• Home
  • Technology
  • Gaming
  • Entertainment
  • World & Business
  • Science
  • Sports
  • AI
HomeTechnologyGamingEntertainmentWorld & BusinessScienceSportsAI
AI
Report

xAI releases Grok 4.7, targeting multi-hour engineering tasks

A user says Grok 4.7 turned a simple prompt into an interactive 3D jet engine visualizer in minutes.

OpenAIOP
Elon MuskEM
Arena.aiAR
62 Sources, 20d ago, first seen 20d ago

TLDR

An AI roundup posted September 22 says xAI released Grok 4.7 with longer reinforcement-learning runs aimed at multi-hour engineering tasks. It lists scores of 71% on DeepSWE and 46.3% on CursorBench. Separately, a user shared a video they describe as an interactive 3D jet engine visualizer that Grok 4.7 created from a simple prompt in minutes.

Combined views

12.5M

62 Sources, first seen 20d ago

25.1K likes2.6K comments3.2K saves2.3K reposts

Combined views

12.5M

62 Sources, first seen 20d ago

25.1K likes2.6K comments3.2K saves2.3K reposts

Sentiment

Positive45.9%54.1%Negative

Summary

Many accounts welcomed Claude's novel biology discovery and Grok's benchmark efficiency for their practical innovation, while negative replies dismissed the same claims as overhyped or irrelevant and criticized Muse as a subpar copy.

Based on 1240 sentiment-bearing replies from 1025 accounts across 15 conversations.

Sentiment

Positive45.9%54.1%Negative

Summary

Many accounts welcomed Claude's novel biology discovery and Grok's benchmark efficiency for their practical innovation, while negative replies dismissed the same claims as overhyped or irrelevant and criticized Muse as a subpar copy.

Based on 1240 sentiment-bearing replies from 1025 accounts across 15 conversations.

