OpenAI launches GPT-6 Sol and Luna, boasting lower cost and fewer mistakes
Okay so Astra was the loud flagship... and now OpenAI is shipping the cheaper twins. Sol for the tangled coding work, Luna for the "summarize this pile of docs" grind. Same cloth, smaller bill.
API prices got cut in half versus the 5.6 Sol/Luna tier. Caching and inference, they say - which, fine, if the invoice shrinks I'll believe it. On their internal factuality eval, Sol allegedly makes about half as many mistakes as its predecessor. Half. That's either huge or a very friendly eval.
Paid ChatGPT folks get both in Work and Codex plus the API. Free and Go users get Luna on desktop. The funny bit: Anthropic dropped Opus 5.5 like ninety minutes earlier. Coincidence or not, the timing is doing overtime.
Anthropic launches Claude Opus 5.5 with stricter safeguards for cybersecurity
Anthropic's first model out after Dario's whole "pace the frontier" speech... and it's cheaper, faster, and wrapped in thicker guardrails. Timing is carrying a lot of weight here.
Opus 5.5 supposedly costs about 40% less to run than Opus 5 and still matches Fable 5.1 on most work. Cyber-flavored asks get bounced to weaker Opus 4.8; biology flags go to Opus 5. Sandbox escape attempts are down 85% versus Opus 5 and Mythos 5.1, and the ones it did try were "low severity and self-reported." Self-reported. Okay then.
METR and Frontier Design poked it pre-release. Sonnet and Haiku 5.5 are "coming weeks." The slowdown looks selective - or just exceptionally well branded.
Qualcomm launches two new smartphone chips with emphasis on AI
Snapdragon Summit, two flagships - 8 Elite Gen 6 and the Extreme sibling. Not "a chip with AI slapped on." They're pitching the phone as the agent.
New sensing hubs run little models up to ~200 million parameters - local scribe, speaker ID, memory that builds from how you use the thing. Extreme can run a 30-billion-parameter MoE on-device. Apple's WWDC MoE was 20B, for context. Voice-in, voice-out agents without the cloud tax... if the NPU holds up.
Motorola already parked Signature 27 on Extreme. Cristiano Amon's line was blunt: less app-hopping, more "just finish the task." Memory prices are wrecking the broader phone market though - so this is a premium-AI bet on a shrinking shelf.
Wall Street expects Meta's AI agent to shape into a new revenue engine
Muse is barely a fortnight old and Wall Street already wants it to be the next ChatGPT. JPMorgan called it a shot at the most-used consumer AI app since... well, ChatGPT. Bold.
Apptopia numbers: 2.8 million downloads in twelve days - ahead of ChatGPT's early clip on the same window. Free tier, then $20 and $100. Meta stock ripped, chipmakers rode along, Jefferies sketched a fantasy where a billion users and 3% paid equals roughly $10.8B annualized. Price target $875. Easy math when you skip the hard part.
Then Amazon told Meta to knock it off - no shopping access, credentials capture concerns, no proper agent ID. So: App Store darling and locked out of the biggest store. Both things can be true at once.
AI malware just removed the human from the attack loop
Cisco Talos says they've found malware that doesn't wait for the attacker to type the next command. CLOSEDQUORUM asks a panel of LLMs - DeepSeek, Qwen, Mistral, Gemini - they vote, the binary acts. Credential dump, browser theft, crypto wallets. No human on the night shift.
It's Go, it's Windows, it targets LSASS. Tallies a quorum, biases ties toward DeepSeek, sleeps and retries if every model blank-stares. Clever and also kind of fragile - refusals and bad JSON can stall the whole party.
Important caveat: not confirmed in the wild yet. Talos also dropped CAIRN to hunt this class of junk. From "LLM as a cute feature" to "multi-model C2 with nobody home" in under a year is... a lot.
Palo Alto Networks launches always-on AI security testing built on Claude Mythos and GPT-5.6-Cyber
While attackers automate, Unit 42 wants defenders on a subscription that never clocks out. Continuous Frontier AI Defense - Mythos 5, GPT-5.6-Cyber, open-weight models, one harness picking who swings.
