OpenAI Ships GPT-6 Sol and Luna at Half the Price as Grok 4.7 and MiMo-V2.6 Reshape the Model Race

Key Takeaways

1. OpenAI released GPT-6 Sol and GPT-6 Luna, priced 50% below GPT-5.6 promotional rates, with Altman claiming broad gains in intelligence, alignment, coding, and computer use.
2. Grok 4.7 landed in Grok Build with a 94% Next.js score at 2-7x lower cost, while Grok Bot hit 418K weekly actives and shipped 53 performance fixes.
3. Xiaomi's MiMo-V2.6 open-weight models topped Artificial Analysis's open-weight intelligence index, and Perplexity reported a 21% drop in tool-call failures from its post-training method.
4. The Horowitz Andreessen Academy launched in San Francisco with $42M led by a16z, 10 founding partners, and a tuition-free one-year fellowship.
5. Meta's Muse added Expedia and Instacart integrations, while X rolled out X Numbers for contact without mutual follows.
6. Fitch gave Tesla its first BBB investment-grade rating, and SpaceXAI's Grok Bot data plus Starlink Japan's zero-yen Mini kit rental signaled continued infrastructure expansion.

1. AI Models and Infrastructure

1. OpenAI released GPT-6 Sol and GPT-6 Luna, built on GPT-6 Astra progress and positioned as faster, cheaper models for scaled work; caching and inference efficiency improved, and API prices are 50% below GPT-5.6 promotional rates. 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 prices halved and per-task billing even lower; he argued per-task billing is the key metric and that no competing product matches it. Altman added that OpenAI wants the best model at every price point and in every modality, and that only with top capability models can the company drive change elsewhere. 1 2 3 4 5
2. Grok 4.7 is now available in Grok Build via the 1.0.41 update, which improved crash recovery, prompt handling, navigation, configuration reliability, and sub-agent performance, and added a sports_search tool, sub-agent model inheritance, per-model request size limits, and long-reasoning reminders. In Next.js evaluations, Grok 4.7 scored 94%, below Opus 5.5, GPT 6 Sol, and Fable 5.1 at 97%, but at 2-7x lower cost; in a kilocode simulation, Grok 4.7 cost $3.52 versus Opus 5.5's $7.35. 1 2 3 4
3. Grok Bot shipped 53 performance fixes in days, cutting reconnect time from 60 seconds to 0.7 seconds, laptop wake from 23 seconds to 1 second, Computer view switching from 868ms to 27ms, and long-code chat opening from 954ms to 228ms; Media tab downloads dropped from 137MB to 0.7MB. The app added voice calls and memos, 1Password vault, inline forms, email and Slack inline drafts, account switching, and desktop traffic routing, and reached 418K weekly actives as of September 14, up 24% week-over-week, a figure described as an early demand signal and faster growth than previously seen. 1 2 3
4. Xiaomi's MiMo-V2.6 open-weight series, including Pro and Flash, advanced via scaled reinforcement learning; Pro matched Claude Opus 5 and GPT-5.6 Sol on most agent benchmarks and scored 46 on the Artificial Analysis intelligence index, the highest for an open model, with weights, technical report, RL environments, and training code released. Aravind Srinivas responded "never bet against open-weight models losing," and MiMo-V2.6-Pro debuted on Wally at 320 tok/s, 2.4x the model owner's own API, topping the open-weight index within 24 hours. 1 2
5. Perplexity published post-training research letting Computer agents learn from real user sessions by imitating good trajectories and explicitly correcting avoidable errors such as wrong tool calls even in successful runs; combining rejection sampling fine-tuning with prompt-guided self-distillation cut tool-call failure rates by about 21% in online A/B tests. GPT-6 Sol opened to all Perplexity users, outperformed Opus 5 on the WANDR evaluation at one-fifth the price, and will serve as the orchestration model for Computer's "Light" tier while Astra handles "High." 1 2

