OpenAI Ships GPT-6 Sol and Luna at Half the Price of GPT-5.6, as Grok Bot Hits 418K Weekly Users

Key Takeaways

1. OpenAI released GPT-6 Sol and GPT-6 Luna, priced 50% below GPT-5.6 promotional rates, with Sol already live for all Perplexity users.
2. Grok Build 1.0.41 brings Grok 4.7, while Grok Bot logged 53 performance fixes and 418K weekly active users, up 24% week over week.
3. Fitch assigned Tesla its first BBB investment-grade rating, citing EV leadership, $43.5B in cash and investments, and vertical integration across AI, Robotaxi, and Optimus.
4. Hugging Face ecosystem updates include vLLM's hardware-agnostic layer, GGUF support in transformers, and NVIDIA's open Nemotron 3 Diarization model.
5. Perplexity reported a post-training method that cut tool-call failure rates by about 21% in online A/B tests.
6. OpenAI outlined a policy to give third-party independent evaluators deep access across training, evaluation, and deployment.

1. AI Models and Pricing

1. OpenAI released GPT-6 Sol and GPT-6 Luna, built on GPT-6 Astra's technical progress and positioned as faster, cheaper models for scaled work; caching and inference efficiency improvements let OpenAI pass savings to users, with Sol and Luna API pricing 50% below GPT-5.6 promotional rates 1 2 3. Sam Altman said Sol and Luna show large gains over the 5.6 series in intelligence, alignment, work output, coding, and computer use, with per-token price halved and per-task billing even lower; he argued per-task billing is the key metric and that no competing product matches it 4 5. Greg Brockman framed the release as carrying Astra's SOTA progress in professional work, factuality, coding, computer use, and alignment into more affordable models 6.
2. GPT-6 Sol is now available to all Perplexity users, outperforming Opus 5 on the Wide-And-Deep-Research (WANDR) benchmark at one-fifth the price; Sol will serve as the orchestration model for Perplexity Computer's "Light" tier while Astra continues on "High," with Aravind Srinivas calling OpenAI's cadence a run of Pareto-optimal models 1. In a separate evaluation, Grok 4.7 scored 94% on a Next.js benchmark versus 97% for Opus 5.5, GPT 6 Sol, and Fable 5.1, but at 2–7x lower cost, which the author called quite good for a smaller model 2. A kilocode simulation showed Grok 4.7 costing $3.52 versus $7.35 for Opus 5.5 3.
3. On the tooling side, swyx said 5.5 Opus has become AINews' new default model, citing more concise, tasteful reporting and less "slopese" than 5 Opus after a side-by-side comparison with 6 Sol 1. Ben Tossell reported Opus 5.5 is live on Factory, with Medium as a strong default tier, 20–25% fewer output tokens than Anthropic Opus 5 at the same effort, and clear actionable answers on long investigation tasks 2. Deedy said Opus 5.5 produced a launch video for an inference startup in about one minute for roughly $2, versus weeks or months and about 1000x the cost through an agency, and argued video will reshape how products are marketed and explained 3.

2. Grok, X, and Platform Updates

1. SpaceXAI released Grok Build 1.0.41, bringing Grok 4.7 into Grok Build with improved crash recovery, prompt handling, navigation, configuration reliability, and sub-agent performance, plus a new sports_search tool for NFL live scores, standings, schedules, and player data, sub-agent model inheritance settings, per-model request size limits, and long-reasoning reminders 1. Grok Bot shipped 53 performance fixes over several days: reconnect time dropped from 60 seconds to 0.7 seconds, laptop wake from 23 seconds to 1 second, switching back to the Computer view from 868ms to 27ms, opening chats with long code blocks from 954ms to 228ms, and the Media tab stopped re-downloading (137MB to 0.7MB) 2. The app also added voice calls and voice memos, 1Password vault, inline forms, email and Slack inline drafts, account switching, and desktop traffic routing 3.
2. Grok Bot reached 418K weekly active users as of September 14, up 24% week over week, based on a SpaceXAI employee demo in London last week and described as an early demand signal for a key product; the author said usage growth exceeds anything they have seen before 1. X launched X Numbers, letting anyone be contacted without mutual follows or accepted requests, with users sharing their number themselves for optional messages or calls 2 3.

