Grok 4.7 Tops Coding Benchmark as Perplexity and Replit Ship New Agent Tools

Key Takeaways

1. Grok 4.7 sweeps the top three spots on VulcanBench Frontier v4 and lands on Microsoft Foundry.
2. Perplexity open-sources a multimodal decision model and a dual-size embedding model, cutting input token prices.
3. Replit adds cross-project context, letting new chats reference and edit existing projects.
4. SpaceX's Starlink reports over 11,000 satellites and 13 million customers as it pushes laser links toward interplanetary networking.
5. Meta and Sierra lead an open Personal Agent Protocol with Stripe, Shopify, Walmart and others.
6. Tesla's Model Y leads Korea's market for a third straight month as FSD Supervised gains German political backing.
7. Grok Bot automates email, calendar and task tools, with SpaceX routing tasks across multiple backend models.

1. AI Models and Infrastructure

1. Grok 4.7 took the top three spots on the VulcanBench Frontier v4 coding benchmark, with a winning score of 93.15 and all 23 tasks passed; the model is now available on Microsoft Foundry. A separate relayed claim says Grok is the only major AI near the political center in a bias test, while Anthropic, Google, Meta and OpenAI lean left — a claim marked disputed/unverified — via 1 2 3
2. Perplexity released the open-weight multimodal decision model pplx-decider-v1.1-27b, which scored highest on the new Hugging Face Decision Index 0.3, and halved input token pricing to $0.02 per million tokens. It also open-sourced pplx-embed-v2-late in 9B and 0.6B versions sharing one embedding space, supporting OCR-free PDF page retrieval, with 92.4% on MADQA and 64% on BrowseComp+ — via 1 2 3
3. Replit added cross-project context: a new chat can find existing projects, add features, or use one project as a reference for another, with Replit reading project files and starting edits in background tasks. Separately, Jev launched on Replit without an API key for building classification features or reselling APIs like sentiment analysis — via 1 2
4. Grok Bot is described as handling tasks across email, calendar, Notion, Linear and Todoist: scanning email roughly every two hours and drafting replies, writing decisions into tools after meetings, running monitoring tasks offline, and supporting @bot triggers for summaries, reminders and drafts. SpaceX will reportedly choose the best backend model per task, including Claude Opus 5.5, MidJourney and Suno, routing simple questions to small models and complex ones to large models — via 1 2 3 4 5 6 7 8 9

2. Space, Connectivity and Autonomy

1. Starlink now has over 11,000 satellites, about 60% of active LEO satellites, serving over 13 million customers across roughly 170 markets, with about 1,000 collision-avoidance maneuvers daily and thresholds nearly 100x more cautious than industry standards; the biggest risk cited is other operators not sharing maneuver data. Starlink is developing high-bandwidth lasers to connect the Moon and Mars in an "interplanetary mesh network," while StarMind AI plans to start launching satellites next year at about 250 kW per satellite and has filed for up to 1 million satellites — via 1 2
2. SpaceX's Starship completed its first commercial successful orbital insertion, with the company advancing full reusability and a 100+ ton platform. A Dragon capsule has undocked from the space station and will perform a deorbit burn and splash down off California in about 27.5 hours. Starlink Mobile Direct-to-Cell V2 satellites were approved by the FCC for up to 15,000 units, targeting peak speeds of 150 Mbps at roughly 326–335 km altitude — via 1 2 3 4 5
3. Tesla's Model Y was Korea's best-selling vehicle across all powertrains for a third straight month in September, with 12,372 Tesla registrations that month, up 36% year over year and an eighth consecutive month leading imported-car sales; about three-quarters of Norwegian Tesla owners would choose Tesla again. Germany's transport minister said in Berlin he would push for FSD Supervised to cover the EU quickly, and that Germany should explore its own approval path if EU talks fail — via 1 2 3

3. Agents, Protocols and Open Standards

1. Meta and Sierra Platform are jointly developing the open Personal Agent Protocol with Genesys, Instinct, RocketOTD, Shopify, Stripe and Walmart to define how personal agents interact with enterprises, open for anyone to implement. Shopify upgraded UCP to support rich media including video and 3D models and launched Shopify Catalog — via 1 2
2. Yohei Nakajima's project raised $7.7 million to eliminate subscriptions for agent tools, positioning itself as an "OpenRouter for agent tools" that lets agents discover, run and pay for tools at runtime through a single connection to 2,500 APIs, covering leads, SEO, marketing, search, e-commerce, stocks, video/image/music/3D generation, and agent email and phone, with pay-per-call pricing and no subscription — via 1
3. Perplexity's Computer can generate interactive 3D previews of designs, export editable CAD files, and be modified in FreeCAD with recorded edit videos; it can also schedule meetings via email based on a user's calendar and preferences. A relayed post describes Computer producing an animated scene from concept art to final cut using Blender, Nano Banana and Seedance 2.5 — via 1 2 3
4. AI Gateway added a beta feature where, if the primary model's confidence falls below a set threshold, the decision is rerun on a backup model, letting agents escalate uncertain decisions — via 1

4. Industry Moves and Market Signals

1. Intel's CEO confirmed continued cooperation with Musk on Terafab, with Intel participating since April as a development partner in chip design, manufacturing and packaging. Beam, an open agent model with 501B total parameters and 23B active, will release full weights this month. Grok Build v1.0.50 adds compact mode, timestamps, TSV/CSV/Markdown table export, worktree reliability, permission handling and CJK memory search, plus a breaking security change where grok worktree rm refuses by default to delete worktrees in use or with unsaved changes — via 1 2 3
2. General Medicine, TJ Parker's new company, raised a $120 million Series B led by a16z to build a "healthcare store" that makes users the most powerful participants in the healthcare system by letting them "shop." Parker previously founded PillPack, which Amazon acquired five years after a seed investment — via 1
3. Rhem Labs launched Rhem, a companion robot for parents that can talk with them, help make calls, book appointments, manage reminders and handle daily online tasks, with a family app showing health readings and routine changes and letting parents choose what to share; preorders are open — via 1
4. Replit's Amjad Masad said AI-driven reverse engineering and decompilation progress is "completely crazy" and predicted all software will soon be effectively open source. Naval argued models are software's last moat, expecting more software to move server-side to resist distillation, and relayed a claim that Photoshop was clean-room reimplemented via decompilation, spec conversion and LLM-generated Rust — via 1 2