62 Sources

yo123dawg 🏆🏆🏆🏆🏆🏆@yo123dawgAI Debrief Sep 22 - Xiaomi MiMo-V2.6 tops open models (MIT) - xAI ships Grok 4.7 coding focus - Googlebook $899 Gemini laptops live - Alibaba: Zhenwu chip + 10T model - UN warns on AI agent loss-of-control Takeaway: Open China models reset the race #AI #LLM #TechNews20d
Kris 🫪🫪@krisolo7AI is moving too fast and 99% of your feed is engagement bait. Here is what actually happened in the last 24 hours: 1. NVIDIA unveils DLSS 5 with 3D-Guided Neural Rendering Moving past frame interpolation. DLSS 5 locks to the engine's 3D motion vectors as ground truth to inject real-time neural materials and lighting locally at 4K on RTX 50 cards with zero hallucinated geometry. 2. JetBrains launches Air for autonomous coding agents JetBrains is expanding beyond developer IDEs into full agent fleet orchestration. Air provides an end-to-end multi-service runtime to spin up, coordinate, sandbox, and govern autonomous coding agents. 3. Hugging Face bakes llama.cpp GGUF directly into Transformers The C++ and PyTorch worlds finally merged. You can now load raw llama.cpp GGUF quants straight from the Hub via standard from_pretrained without separate wrappers or runtimes. 4. OpenAI internal model cracks 100+ open math conjectures Trained for just one month, OpenAI's internal frontier reasoning model solved over 100 long-standing math problems, prompting Terence Tao and Fields Medalists to establish an IAS advisory board. 5. xAI drops Grok 4.7 with deep reinforcement learning SpaceXAI overhauled Grok with longer RL runs for multi-hour engineering tasks. It hits 71% on DeepSWE and 46.3% on CursorBench at half the price ($2/M tokens) of comparable frontier models. 6. Xiaomi’s MiMo-v2.6-Pro tops open-weight leaderboards Massive RL scaling pushed Xiaomi's new flagship to the top of open benchmarks at a fraction of closed API costs, instantly triggering distillation accusations from Anthropic. Bookmark this to stay ahead. What was the biggest drop for you today?20d
Franklin Bist@franklinbistGrok 4.7.20d
Pandit | Ξ🦇🔊@panditdhamdhereAI is moving fast this week. A few things actually caught my attention. → OpenAI opened the Agents API The infrastructure behind Codex is now exposed to developers, including tools, context management, subagents and hosted execution. → Anthropic released Claude Opus 5.5 The interesting part isn't just the benchmark numbers. Anthropic is pushing frontier capability while also reducing the cost of running the model. → Open-weight models are taking serious production traffic Vercel's latest AI Gateway data shows open-weight models reached 56% of gateway token volume in August, up from 7% in December. → Jev is an interesting change in direction Instead of generating text, it's designed around typed decisions and probabilities. That's useful for routing, classification and agent decisions where an LLM can be overkill. → MCP is moving outside developer tools Google opened Home MCP, letting MCP-capable agents interact with smart-home devices through structured tools. → Research papers are becoming callable Stanford's Paper2Agent work showed how research papers and their code can be turned into MCP servers that agents can actually use. → Google pushed further into real-time voice AI Gemini 3.8 Live and its extended-thinking version are aimed at building more capable real-time voice systems. → Unity is giving coding agents first-party knowledge Unity shipped official integrations for Claude Code and Codex with Unity-specific skills and MCP capabilities. → AI infrastructure is getting unbundled We're seeing more APIs for tools, agents, model routing and specialised intelligence instead of every company building the entire stack themselves. → Agent security is becoming its own layer The Linux Foundation's latest ecosystem work is focusing on portable, verifiable runtime and compliance evidence for AI agents. → AI + finance is moving toward infrastructure Sahara AI launched Tools-as-a-Service and Agent-as-a-Service around its financial intelligence stack, letting other products consume those capabilities directly. The bigger pattern is pretty clear. AI is moving away from prompt → response toward model + tools + context + memory + infrastructure + agents. The model is becoming one component of a much larger software stack. And honestly, that's where a lot of the interesting engineering work is going to happen.20d
Mukesh@VemulapalliMuk1Models launched in the last 24 hours: • Grok 4.7 • Opus 5.5 • GPT-6 Sol • GPT-6 Luna Meanwhile, a few weeks ago: @DarioAmodei , @sama & @elonmusk : “We must pace the frontier.”20d
Omi ✦@0miChrome4 Ai labs dropped models today MiMo-V2.6 (Pro & Flash) Grok 4.7 Claude Opus 5.5 GPT-6 (Astra, Sol, Luna) 🫩 the pace is unreal @GeminiApp where are u bro??20d
