← Dailies

Daily · 2026-09-23 00:00 UTC

OpenAI's GPT-6 Sol and Luna halve token prices; Anthropic ships Opus 5.5 — 2026-09-23 00:00

Ten stories: OpenAI and Anthropic both cut model prices on the same day, Xiaomi open-sources MiMo-V2.6, Singapore backs a 22-leader frontier AI safety statement, and Rabbit ships a hardware-free agent.

Source ↗

🎧 News Digest — 2026-09-23 00:00 UTC — 10 items (text digest this run)

1. OpenAI's GPT-6 Sol and Luna halve token prices — OpenAI added two cheaper models to its lineup: GPT-6 Sol at $2 per million input tokens and $10 per million output tokens, and Luna at $0.10 input and $0.50 output — a 50 percent cut against GPT-5.6 Sol and Luna. OpenAI attributes the drop to caching and inference improvements. Sol is aimed at recurring complex work such as building features, reviewing code, debugging and analysing data; Luna at high-volume, well-defined tasks. Cached input tokens now carry a 90 percent discount, alongside a new prompt-caching dashboard. Both ship in ChatGPT Work and Codex for Plus, Pro, Business, Enterprise and Edu subscribers, with Free and Go users getting Luna through the desktop app but neither model initially in regular chat; the API exposes them as gpt-6-sol and gpt-6-luna, and Terra has been retired. OpenAI says that on AutomationBench, which tests business workflows across 47 tools, Sol at its highest effort level beats Claude Opus 5 at maximum effort while costing 9 percent as much per task. The Decoder · Sep 22
2. Anthropic releases Claude Opus 5.5 at a lower running cost — Anthropic's newest model, released Tuesday, sets what the company calls a new state of the art in coding and knowledge work performance. Output tokens are charged at $20 per million, down from $25 for the previous model, and the model is faster to run. Opus 5.5 is less likely to use jargon and more likely to put important information at the start of its messages. The launch comes two months after Opus 5, which shipped on July 24, and Anthropic says the next tiers, Sonnet 5.5 and Haiku 5.5, will follow in the coming weeks. Anthropic describes Opus 5.5 as comparable to Mythos in its biology and cybersecurity capabilities, so it carries the same safeguards as the Fable model; alignment testing and pre-release evaluation were carried out with outside organisations including METR and Frontier Design. TechCrunch · Sep 22
3. Xiaomi open-sources the MiMo-V2.6 model family — The series includes two natively omnimodal models: a flagship MiMo-V2.6-Pro and a smaller, efficient Flash variant, plus a Pro-UltraSpeed variant that Xiaomi says delivers up to 20 times faster output than Pro at the same quality. Both accept text, image, video and audio within one family and carry 1 million-token context windows. Pro's published architecture lists a 681-million-parameter vision encoder, including a 308-million-parameter AudioTokenizer and a 127-million-parameter audio patch encoder capable of distinguishing speech. Xiaomi positions the larger models for agentic use cases, particularly computer use, where one model loop can read a task instruction and a UI screenshot, reason about them and decide what to do next. SiliconANGLE · Sep 22
4. Singapore backs a 22-leader frontier AI safety statement — Prime Minister Lawrence Wong said on Tuesday that Singapore supports a joint statement on AI safety initiated by Norwegian Prime Minister Jonas Gahr Støre and Finnish President Alexander Stubb. The letter, still open for endorsement, has been signed by Wong and 21 other world leaders, including Canadian Prime Minister Mark Carney and Australian Prime Minister Anthony Albanese. It calls on companies to develop transparent safety protocols, including mandatory pre-deployment testing and independent evaluation, with qualified evaluators granted sufficient access to assess risks, and urges greater cooperation on common standards and shared reporting of serious safety incidents. It also asks UN member states to explore creating an international institution that would set standards for frontier AI, enable verification and convene states when capability thresholds are crossed. CNA · Sep 22
5. Snorkel AI triples its valuation to $3.5 billion — The startup, which helps AI labs and corporations build training datasets and simulated environments, raised a $350 million Series E led by Insight Partners and S32 at a $3.5 billion valuation — nearly triple the $1.3 billion it was valued at when it raised $100 million in a Series D 17 months ago. Existing investors Addition, Lightspeed, Greylock, GV and Wells Fargo also participated. Snorkel says its annualised revenue run rate now stands at $375 million, an eighteenfold increase over 12 months, driven by demand for high-end training data; it shifted last year from data-labelling software to supplying completed datasets, an offering it calls data-as-a-service. TechCrunch · Sep 22