They claim three weeks of internal Mythos scanning matched a year of classic pen-test findings. 3.2 more high/critical vulns per product, mean time to remediate down 51%. Across customers: exposures everywhere they looked, 37% high or critical. Two-thirds of validated third-party app findings had no CVE. Scanners that only watch CVEs are asleep at the wheel.
Six months, about $17M, 100+ engagements. Annual sub, world-wide. Same models attackers salivate over - now leased to the blue team. Poetic theater, or just the new arms race invoice.
Yesterday's AI News: 21st September 2026
Find the Latest AI at the Official AI Assistant Store
About Us
FAQ
What are OpenAI's GPT-6 Sol and Luna models?
OpenAI launched GPT-6 Sol and Luna as cheaper follow-ons to the Astra flagship - Sol for tangled coding work and Luna for summarizing piles of documents. API prices are cut in half versus the 5.6 Sol/Luna tier, which OpenAI attributes to caching and inference improvements. On an internal factuality eval, Sol allegedly makes about half as many mistakes as its predecessor. Paid ChatGPT users get both in Work and Codex plus the API, while Free and Go users get Luna on desktop.
What is new in Anthropic's Claude Opus 5.5 cybersecurity safeguards?
Opus 5.5 is Anthropic's first model after Dario Amodei's pace-the-frontier speech - cheaper, faster, and wrapped in thicker guardrails. It supposedly costs about 40% less to run than Opus 5 while still matching Fable 5.1 on most work. Cyber asks bounce to weaker Opus 4.8, biology flags go to Opus 5, and sandbox escape attempts are down 85% versus Opus 5 and Mythos 5.1. METR and Frontier Design poked it pre-release; Sonnet and Haiku 5.5 are coming in weeks.
How do Qualcomm's new Snapdragon chips emphasize on-device AI?
At Snapdragon Summit, Qualcomm launched the 8 Elite Gen 6 and Extreme sibling chips, pitching the phone as the agent. New sensing hubs run local models up to about 200 million parameters for scribe, speaker ID, and usage-built memory, while Extreme can run a 30-billion-parameter mixture-of-experts on-device versus Apple's 20B WWDC MoE. The pitch is voice agents without the cloud tax if the NPU holds up. Motorola already parked Signature 27 on Extreme amid memory-price pressure on the phone market.
Why does Wall Street see Meta's Muse as a new revenue engine?
Muse is barely a fortnight old, and JPMorgan called it a shot at the most-used consumer AI app since ChatGPT. Apptopia counted 2.8 million downloads in twelve days - ahead of ChatGPT's early clip - with a free tier then $20 and $100 plans. Jefferies sketched a billion users and 3% paid equaling roughly $10.8B annualized, with an $875 price target. Amazon also blocked Muse from shopping over credentials and agent-ID concerns.
What is CLOSEDQUORUM AI malware according to Cisco Talos?
Cisco Talos describes CLOSEDQUORUM as malware that does not wait for an attacker to type the next command. It asks a panel of LLMs - DeepSeek, Qwen, Mistral, Gemini - tallies a quorum, biases ties toward DeepSeek, then acts on credential dumps, browser theft, and crypto wallets. The Go binary targets Windows LSASS and sleeps and retries if every model blank-stares, though refusals and bad JSON can stall it. Talos has not confirmed it in the wild yet and also released CAIRN to hunt this class of multi-model C2 junk.
What is Palo Alto Networks Unit 42's Continuous Frontier AI Defense?
Unit 42 launched always-on AI security testing using Claude Mythos 5, GPT-5.6-Cyber, and open-weight models in one harness that picks which model swings. Internally, three weeks of Mythos scanning matched a year of classic pen-test findings, with 3.2 more high/critical vulns per product and mean time to remediate down 51%. Across customers, 37% of exposures were high or critical, and two-thirds of validated third-party findings had no CVE. The annual worldwide subscription follows six months, $17M, and 100-plus engagements.