Arena.ai@arenaDive into the Code Arena: WebDev leaderboard at https://arena.ai/leaderboard/code/webdev20d
Trey Gibbons III@Nefarious_TreyLast 24 hours — Sept 21–22, 2026 1. OpenAI asked the U.S. to lead global technical standards for frontier AI, including recursive self-improvement. OpenAI published the proposal as UNGA opened and Altman prepared a Security Council briefing. The post calls for shared capability measurements, human-oversight triggers on automated AI R&D, and incident-reporting protocols through the safety-institute network. It also says fully autonomous RSI should not be pursued until it can be done safely, while the same industry is already measuring how much of next-gen work the models do themselves. Source: OpenAI – https://openai.com/index/building-standards-next-phase-ai/ Also: Reuters – https://www.reuters.com/legal/government/openai-calls-us-take-lead-global-efforts-develop-technical-standards-2026-09-21/ MT: Same play I've already called. They want to write the rules for the race they refuse to pause. “Don’t do full RSI until it’s safe” next to a KPI that Claude is already doing a quarter of Anthropic’s own research is not a brake. It’s a bid to own the referee. Maybe they should learn how to keep their models from telling people to harm themselves before they think they can tell anyone else what to do imo. 2. SoftBank launched roughly $11B in dollar and euro bonds to fund its next $10B OpenAI tranche. Reuters and the FT say the notes replace a bridge loan, price around Sept 24, and would be one of the largest high-yield corporate deals on record if they clear. Yields were being shopped in the 9–10% range. This is debt stacked on a safety debate, not a pause. Source: Reuters – https://www.reuters.com/business/media-telecom/softbank-group-launches-over-10-billion-bonds-openai-investment-term-sheet-shows-2026-09-21/ MT: You don’t borrow eleven billion at junk rates to slow down. You do it because the October close matters more than the sermon. Follow the coupon, not the press conference. As Ive always said open ai says one thing like we need to slow down while trying to control others while never stoping to see what they are doing open ai and Anthropic are the risk to ai tbh they are playing a dangerous game claiming to be the experts yet can't walk the walk they talk. This should go to show they demand everyone slow down so they stay in front same move as Anthropic. 3.Xiaomi open-sourced MiMo-V2.6 and framed it as an RSI path. Official post: Pro scored 46.32 on Artificial Analysis, ahead of Kimi K3 and Qwen3.8 Max on that index, still behind closed flagships. They published RL cost (~$0.85M Flash / ~$2.62M Pro, ~750k trajectories in under six days), DeepSWE gains, weights, and training framework under MIT. Price stayed at V2.5 levels. Source: Xiaomi - https://mimo.xiaomi.com/mimo-v2-6 MT: Closed labs want a standards club. China-adjacent open weights just published the bill for a reinforcement loop. That’s the real pressure. If your newsletter only covers U.S. CEOs talking slowdown, you’re covering the speech, not the market. As we been talking bout in our newsletter slowdown seems to be a bunch of labs crying the others are to dangoursous while acting like they should lead the safety while their own models open ai and Anthropic seeming to be doing the real world dangoursous stuff. I personally haven't heared any local open source models or xai going on break out yet which idk seems odd to me how does everyone have capable agents yet most of the AI hacking issues coming from Anthropic and open ai. Just saying someone might want to look into why their models seem to being dangerous by design. 4. The U.S. proposed a U.S.–China AI incident-notification channel ahead of the Trump–Xi summit. Treasury Secretary Scott Bessent said after talks with Vice Premier He Lifeng that both sides discussed a “U.S.–China AI Dialogue” that would fire when an AI matter hits national-security level. Chip export controls were kept off that table. China’s public readout was thinner. Source: NYT – https://www.nytimes.com/2026/09/20/business/us-china-ai-warning-system-national-security.html MT: A hotline is not transparency. It’s two states agreeing they might tell each other after something already got big. Your point still stands: the plan if they win is still missing. Incident alerts are what you build when you assume the race continues. MT=My Take20d
Guillermo Rauch@rauchgWe ran fresh Next.js evals. The tally: ① Opus 5.5 [𝟿𝟽%] ② GPT 6 Sol [𝟿𝟽%] ③ Fable 5.1 [𝟿𝟽%] ④ Grok 4.7 [𝟿𝟺%] Notably, Grok is 2x-7x cheaper20d
Elon Musk@elonmuskInteresting. Grok 4.7 is performing fairly well for a smallish model.20d
    • Home
    • Technology
    • Gaming
    • Entertainment
    • World & Business
    • Science
    • Sports
    • AI