6. Rabbit launches a standalone AI agent that doesn't need its R1 hardware — Rabbit is rolling out OS3, an agent it says can operate locally across Windows, Mac and Linux devices. Users can add up to five devices to a single account along with their preferred AI models, and OS3 determines which devices, files, apps and models it needs to complete a task. It is reachable through a dedicated desktop site, a paired messaging app such as Telegram or iMessage, and Rabbit hardware. Founder Jesse Lyu told Wired that the startup has stopped manufacturing R1 devices but is not abandoning AI hardware — it is focusing on a new vibe-coding "cyberdeck" that will run OS3, which he says will launch within months. Rabbit claims OS3 will not store, copy, use or sell user data, though it keeps chats and memories derived from them on its servers. The Verge · Sep 22
7. Qualcomm's Snapdragon 8 Elite Gen 6 arrives with an Extreme version too — Qualcomm announced two flagship phone chips, with spec differences limited mostly to improved AI processing, video capture and gaming performance on the Extreme model. Both are built on a 2nm process and share their modems, display specs, security features, GPS, and memory and storage support; the CPU details given are identical, with two big cores at up to 5GHz and six performance cores at up to 4GHz. The 8 Elite Extreme can drive 8K at 60fps and 4K at 240fps capture and supports the Advanced Professional Video codec, and it gets a larger shared memory pool for the NPU plus AI-dedicated matrix cores in the GPU. Motorola is the first to confirm a device, with its Signature 27 flagship. The Verge · Sep 22
8. Big tech says AI will cure cancer — clinicians say where is it? — The Guardian spoke to oncologists and researchers about the gap between executive promises and clinical reality. Sherene Loi, a medical oncologist specialising in breast cancer at the Peter MacCallum Cancer Centre in Melbourne, said that until AI can think and run experiments at scale itself, she is not sure a cure for all cancer or medical illness arrives in the next 10 years. Emilia Javorsky, a physician and scientist at the Future of Life Institute, writes that 13 years into the AI drug discovery movement there is still no single FDA-approved drug that cleared the full bar of regulatory approval, reimbursement and clinical adoption. Real progress is visible elsewhere: radiologist Ajit Goenka of the Mayo Clinic has been using machine learning to detect pancreatic tumours in scans that were read as normal, and neuropathologist Felix Sahm of University Hospital Heidelberg leads EUcanAI, a European collaboration using agentic AI to speed up brain tumour treatment. The recurring constraint, researchers said, is not intelligence but data, regulation and incentives. The Guardian · Sep 22
9. A layer derived from lab-grown neurons promises 5x faster, 80% cheaper AI video — The Biological Computing Company wants to team up with AWS to sell a text-to-video model it says runs five times faster and about 80 percent cheaper thanks to a small software layer derived from real nerve cells. The layer adds less than 0.1 percent to the base model, but the startup will not say which model that is. The Decoder · Sep 22
10. British Columbia sues OpenAI and Sam Altman over the Tumbler Ridge school shooting — The province filed suit in San Francisco federal court on Monday, naming OpenAI and its chief executive as defendants and seeking damages to fund recovery efforts after the February attack, as well as an order directing changes to how the company handles what its products surface. The lawsuit alleges the attack could have been prevented if OpenAI had warned law enforcement that the shooter, Jesse Van Rootselaar, had used ChatGPT to plan it; she had been flagged by the company's safety team after conversations about gun violence. Nine people, primarily children, were killed. Altman published a letter in April saying he was deeply sorry that OpenAI had not contacted law enforcement and promising reforms, which the province says were never followed through. OpenAI spokesperson Drew Pusateri called the shooting an unspeakable tragedy and said the company remains committed to working with government and law enforcement officials. The suit follows separate California filings by more than 30 family members of victims. The Guardian · Sep 22

🧪 Try This Next

• 🔧 Compare GPT-6 Sol and Luna pricing against whatever you run today — the cache discount on cached input tokens is now 90 percent
• 🎥 Read the Guardian's AI-and-cancer investigation before repeating any "AI will cure cancer" claim — the regulatory and trial bottlenecks are the binding constraint, not model intelligence
• 📝 Write one page on what your agent is allowed to do without asking you — Rabbit's OS3 and the 22-leader safety statement are this week's two live precedents

Content on Anagnorisis is summarized, paraphrased, and editorialized from publicly available sources for length and clarity. Original sources are linked where available. All trademarks belong to their respective owners.