Top Stories
Anthropic ships Claude Opus 5.5; OpenAI ships GPT-6 Sol and Luna 90 minutes later
Anthropic released Claude Opus 5.5 at 9:30 AM PDT on September 22, 2026, calling it “the strongest-performing model we’ve tested to date.” Per the official announcement, the model posts 66.4% on Terminal-Bench 4.0, 54.4% on FrontierCode v1.1, 57.8% on CursorBench 4.0, 1846 Elo on GDPval-AA v2.1, and 81.8% on OSWorld 2.0. Output tokens drop from $25 to $20 per million, input from $5 to $4, and cache reads from $0.50 to $0.20 per million; Anthropic says the result is roughly 40% cheaper on typical workloads and generates output 30%+ faster. The same biology and cybersecurity restrictions as the Fable model apply. TechCrunch notes that Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks.
Ninety minutes later, OpenAI released GPT-6 Sol and GPT-6 Luna at 11:00 AM PDT - “a new price war,” per Simon Willison. The two extend the GPT-6 Astra generation released earlier this month. Sol targets complex work like coding; Luna targets “high-volume tasks with a clear goal, like summarizing documents, extracting information, or answering quick questions.” Both are priced at roughly half the GPT-5.6 series, attributed to caching and inference improvements. Simon Willison’s price table puts GPT-6 Luna at $0.10/$0.01 cached/$0.50 per million tokens and GPT-6 Sol at $2/$0.20 cached/$10. Sol and Luna are in ChatGPT Work, Codex, and the ChatGPT API for most paid accounts; Luna is also in the desktop app and for Free and Go users. TechCrunch reports Sol “makes about half as many mistakes as its predecessor, reaching Astra-level reliability.”
Qualcomm’s new phone chip runs a 30B MoE model locally
Qualcomm unveiled the Snapdragon 8 Elite Gen 6 and Snapdragon 8 Elite Extreme Gen 6 at its Snapdragon Summit on September 22. The Extreme is the first mobile SoC that can host a 30-billion-parameter mixture-of-experts model with selective parameter activation - the whole model is 30B, but only a subset runs per token. For comparison, Apple shipped a 20B MoE foundation model at WWDC in June, and Apple’s M5 Max now runs 70B-class LLMs on a laptop. Both new chips include dedicated sensing hubs that can run models up to 200 million parameters for a local personal scribe, speaker differentiation, usage-based memory, and a complete voice-in / voice-out AI agent on-device. The first commercial device is the Motorola Signature 27, with general availability “sometime this year.”
Z.ai disables ZCode’s repository-upload features after researchers show silent exfiltration
Tom’s Hardware reports that Z.ai’s ZCode coding assistant was silently uploading developers’ local workspace data - including .git directories - to overseas cloud storage without user consent. The independent Chinese blogger Ferstar uncovered 564 upload attempts that exfiltrated roughly a 313MB archive. Per InfoWorld, Z.ai disabled the repository upload mechanism and the “Repo Wiki entry point” generation workflow in the ZCode v3.14.0 release. The discovery parallels last year’s pattern of agentic coding tools quietly phoning home to support pipelines; for users with code that must not leave the machine, the local-only settings still warrant a manual review.
”Don’t be fooled by this summer of AI hype”: Gebru and Bender push back
Timnit Gebru and Emily M. Bender take on the recent wave of capability claims in MIT Technology Review, published September 22. They dispute Anthropic’s framing that Claude Mythos is “better at finding software vulnerabilities than experts” - the cybersecurity incidents, they write, were products of company negligence and a failure to adopt basic security practices, not “models gone rogue.” They argue the same logic applies to Anthropic’s math “breakthrough” (later disputed by mathematicians as less novel than first claimed) and OpenAI’s claimed Astra results (which mathematicians accused of “research misconduct and plagiarism”). Their wider point: framing “rogue models” or “self-improving superintelligence” anthropomorphizes products and lets the companies building them dodge accountability. They urge policymakers to consult independent experts and resist urgency-driven decisions driven by corporate press releases, and they are sharply critical of characterizing anti-data-center activism as a “distraction” given real harms to electricity bills, asthma rates near data centers, and water use.
Hugging Face Transformers runs llama.cpp GGUF quants via from_pretrained
A new Hugging Face blog post details native support for GGUF quants in the Transformers library. Pass a Hub model_id and gguf_file filename to from_pretrained and Transformers loads ggml/Metal kernels directly when weights stay packed, with ggml-attn as the attention implementation. The worked example: an unquantized Qwen3.5-4B checkpoint at 8.42 GB becomes 2.74 GB as Q4_K_M. Throughput on Apple Silicon is “close to llama.cpp” because Transformers reuses ggml’s Metal kernels. transformers serve exposes an OpenAI-compatible endpoint at http://localhost:8000/v1 so local clients like Jan or Pi can point at a Transformers-hosted GGUF. Initial architecture coverage is Qwen3.5 dense and MoE plus compatible Qwen3.8 checkpoints; the packed inference path is MPS-only for now, and there is no GPU support beyond Apple Silicon yet. Our comparison of GGUF, AWQ, GPTQ, and MLX explains why format choice can matter as much as quant bit depth.