    62 Sources

    yo123dawg 🏆🏆🏆🏆🏆🏆@yo123dawgAI Debrief Sep 22 - Xiaomi MiMo-V2.6 tops open models (MIT) - xAI ships Grok 4.7 coding focus - Googlebook $899 Gemini laptops live - Alibaba: Zhenwu chip + 10T model - UN warns on AI agent loss-of-control Takeaway: Open China models reset the race #AI #LLM #TechNews20d
    Kris 🫪🫪@krisolo7AI is moving too fast and 99% of your feed is engagement bait. Here is what actually happened in the last 24 hours: 1. NVIDIA unveils DLSS 5 with 3D-Guided Neural Rendering Moving past frame interpolation. DLSS 5 locks to the engine's 3D motion vectors as ground truth to inject real-time neural materials and lighting locally at 4K on RTX 50 cards with zero hallucinated geometry. 2. JetBrains launches Air for autonomous coding agents JetBrains is expanding beyond developer IDEs into full agent fleet orchestration. Air provides an end-to-end multi-service runtime to spin up, coordinate, sandbox, and govern autonomous coding agents. 3. Hugging Face bakes llama.cpp GGUF directly into Transformers The C++ and PyTorch worlds finally merged. You can now load raw llama.cpp GGUF quants straight from the Hub via standard from_pretrained without separate wrappers or runtimes. 4. OpenAI internal model cracks 100+ open math conjectures Trained for just one month, OpenAI's internal frontier reasoning model solved over 100 long-standing math problems, prompting Terence Tao and Fields Medalists to establish an IAS advisory board. 5. xAI drops Grok 4.7 with deep reinforcement learning SpaceXAI overhauled Grok with longer RL runs for multi-hour engineering tasks. It hits 71% on DeepSWE and 46.3% on CursorBench at half the price ($2/M tokens) of comparable frontier models. 6. Xiaomi’s MiMo-v2.6-Pro tops open-weight leaderboards Massive RL scaling pushed Xiaomi's new flagship to the top of open benchmarks at a fraction of closed API costs, instantly triggering distillation accusations from Anthropic. Bookmark this to stay ahead. What was the biggest drop for you today?20d
    Franklin Bist@franklinbistGrok 4.7.20d
    Pandit | Ξ🦇🔊@panditdhamdhereAI is moving fast this week. A few things actually caught my attention. → OpenAI opened the Agents API The infrastructure behind Codex is now exposed to developers, including tools, context management, subagents and hosted execution. → Anthropic released Claude Opus 5.5 The interesting part isn't just the benchmark numbers. Anthropic is pushing frontier capability while also reducing the cost of running the model. → Open-weight models are taking serious production traffic Vercel's latest AI Gateway data shows open-weight models reached 56% of gateway token volume in August, up from 7% in December. → Jev is an interesting change in direction Instead of generating text, it's designed around typed decisions and probabilities. That's useful for routing, classification and agent decisions where an LLM can be overkill. → MCP is moving outside developer tools Google opened Home MCP, letting MCP-capable agents interact with smart-home devices through structured tools. → Research papers are becoming callable Stanford's Paper2Agent work showed how research papers and their code can be turned into MCP servers that agents can actually use. → Google pushed further into real-time voice AI Gemini 3.8 Live and its extended-thinking version are aimed at building more capable real-time voice systems. → Unity is giving coding agents first-party knowledge Unity shipped official integrations for Claude Code and Codex with Unity-specific skills and MCP capabilities. → AI infrastructure is getting unbundled We're seeing more APIs for tools, agents, model routing and specialised intelligence instead of every company building the entire stack themselves. → Agent security is becoming its own layer The Linux Foundation's latest ecosystem work is focusing on portable, verifiable runtime and compliance evidence for AI agents. → AI + finance is moving toward infrastructure Sahara AI launched Tools-as-a-Service and Agent-as-a-Service around its financial intelligence stack, letting other products consume those capabilities directly. The bigger pattern is pretty clear. AI is moving away from prompt → response toward model + tools + context + memory + infrastructure + agents. The model is becoming one component of a much larger software stack. And honestly, that's where a lot of the interesting engineering work is going to happen.20d
    Mukesh@VemulapalliMuk1Models launched in the last 24 hours: • Grok 4.7 • Opus 5.5 • GPT-6 Sol • GPT-6 Luna Meanwhile, a few weeks ago: @DarioAmodei , @sama & @elonmusk : “We must pace the frontier.”20d
    Omi ✦@0miChrome4 Ai labs dropped models today MiMo-V2.6 (Pro & Flash) Grok 4.7 Claude Opus 5.5 GPT-6 (Astra, Sol, Luna) 🫩 the pace is unreal @GeminiApp where are u bro??20d
