and it opens as a sidebar next to the file you have open [1]. Android Authority lists those plans as Pro, Max, Team and Enterprise [2]. Claude can read the open doc, sheet or deck, see what you have selected and change the file directly, the company says [1].

In Docs it can fix a sentence in place or propose larger rewrites as suggestion cards you apply or dismiss [1]. In Sheets it writes formulas, builds pivot tables and native charts, adds tabs, and can pull a range into Python for joins or cleaning and write the result back [1]. In Slides it builds new slides from the deck's own layouts and flags elements that overlap or run off the slide [1]. The default mode, "Ask before edits", shows each change as an approval card; "Accept all edits" lets Claude work through a task without stopping [1][2].

A second release runs the other way: new Google Docs, Sheets and Slides connectors, also in beta, let Claude create and edit Google files from the Claude chat [1][3]. Claude's access matches your existing Google sharing permissions, Anthropic says, and on Team and Enterprise plans an owner has to enable the connectors first [1]. Admins can deploy the add-on to a domain or to selected groups from the Google Admin console [1].

No separate price has been announced; access comes with existing paid plans, according to [3]. VentureBeat notes that Gemini and Microsoft Copilot already offer similar editing in their own suites [3].

ChatGPT is getting a new interface OpenAI calls Intelligent UI, rolling out with a new GPT-6 model, according to [1]. Answers can include interactive elements such as tappable buttons, a calculator built for the task, interactive charts and editable graphs [1]. OpenAI's stated goal is to make "learning complex topics easier" [1]. Users who want fewer visuals will be able to dial them back [1].

The rollout reaches Pro, Plus, Business and Enterprise users first and Free and Go users the following day, according to [1]. Xataka puts the Free and Go start at 8 October and reports that paid plans get a version based on GPT-6 Sol, while Free and Go use GPT-6 Luna [2]. The change covers the Chat experience only; the models used in Work and Codex do not change in this release, according to [2].

Xataka also reports that the model can start answering while it is still working in the background. By OpenAI's internal figures, on questions that need a web search, GPT-6 Instant starts answering 44% sooner on average than GPT-5.6 Instant [2]. Images in answers can come from the open web, and the model decides when it needs to generate a new one [2]. OpenAI says the extra visuals should not make users hit their usage limits more often, according to [2].

If you rely on ChatGPT for text you paste elsewhere, check whether the new answers still copy cleanly.

OpenAI's Decisions API is now in public beta, and OpenAI says it expects general availability in the coming weeks [1]. It was first announced at last week's DevDay [2]. The API takes text, images or both and returns typed answers rather than prose: the probability that a condition is true, a choice from a fixed set, or a score against ordered levels [1]. OpenAI says it answers about 10x faster than the Responses API [1]. The only model is gpt-6-luna, on a dedicated /v1/decisions endpoint [1].

Pricing is input-only: $0.10 per 1M input tokens, with no output-token or cache charges [1]. Zero Data Retention and HIPAA use are supported for eligible customers, with regional processing in the United States and Europe [1]. Simon Willison notes that Jev, the decision model from TypeSafe AI, charges 4.2 cents per million input tokens, but OpenAI's model also accepts images [2][4].

Liquid AI released d1-3B and an experimental d1-omni-600M; it says d1-3B answers a question in 16 ms on an NVIDIA Jetson AGX Thor [3]. The Strands Agents team released Strands Decider 2B as open source with weights and training data, and reports a median of around 115ms per decision on an RTX 3090 [4]. Strands says these models suit routing, tool selection, guardrails and policy classification, and are worse than reasoning models at complex problems [4].

For a small team, the practical use is cheap triage: route a ticket, flag a risky tool call or label a message before a larger model sees it.

Google has opened its SynthID Detector to everyone, available globally in English from today [1]. Anyone can check whether an image, video or audio file was made with AI from Google or its partners, which Google lists as OpenAI, NVIDIA, Kakao and, soon, Apple [1]. Until now, people outside Google's trusted tester program could only check by asking Gemini, and the old detector only read Google's own watermark, according to [3].

Google says it has watermarked over 180 billion images and videos and 240,000 years of audio since SynthID launched in 2023 [1]. Its built-in checks in Search, the Gemini app and Chrome now handle over 1 million requests daily, the company says [1]. The site accepts common image formats including JPG, PNG, WEBP and HEIC, video in MP4, MOV and WEBM, and audio including WAV and MP3 [2].