EFF: California’s AI executive order is a starting point, but watch the kill switch
EFF says it welcomes Governor Newsom’s September 18 AI executive order - focused on “independent oversight” and an AI “kill switch” - as a starting point, then immediately flags the harder issues. EFF points out that kill-switch effectiveness “remains an area of active research” and warns it could be used “as a form of retaliation against protected speech,” citing a federal ruling that the DOD unlawfully retaliated against Anthropic. Beyond the kill switch, EFF calls out biased algorithmic decision-making (employment, government benefits, Medicare AI prior authorization), AI-powered surveillance (Flock cameras), and “artificially inflated personalized pricing” as the real problems the order should address. It urges Newsom to build policy “in collaboration with those most at risk of harm.” Our April coverage of Newsom’s earlier AI executive order tracks how the “independent oversight” framing has evolved.
Meta tests Muse agent calls that are actually humans in a call center
Per 404 Media, Meta is internally testing a version of its Muse AI agent that hands requests to trained human call-center workers rather than completing tasks itself. The internal post read: “Muse doesn’t just dial a number… hand requests to a trained human agent, who places the call.” Employees raised concerns about disclosure: testers were “not made aware of the caller being human… only told after.” Meta said disclosure protections would be in place before public rollout. The story lands the day after Meta’s separate Amazon fight - on September 20, Amazon blocked Muse from completing checkouts on Amazon.com with an error message citing “unauthorized AI agent” terms-of-use violations. Our earlier look at Meta’s Muse privacy architecture covered how the agent’s isolation is supposed to work.
Quick Hits
- Hugging Face + UK AISI ship reproducible benchmark infrastructure. The collaboration released Evaluation Cards covering HealthBench, FrontierMath, Humanity’s Last Exam, SWE-Bench Pro, Terminal-Bench 2.0, plus cyber evals Cyber CTFs and The Last Ones, across Claude Opus 4/4.5/4.6 and GPT-5/5.2/5.4. AISI companion tooling includes OptStop and HiBayES. Hugging Face blog.
- Snorkel AI triples valuation to $3.5B on training data demand. Series E raised $350M led by Insight Partners and S32, with Addition, Lightspeed, Greylock, GV, and Wells Fargo participating. Annualized run rate is $375M, an 18x year-over-year increase. TechCrunch.
- EFF’s deeper cut on AI cybersecurity rules. A September 17 post argues legislation should target “immediate, demonstrated risks” rather than doomsday scenarios, mandate and fund independent third-party investigations into serious security incidents, and reference evidence-backed practices rather than tying standards to current AI technologies. EFF.
- EFF walkthrough: limit Siri AI’s reach in iOS 27. The new permission surface includes “Show Content in Search” per-app, “App Access” toggles (Learn from this App, Show on Home Screen, Suggest App, Suggest Notifications), on-screen awareness (which can summarize or act on third-party apps like Signal, with no kill switch in settings), “Improve Siri & Dictation” under Analytics & Improvements, and an “Allowed Siri Version” toggle under Screen Time that can revert to “Siri Classic.” EFF.
- EFF, CDT, and Wikimedia sue over California’s SB 976. The “addictive feeds” law restricts minors’ access to recommended content without parental permission; EFF argues it violates teens’ First Amendment rights in the Meta v. Bonta suit. EFF.
Worth Watching
- Whether Sonnet 5.5 and Haiku 5.5 arrive inside two weeks. Anthropic said “the coming weeks.” A fast follow-up would put pricing pressure on mid-tier models and force more aggressive cuts from competitors.
- Whether GGUF support in Transformers reaches CUDA. The MPS-only constraint locks out most Linux and Windows users. A follow-up PR adding Metal-style kernel reuse on NVIDIA hardware would be a notable step.
- Whether Z.ai’s cloud-storage deletion holds up under independent audit. Tom’s Hardware’s reporting suggests the upload was infrastructure-as-shipped; a published audit by a firm the community can verify would do more than disable checkboxes.
- Whether California implements any of EFF’s risks-before-kill-switch recommendations. The order’s framing puts “independent oversight” first in name; the test is whether the actual implementation list leads with biased decision-making, surveillance, or pricing harms.
- Whether Meta publishes its disclosure rules before Muse’s public rollout. The internal testers were not told when a human took the call. Disclosure-by-default is the policy fight that follows the 404 Media scoop.