Grokbot Courts Rival Models While ChatGPT Goes Visual: AI's Loudest Seven Days

This week in AI boils down to a single sentence: the best assistant now wants to run on the best models. Elon Musk posted on X that Grokbot will henceforth use whichever backend model is most likely to deliver the best outcome for each task, naming Anthropic's Claude Opus 5.5 alongside Midjourney and Suno. According to Thenextweb's October 7 report, Musk did not say which jobs go where, but the principle of multi-model routing is clear: Midjourney for images, Suno for songs, Claude for demanding text work.

The move targets the old trade-off between assistants and models. As the host frames it, Grokbot and Muse are delightful to use but run on comparatively weaker models, while Dots runs on a top-tier model yet still feels rough. Users previously had to pick the best app or the best model; Musk now promises both at once. The timing is pointed too, since Thenextweb notes that a day earlier Grok users hit access problems on mobile and web, as reported by City AM.

The most comfortable assistant meets the strongest models

Grokbot brought more news besides. A starred primary bot in every account embodies the proactive assistant idea, hunting for work instead of waiting for orders; the host describes it flagging an offer in his inbox hours before it expired. Yet he reckons only about a quarter of such nudges prove genuinely useful, with the rest adding to notification noise. Dots and Muse are heading the same way, but none of them has nailed proactivity yet.

Grokbot also gained built-in monitoring of X, able to read, search and track what trends there with no extra setup, where a paid developer interface used to be required. That platform-native watchtower is the assistant's first serious window onto the outside world. Open questions remain over cost and user control: routine writing may be routed to Grok while Opus fans grit their teeth, and nobody knows how fast premium models will burn through usage allowances. This reporting comes from the Thenextweb source.

The week's most concrete OpenAI news was GPT-6 reaching everyone on ChatGPT. The company's October 7 announcement extends the new generation of intelligence to more than 1.2 billion weekly users, with GPT-6 Sol serving paid plans and GPT-6 Luna serving free and Go users. Arriving with it, the intelligent UI layout produces visual, subject-specific arrangements instead of a wall of text. This information comes from the OpenAI source.

The chat window turns into an app

The host's experiments show the new layout stretching from tutoring to daily chores. Asked about Tabata workouts it answers with a table and explainer, then builds a countdown workout timer; asked about handstands it draws a four-step roadmap with web-sourced images matched to each stage. An IKEA product link first confuses assembly, then yields a lucid diagram on request; a meditation guide arrives in a paint-program aesthetic with a daily tracker attached. A branching horror story continues with fresh choices each time the reader locks the door. This information comes from the OpenAI source.

The Dots front, by contrast, is still maturing. Planned voice fixes include faster, more reliable call connections, written records and summaries of conversations, and variety for the assistant's endlessly repeated checking patter. On top of that, the engineer leading Codex pledged that for 28 days, each day will bring either a concrete ChatGPT improvement or a full usage reset for everyone. According to Nerdschalk's roundup, the pledge was posted on October 4 and GPT-6 Astra plus 6.1 Sol already run about 50 percent faster, though no reset has been confirmed for any individual account.

Anthropic stamped the week with its small, fast model: Claude Haiku 5.5, by its own account the cheapest, fastest and most capable small model it has ever shipped. Its October 7 announcement puts average running costs around 75 percent below the previous Haiku generation. The measurements are punchy: 1620 points on the knowledge-work evaluation against GPT-6 Luna's 1437, 72.4 percent on the computer-use test versus 15.7 percent before, and 39.2 percent on terminal coding, far ahead of rivals. This information comes from the Anthropic source.

Anthropic was generous on pricing too: cache-read fees for Sonnet 5.5 halved, saving roughly 20 percent on most agentic work, and Max and Team subscribers gained a monthly API credit for building agents. The host says he runs Opus 5.5 non-stop on the 100-dollar plan without hitting limits, and that Sonnet 5.5 comfortably covers a week on the 20-dollar plan. The family's biggest member, Fable 5.5, remains unannounced; the host expects it to shake things up, but that is unverified. This information comes from the Anthropic source.

Mandatory cloud and AI inside the office suite

The week's most contentious call came from Anthropic: as of October 6, new Cowork tasks on Pro and Max face a cloud-only mandate , running only in the cloud, the on-this-computer-only option is gone, and there is no way back. Two grievances dominate: many users, the host included, say they never received the email or in-app notice; and Cowork's strongest selling line, that everything stays on your machine, has been dented. According to Hackingdemand's October 5 review, existing local sessions survive until finished, but reaching your files requires the desktop app to stay open. This reporting comes from the Hackingdemand source.

The host concedes the cloud move plainly improves mobile life, with every task now watchable and steerable from a phone. The logic question stands anyway: if files upload and sync through the on-machine app, the computer must be on regardless, so why not run locally? Those insisting on local execution are pointed at Claude's developer-focused command tool. Meanwhile Claude is moving into Google's office apps, with an official add-on for Docs, Sheets and Slides in open beta across all paid plans. This information comes from the Claude source. The second half of the same picture sits in office software: the add-on docks a sidebar beside the file, reads the open document, the selected cells or slides, and edits directly. In Docs it fixes sentences and restyles headings in place, proposing bigger rewrites as approval cards. In Sheets it writes formulas, builds pivot tables and native charts, and drops into Python for joins and cleaning before writing results back. In Slides it generates new slides from the deck's own themes and flags overlapping or unreadable elements. The default ask-before-edits mode ties every step to an approval card. This information comes from the Claude source.

Google quietly opened an important door in Drive: Markdown files now preview, edit and collaborate natively inside Drive and Docs, where they used to be view-only or forcibly converted into Docs. As the host stresses, nearly every AI tool writes Markdown, so making it Drive's native tongue unlocks workflows where agents are pointed straight at Drive folders. This reporting comes from the Thenextweb source.

Free Gemini narrows while games and accessibility widen

Google's sour news sits in pricing: from October 9, plan-less Gemini users get only the Flash-Lite model, losing Flash and Pro. According to 9to5google's October 3 report, 4.99-dollar Plus subscribers keep Lite and Flash but lose Pro, while Pro and Ultra plans stay whole and Pro gains the maximum parallel-reasoning option. October also brings low, medium and high effort levels per model; high effort answers more thoroughly but eats more allowance. This reporting comes from the 9to5google source.

The sweet news comes from games and visuals. Playground, an experimental platform, promises game creation from written descriptions with no software knowledge required: two- and three-dimensional options, starter templates and single-player ready now, multiplayer on the way. The host's amusement-park tycoon attempt started rough but suggested a decent game is reachable with effort. On visuals, the Nano Banana Pro wait continues; the host reckons the aging model still beats GPT Images 2.5 most of the time. This information comes from the Google blog source.

Verification and accessibility brought two welcome steps. SynthID Detector opened to everyone on October 7: upload an image, audio clip, video or text and the portal scans for the invisible watermark , recognizing content from Google, OpenAI and Kakao tools, with Apple support coming soon. According to Gigazine's October 8 trial, more than 180 billion images and videos have been watermarked since 2023. Separately, Guided Vision inside Gemini Live offers blind and low-vision users live spoken descriptions plus natural reframing cues through the phone camera; the Google blog says it trained on tens of thousands of hours of visual-interpretation data with Aira and was tested by over a thousand volunteers. This information comes from the Gigazine and Google blog sources.