There are limits. You must log in with a Google, OpenAI or Apple account, and Google says each user gets "a daily usage quota of approximately 10 image, video, and audio checks" [3]. The result is a plain message on whether SynthID was used, without the highlighted regions of the internal tool [3]. Meta uses different watermarking, some providers use none, and open-weight models run at home leave no watermark, so a clean result does not prove a file is real [3].

For anyone vetting a supplier's images or a viral clip, this is a free first check, not a verdict.

OpenAI says College Planner is coming "soon" to ChatGPT for Teens, the default experience for accounts identified as under 18 [1][2]. It brings application requirements, deadlines, tasks and financial-aid steps for each school on a student's list into one plan [1]. The first US version targets students in grades 10–12 planning to attend a four-year college [1]. OpenAI is also supporting College Advising Corps, a nonprofit that places advisers in under-resourced high schools [1]. The Verge reports new study tools too: multi-photo capture on iOS that turns photos into one PDF, and flashcards generated from notes [2].

OpenAI published usage figures with the launch. It says teens spend under 15 minutes a day on ChatGPT on average and fewer than 2% spend more than three consecutive hours [1]. In one week, nearly 1.2 million teens used Learning Visualizations and more than 180,000 used Study Mode, the company says [1].

The announcement follows a Common Sense Media assessment that called ChatGPT for Teens an "unacceptable risk" [3]. The group says it "doesn't send alerts to parents when it should" and "still does kids' homework" [3]. OpenAI says much of the testing may have run before parental controls finished activating; Common Sense says some accounts were linked well beyond that window and still produced no alerts [3].

Parents should treat the alerts as unproven until an independent test shows they work.

Google has released EmbeddingGemma 2, an open embedding model that maps text, code, images, video and audio into one shared space [1]. It has 740 million parameters, is built on the Gemma 4 architecture and ships under the Apache 2.0 license, which allows commercial use [1]. Text-only work needs as little as 270M parameters; the vision (170M) and audio (300M) encoders are optional [1].

Vectors can be cut from 768 dimensions down to 512, 256 or 128, which Google says gives up to 6x less storage in a local vector database [1]. On code search, Google reports an MTEB Code score of 78.68, up from 68.76 for the first EmbeddingGemma [1]. That first version passed 20 million downloads, the company says [1]. Weights are on Hugging Face and Kaggle, and the model runs in tools including transformers, sentence-transformers, llama.cpp, Ollama and LM Studio [1].

Simon Willison argues the license matters more for embeddings than for chat models [2]. Applications store thousands or millions of vectors; if a hosted, closed model is retired, the owner has to pay to re-embed everything [2]. With open weights, he notes, you can keep paying a provider but fall back to running the same model yourself [2].

For a small team building search over its own files, this is an on-device model whose weights you keep even if a host stops serving it.

Anthropic has expanded its Cyber Verification Program into three access tiers that lift some of the cyber blocks on its models for vetted security teams [1]. Each tier includes Claude Opus 5.5, Claude Sonnet 5.5 and Claude Mythos 5.1 [1]. Project Glasswing, which gave Mythos access to organisations securing critical software, is folded into the program [1][2].

Defense Access covers incident response, malware reverse-engineering and vulnerability analysis. Company and government security teams, open-source maintainers and individual researchers with a record of reported vulnerabilities can apply, and Anthropic aims to answer within a few days [1]. Red Team Access adds authorised penetration testing, is open to organisations only and takes a few weeks to review [1]. Specialized Access, for systems such as power grids and flight software, is reviewed with the US government [1].

Enrolled organisations must accept data retention so Anthropic can monitor misuse [1]. In Anthropic's own test, Opus 5.5 ran 10 multi-stage CyScenarioBench challenges five times in each tier: without program access every task was blocked on the first prompt, Defense Access blocked 46 of the 50 trials, and Red Team Access blocked none and completed 34 [1][2]. SiliconANGLE reports the program is available through the Claude Platform, Vertex AI and Microsoft Foundry, and on Bedrock only for customers eligible for Enterprise Frontier Safeguards [2].

Generally available models can still be used for code review, patching known issues and alert triage, Anthropic says [1].