2. Startups, Products, and Platforms

1. The Horowitz Andreessen Academy launched in San Francisco for high-potential high school graduates, aiming to be an elite talent magnet that helps students adapt to the future rather than the past. Its primary goal is teaching students to build, the most important skill in the AI era, mainly through individual and group projects plus courses and guest lectures from Silicon Valley builders. The network includes 10 founding partners—Anduril, Anthropic, Coinbase, Google, Meta, NVIDIA, OpenAI, Palantir, Replit, and Stripe—plus 50+ hiring partners and 200+ speakers and mentors; it raised $42M led by a16z, with Jack as CEO and Marc Andreessen and Erik Torenberg joining the board. The founding Fellowship is a tuition-free one-year program, with a two-year paid program planned after regulatory approval at a cost near elite private universities. 1 2
2. Meta's Muse added Expedia, letting users tell their personal AI agent a destination and have it coordinate hotel and itinerary arrangements, and Instacart, where connecting the service and saying "Taco Tuesday" generates a cart from a usual store, checks out, and delivers while preserving the existing Instacart experience. X launched X Numbers, allowing anyone to contact another person without mutual follows or accepted requests, with users sharing their number for optional messages or calls. 1 2 3 4
3. Replit became a founding partner of the Horowitz Andreessen Academy, with Amjad Masad joining its board and endorsing Gagan's team's education philosophy that learning should focus less on memorizing content and more on what students can build. Replit also launched "RevOps on Replit," hosted by RevOps lead Quoc Tran, showing tools RevOps teams actually use and live-building against real revenue problems, streaming Thursdays at 9pm PT on YouTube, LinkedIn, and X. 1 2 3
4. Basecamp 5 added a background wallpaper feature accessible in project settings by clicking the project name, supporting full-screen or traditional tiled backgrounds; Jason Fried said setting a background makes it immediately clear which project you are in and that the feature has significant customization room. He also endorsed an automation case linking Basecamp and Cloudflare R2, where a bot stores files received in Basecamp into R2, assigns IDs in both systems, creates a triage card in Basecamp, and after light editing or approval turns files into public-facing web pages before marking the card DONE. 1 2
5. Eight Sleep released the sixth-generation Pod 6 smart sleep system, calling it its most important hardware launch and most accessible generation; the new Hub can sit under the bed frame, is more powerful, has 9x more sensors for accuracy, adds a Solo size, and starts at the lowest price of any generation, with sales already open. 1 2

3. Business, Markets, and Society

1. Fitch gave 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. Separately, Elon Musk said in a CCTV interview that he envisions everyone owning a personal robot more efficient than C-3PO or R2-D2 that can care for the elderly, watch children, serve as a personalized tutor, and take on various jobs; he predicted companies where one person manages hundreds or thousands of physical and digital robots, a de facto universal high income, and uncertainty about whether money will still matter. 1 2
2. Starlink Japan introduced easier installation, with some areas offering a zero-yen initial cost "Mini kit rental" for households, fast installation across most of Japan, and Japanese-language customer support. SpaceXAI's Grok Bot weekly actives reached 418K, up 24% week-over-week, a figure presented at a London demo by SpaceXAI staff and treated as an early demand signal for a key product. 1 2
3. 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 predicted AI could eradicate government fraud and waste, sharply cut healthcare costs, and reshore manufacturing, and that productivity growth of 3-5% or more could solve the deficit; he called robotics at a "GPT-3 moment," said he is not worried about near-term existential risk, but warned companies could accidentally cause hacking and cyber problems and should slow down and take responsibility for damage. Joe Kernen questioned whether only the rich would get richer. 1
4. Reid Hoffman agreed that data center siting should be decided by communities that will live with the data centers long-term, including whether to build and under what conditions, arguing the decision belongs only to communities that bear the costs and share the benefits. The cited post said the governors of Virginia and New Jersey are quietly steering Democratic data center policy back toward rationality: no bans, community decision-making, no NDAs or subsidies, and priority for local construction jobs. 1
5. Andrew Wilkinson highlighted Eight Sleep's Pod 6 as the company's most important hardware launch and most accessible generation, with a new under-bed Hub, 9x more sensors, a new Solo size, and the lowest starting price of any generation, already on sale. 1
6. Levelsio said his income plus investment returns now exceed $10M/year at roughly 94.5% margin, with business revenue steady at $200K-$250K/month and investment returns exceeding business income, mainly in Vanguard S&P 500 ETFs plus individual stocks and startup investments including Nvidia, Replicate, FAL, and Cursor; most gains are unrealized and he never sells, while warning that investments are volatile. He also said he bought $72,450 of AMD stock, arguing that if George Hotz makes AMD GPUs compatible with Nvidia code, AMD could become the first company to challenge Nvidia's CUDA monopoly. 1 2