3. Infrastructure, Open Source, and Evaluation

1. Hugging Face ecosystem updates: vLLM is introducing a hardware-agnostic layer to preserve frontier performance while improving portability, work presented by IBM, Meta, and Hugging Face contributors in a new PyTorch Foundation blog post 1. Transformers can now run GGUF models directly, bringing ggml's Metal kernels into the transformers ecosystem for better compatibility and performance 2. The oMLX author joined Hugging Face full-time, saying the project stays in its original repository under Apache 2.0 with the author continuing to lead it, contributing upstream where sensible and experimenting in oMLX for missing community features, thanking 264 contributors 3.
2. NVIDIA open-sourced Nemotron 3 Diarization under a commercially friendly license for reliably tracking speakers in real-time conversation, handling overlapping speech and up to 8 speakers at 100M parameters, now on Hugging Face with day-zero Transformers integration; the poster said quality is good at one-second speech chunks and suited to voice agents, and demoed it with Pollen Robotics' Reachy Mini and a speech-to-speech pipeline on DGX Spark where the robot recognized a new voice, asked, and remembered the name 1 2 3. Sluicebox built supplier agent Lucy on NVIDIA Nemotron 3 Ultra to collect and verify supplier data for engineers assessing data-center carbon footprint choices, reporting improved accuracy, 51–80% lower cost, and up to 2x faster responses 4.
3. Perplexity published post-training research in which its Computer agent learns from real user sessions, imitating good trajectories and explicitly correcting avoidable errors such as wrong tool calls even when the overall trajectory succeeded; combining rejection sampling fine-tuning (RFT) with prompt-guided self-distillation cut tool-call failure rates by about 21% in online A/B tests 1. OpenAI said it will support third-party independent evaluators with deep access across training, evaluation, and deployment so they can challenge assumptions, find risks it may have missed, and reach independent conclusions on safeguard effectiveness, listing four priority areas for deeper evaluation alongside principles for rigorous, safe, independent work 2.

4. Business, Robotics, and Industry Views

1. Fitch assigned Tesla its first BBB investment-grade credit rating, citing Tesla's EV leadership, strong profitability, $43.5B in cash and investments, and "unparalleled" vertical integration as it expands into AI, Robotaxi, and Optimus 1. In a CCTV interview, Musk described Optimus as a personal robot more capable than C-3PO or R2-D2 that could care for the elderly, watch children, serve as a personalized tutor, and take on various jobs, predicting companies where one person manages hundreds or thousands of physical and digital robots, de facto universal high income, and uncertainty over whether money will still matter 2.
2. On AI and wealth inequality, Joe Lonsdale said on CNBC that the 2030s will be a "deflationary decade" in which everyone gets very rich if nothing goes wrong, citing the second industrial revolution from 1870 to 1900 when the average American working-class person's real wealth doubled within a generation; he expects AI to root out government fraud and waste, sharply cut healthcare costs, and reshore manufacturing, with productivity of 3–5% or higher solving the deficit, and called robotics at a "GPT-3 moment" while warning companies could accidentally cause hacking and cyber problems and should slow down and take responsibility for harm, with Joe Kernen questioning whether only the rich get richer 1. Musk shared a Jeffrey Katzenberg article on AI and creativity in which Katzenberg said his 2023 prediction that AI tools would cut world-class animation time and cost by up to 90% within three years, argued reasoning analyzes existing information while creation generates what does not exist, said AI is almost entirely on the reasoning side lacking empathy, commitment, serendipity, and human creativity though the gap may narrow, and urged creators be included through credit, consent, and compensation rather than overridden, with Hollywood accepting AI is not going away and deciding the terms of its existence 2.
3. Musk told a CS undergraduate asking where to have the most impact in the AI era that it may be at either extreme: close to the technology actually building LLMs, close to customers using AI to precisely meet their needs, or both 1. He said he admits he quite enjoys vibe coding and may have been a little wrong before 2, and asked what people would do with two hours a day returned by AI, 730 hours a year, whether spending time with children, exercising, or starting a company, as a conversation he wants to have about the tech future 3. He also agreed with a post saying company profits decline after a successful owner dies or retires 4.

5. Regional and Community Notes

1. Starlink Japan introduced easier sign-up, with a "rental Mini kit" at 0 yen upfront in some areas, fast installation across most of Japan, and Japanese-language customer support 1. Runway's AI Summit is one week away with partial agenda published, including a "Grounding Intelligence in the Physical World" session with Nvidia's Ming-Yu Liu, Google DeepMind's Jon Barron, and Wayve's Alex Toshev 2. Hugging Face is sponsoring its free Open Together community event on Friday, October 16 in San Francisco to kick off Open Source AI Week, with doors at 18:00, 36 community live demos plus food and networking from 18:00–21:00, a DJ dance party from 21:00–24:00, and collectible HF merch for the first 500 attendees 3.
2. Ethan Mollick built a hard sci-fi interstellar warship combat game with Fable/Opus including orbital mechanics, delta-v, heat, and realistic tactics simplified for 2D space, with the game handling complex math, and called it quite fun to play 1. He also questioned whether a major safety AI incident has ever actually occurred around a commercial frontier model release under normal deployment and normal safeguards, noting organizations were deeply worried in 2023 but that he is unaware of a real event 2. Hamel Husain advised that when "golden" eval datasets go stale, teams should find new problems through regular error analysis and keep updating eval sets as products and users change, recommending at least one /eval-audit on existing pipelines using the eval skill he created with @sh_reya, and shared views on the difficulty of scaling human attention and annotation to agent-generated unstructured information and on making formal methods explainable and scalable for human-agent software development 3 4 5 6.