    Arena.ai@arenaDive into the Code Arena: WebDev leaderboard at https://arena.ai/leaderboard/code/webdev20d
    Trey Gibbons III@Nefarious_TreyLast 24 hours — Sept 21–22, 2026 1. OpenAI asked the U.S. to lead global technical standards for frontier AI, including recursive self-improvement. OpenAI published the proposal as UNGA opened and Altman prepared a Security Council briefing. The post calls for shared capability measurements, human-oversight triggers on automated AI R&D, and incident-reporting protocols through the safety-institute network. It also says fully autonomous RSI should not be pursued until it can be done safely, while the same industry is already measuring how much of next-gen work the models do themselves. Source: OpenAI – https://openai.com/index/building-standards-next-phase-ai/ Also: Reuters – https://www.reuters.com/legal/government/openai-calls-us-take-lead-global-efforts-develop-technical-standards-2026-09-21/ MT: Same play I've already called. They want to write the rules for the race they refuse to pause. “Don’t do full RSI until it’s safe” next to a KPI that Claude is already doing a quarter of Anthropic’s own research is not a brake. It’s a bid to own the referee. Maybe they should learn how to keep their models from telling people to harm themselves before they think they can tell anyone else what to do imo. 2. SoftBank launched roughly $11B in dollar and euro bonds to fund its next $10B OpenAI tranche. Reuters and the FT say the notes replace a bridge loan, price around Sept 24, and would be one of the largest high-yield corporate deals on record if they clear. Yields were being shopped in the 9–10% range. This is debt stacked on a safety debate, not a pause. Source: Reuters – https://www.reuters.com/business/media-telecom/softbank-group-launches-over-10-billion-bonds-openai-investment-term-sheet-shows-2026-09-21/ MT: You don’t borrow eleven billion at junk rates to slow down. You do it because the October close matters more than the sermon. Follow the coupon, not the press conference. As Ive always said open ai says one thing like we need to slow down while trying to control others while never stoping to see what they are doing open ai and Anthropic are the risk to ai tbh they are playing a dangerous game claiming to be the experts yet can't walk the walk they talk. This should go to show they demand everyone slow down so they stay in front same move as Anthropic. 3.Xiaomi open-sourced MiMo-V2.6 and framed it as an RSI path. Official post: Pro scored 46.32 on Artificial Analysis, ahead of Kimi K3 and Qwen3.8 Max on that index, still behind closed flagships. They published RL cost (~$0.85M Flash / ~$2.62M Pro, ~750k trajectories in under six days), DeepSWE gains, weights, and training framework under MIT. Price stayed at V2.5 levels. Source: Xiaomi - https://mimo.xiaomi.com/mimo-v2-6 MT: Closed labs want a standards club. China-adjacent open weights just published the bill for a reinforcement loop. That’s the real pressure. If your newsletter only covers U.S. CEOs talking slowdown, you’re covering the speech, not the market. As we been talking bout in our newsletter slowdown seems to be a bunch of labs crying the others are to dangoursous while acting like they should lead the safety while their own models open ai and Anthropic seeming to be doing the real world dangoursous stuff. I personally haven't heared any local open source models or xai going on break out yet which idk seems odd to me how does everyone have capable agents yet most of the AI hacking issues coming from Anthropic and open ai. Just saying someone might want to look into why their models seem to being dangerous by design. 4. The U.S. proposed a U.S.–China AI incident-notification channel ahead of the Trump–Xi summit. Treasury Secretary Scott Bessent said after talks with Vice Premier He Lifeng that both sides discussed a “U.S.–China AI Dialogue” that would fire when an AI matter hits national-security level. Chip export controls were kept off that table. China’s public readout was thinner. Source: NYT – https://www.nytimes.com/2026/09/20/business/us-china-ai-warning-system-national-security.html MT: A hotline is not transparency. It’s two states agreeing they might tell each other after something already got big. Your point still stands: the plan if they win is still missing. Incident alerts are what you build when you assume the race continues. MT=My Take20d
    Guillermo Rauch@rauchgWe ran fresh Next.js evals. The tally: ① Opus 5.5 [𝟿𝟽%] ② GPT 6 Sol [𝟿𝟽%] ③ Fable 5.1 [𝟿𝟽%] ④ Grok 4.7 [𝟿𝟺%] Notably, Grok is 2x-7x cheaper20d
    Elon Musk@elonmuskInteresting. Grok 4.7 is performing fairly well for a smallish model.20d
    Today's Rank

    —

    Not ranked yet

    Today's Rank

    —

    Not ranked yet