This issue collects 24 AI updates from 2026-09-23, led by Anthropic releases Opus 5.5 with lower prices and Fable level performance, OpenAI launches GPT 6 Sol and Luna, pitching lower cost and fewer mistakes, Meta admits Muse's likeness to OpenClaw isn't a coincidence, OpenAI Improves Prompt Caching for GPT 6.
Anthropic has released Opus 5.5, describing it as the strongest performing model the company has tested to date.The new flagship arrives with lower pricing and performance the report places at Fable level, putting cost alongside capability as a headline claim.
OpenAI has launched two new models, GPT 6 Sol and GPT 6 Luna, describing them as cut from the same cloth as Astra while pitching lower cost and fewer mistakes.TechCrunch reports the pair arrived together, a sign that price and reliability, not just benchmark scores, are now the headline claims of a frontier release.
NVIDIA used its official blog on September 23 to highlight AI Day Singapore, a two day event held September 22 to 23 at the Raffles City Convention Centre.The company and its partners showcased AI advancements across Southeast Asia alongside hands on training and expert led sessions.
According to QbitAI, Zhi Huijun is offering a robot priced at 20,000 yuan per unit, one that can take a humanoid or a quadruped form and is open to developers.The report frames the price with a direct question: what kind of robot do you actually get for 20,000 yuan?
Qwen has settled who runs the model family, according to a report from QbitAI, with Liu Dayiheng taking over the top role.The report notes he comes from the same cohort of “talent prodigy” engineers as the widely known ZhiHuiJun, and the change could affect release cadence, open source strategy and community communication.
Banma Smart introduced AutoOmni 2.0 23B A3B, a new generation on device omni modal large model, during the Yunqi Conference on September 23.The company frames the release as making its Yuanshen AI understand "my world" better.
OpenAI and Grab have launched GO Forward with AI, a regional programme intended to help 30,000 partners build practical AI skills across Southeast Asia.The announcement frames the effort around real work scenarios but does not yet detail curriculum, country coverage, timeline, or cost.
OpenAI is seeking independent reviews of its AI safeguards and of serious incidents, according to a report by citybiz, addressing the question of who verifies frontier model safety claims.The move would shift safety assessment from an internal process toward a more externally supervised framework.
Deepal and Volcano Engine have officially announced a partnership that puts ByteDance's Doubao large model into a production vehicle for the first time, in the Deepal S07 AI Laser Edition.The car goes on sale September 28, moving large model cockpit features from launch event demos into a dated production rollout.
A Sina Finance AI industry roundup published on September 23 reports that Qwen has released Qwen Audio 3.1, an update to its audio model line.The available summary confirms only the release itself, with no detail on size, language coverage or licensing.
YouTube is rolling out new AI features inside its Studio app, including tools to generate content ideas and monitor how thumbnails perform.The additions land in the dashboard creators already use for uploads, management and analytics rather than in a separate product.
YouTube's new custom feeds let users describe the videos they want to see in their own words, then use Gemini to build a personalized feed around that request.It hands part of the recommendation logic back to viewers and marks another consumer facing placement for Gemini inside a core Google product.
TechCrunch reports that Meta has acknowledged the resemblance between its AI assistant Muse and OpenClaw is not a coincidence, saying Muse was heavily inspired by OpenClaw.According to the report, that similarity extends to some workspace filenames and content.
QbitAI reports that Alibaba's Qwen AI platform has been upgraded across model services, agent services, and AI applications.The update spans MaaS, agent tooling and industry specific AI solutions, covering the path from model calls through to business deployment.
Analytics India Magazine reports that Meta is testing human contractors to take over calls for its Muse AI agent, filling gaps the agent cannot handle on its own.The story is attributed to a report rather than an official announcement, and the scale, regions, and launch timing remain undisclosed.
Chinese outlet QbitAI reports that Huizhi Intelligence has launched Hellome, positioned as China's first FDE direct agent service platform, with the aim of compressing enterprise AI delivery cycles into weeks.The report frames the launch as part of a shift toward platform based delivery of enterprise AI services.
OpenAI has detailed a set of prompt caching improvements for GPT 6, including higher cache hit rates, new diagnostics, explicit breakpoints and additional controls.The changes aim to cut latency and cost for repeated prompts while making cache behaviour easier to observe and steer.
Qualcomm has launched two new smartphone chips, placing on device AI at the centre of the announcement.The company says its top end chip can run a 30B mixture of experts model locally, pushing larger model inference onto the phone itself.
Lenovo brought its Tianxi AI to Alibaba Cloud's Apsara Conference with a full scenario, multi device product matrix, framing the showing as a push to land "super organization" capability on the device side.The multi device combination is the centerpiece of the appearance.
Snorkel AI has raised a $350 million Series E that triples its valuation to $3.5 billion, TechCrunch reports.The seven year old startup sells a data as a service approach to building AI training data, a layer attracting fresh capital as demand for training data booms.
Ema has raised $77 million in a new round, bringing its total funding to $140 million, and now counts more than 50 enterprise customers including Google and Microsoft.TechCrunch frames the round as evidence that AI agents are beginning to absorb spending that once went to enterprise software and services.
In a TechCrunch interview this week, Greek Prime Minister Kyriakos Mitsotakis said no government is ready for what AI is about to do.The piece is framed around his remark that leaders are already fighting yesterday's battle.
OpenAI is extending access to its Daybreak program to the Government of Ukraine to support cyber defense of civilian infrastructure.The announcement gives no technical scope, delivery details or timeline, and coverage such as the BBC's report centers on the same core point.
NVIDIA has published Nemotron 3 Diarization, a 100 million parameter open weight model that labels who spoke when for up to eight speakers in live and recorded audio.NVIDIA says it ranks first on VoiceArena's Diarization Bench with a 14.72 percent error rate and cuts diarization error by about 40 percent at 1.
