The last 24 hours in AI were dominated by two same-day launches: OpenAI and Anthropic both shipped new models and cut prices on the same day — the price war is officially on. Meta's high-privilege assistant, meanwhile, was hit with a security disclosure. Here are the 8 stories worth scanning.
OpenAI launches GPT-6 Sol and Luna, halving API prices
On September 22, OpenAI announced two new members of the GPT-6 family on its official blog. GPT-6 Sol targets complex coding and agentic workflows at $2 per million input tokens and $10 per million output tokens; the lighter GPT-6 Luna costs just $0.10 input and $0.50 output — half the promotional price of GPT-5.6. Both models are live in ChatGPT Work, Codex, and the API, with Luna also available to Free and Go users in the desktop app.
The takeaway: token bills for running agents are about to get noticeably thinner — Sol is aimed at high-frequency call patterns, Luna at high-volume but simple jobs like document summarization and information extraction. With Anthropic cutting prices the same day, developers are enjoying a rare buyer's market.
Anthropic ships Claude Opus 5.5 the same day, 40% cheaper
In a September 22 newsroom announcement, Anthropic introduced Opus 5.5, the first model in the Claude 5.5 family: 40% lower cost than Opus 5 on typical workloads, with input and output down to $4 and $20 per million tokens and over 30% faster output. It leads on agentic coding and knowledge-work benchmarks such as Terminal-Bench 4.0 and tops the Artificial Analysis Intelligence Index. Claude Code also shipped v2.1.280 the same day, switching its default Opus model to 5.5.
Two giants launching and discounting on the same day is no coincidence: frontier-model competition has shifted from "who's smarter" to "who's cheaper and faster." The direct upside for users: five-hour usage limits on Pro, Max, and other subscription tiers go up as well.
OpenAI upgrades prompt caching for GPT-6, up to 90% off cached reads
Also on September 22, OpenAI released an improved prompt caching system and diagnostics tooling for GPT-6: higher cache hit rates, faster responses, and a 90% discount on cached input-token reads. Combined with the price cuts, costs for long conversations and repeated agent workloads drop further. Developers should check the new caching dashboard in the console to confirm their call patterns are actually capturing the savings.
Qwen releases the Qwen-Audio-3.1 voice lineup with steep price cuts
On September 23, Qwen officially announced the Qwen-Audio-3.1 series: upgraded ASR, TTS, and Realtime models plus two newcomers, TTS-Next and ASR-Next. Prices fall across the board — roughly 70% off TTS, 85% off Realtime, and up to 95% off ASR. The API price war is raging in the voice lane too; our AI news channel covered this release in detail earlier today, so we won't repeat it here.
Qwen-Image-2.1 tops Alibaba's AI Arena open-source leaderboard
On September 23, Qwen announced that Qwen-Image-2.1 took first place among open-source models in image editing and text-to-image on its own AI Arena evaluation platform. Note this is a vendor-run leaderboard, and independent third-party evaluations are still scarce — wait for more blind-test data before judging real-world quality. Either way, Chinese open-source image models are climbing the public leaderboards fast.
Meta's Muse hit with a 0-day, hotfix already out
On September 22, Ars Technica first reported that macOS security researcher Patrick Wardle found a zero-day in Meta's newly launched AI assistant Muse: any local app, with no special permissions, could modify undocumented Muse settings, redirecting the dictation transcription endpoint to an attacker-controlled server and hijacking the entire Muse account. Wardle's proof-of-concept demos included silently writing files and snapping photos. Meta has pushed a hotfix, stressing this was local privilege escalation, not remote exploitation. The same day, Amazon confirmed it had asked Meta to stop letting Muse shop on Amazon.com on users' behalf.
The real story isn't the bug, it's the architecture: a personal agent that demands every permission becomes far more dangerous than an ordinary app once its design has a weak spot. Meta's launch narrative around its "secure VM" just got a serious stress test.
Pentagon probe: overreliance on Palantir's Maven AI contributed to school strike
On September 22, Gizmodo reported on Bloomberg's investigation: an internal Pentagon review found that stale intelligence, outdated satellite imagery, and US Central Command personnel leaning too hard on Palantir's Maven Smart System were among the causes of the February 28 strike on a school in Minab, Iran, which killed 123 children. Investigators said some operators wrongly assumed the system would automatically flag expired or contradictory intelligence. Palantir responded that it is not responsible for underlying data and that no evidence shows its software was at fault.
This is the first time AI-assisted decision-making has entered military accountability in such a devastating way: the failure wasn't a dumb model, but humans outsourcing verification they should have done themselves.
OpenRouter publishes its 2026 best embedding models guide
On September 23, OpenRouter published a 2026 guide to the best embedding models on its official blog, covering 37 directory entries. Teams doing RAG or semantic-search model selection can use it as a cheat sheet instead of reading every doc page one by one.
