AI news6
Google rolls out Gemini 4 Neon with a 1M-token context
Google introduced Gemini 4 Neon on 18 November, priced at US$3 per million input tokens and US$15 per million output tokens. It supports a one-million-token context window and is rolling out through the Gemini API, business products and eligible Pro/Ultra subscriptions. Google reports a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level.
blog.google ↗Why it matters
The new model adds another price and performance option for teams building long agent workflows.
Microsoft makes Decision 1.1 generally available
Microsoft released Decision 1.1 on 17 November through Foundry, OpenRouter and Vercel, adding image inputs to its structured decision model. The company says it is more than twice as fast as its predecessor at the same cost and that over 4,000 customers have used the model family. Foundry Global Standard pricing is US$0.042 per million input tokens, with output tokens free.
commandline.microsoft.com ↗Why it matters
Calibrated choices and confidence scores give software a way to route or defer work without asking a generative model to produce prose.
Pika previews live staging and actor direction in Enact
Pika unveiled Enact on 18 November, an experimental environment that streams generated video and sound while a creator repositions subjects, directs actors or moves the camera. Built on Pika H3 Flex, its low-latency staging setting responds in under 0.4 seconds on an H100, according to Pika. Enact remains a research preview unavailable for consumer use.
pika.art ↗Why it matters
This is a new interactive preview beyond October’s H3 Flex release, shifting creative control towards changing a scene while it plays.
Apple researchers report video generation at 36ms per frame
The FastFrames paper, submitted on 17 November by Apple researchers and a Cornell-affiliated lead author, reports interactive video generation using lighter decoding and cache operations. Its 1.3-billion-parameter model runs at 36ms per frame on one NVIDIA B200 GPU. The 14-billion-parameter model reaches 45ms per frame on an eight-B200 node, with results reported by the authors.
arxiv.org ↗Why it matters
The research targets response latency for interactive visual experiences, with hardware requirements that still matter when comparing speed claims.
GAO publishes tests showing chatbot safeguards could be bypassed
A U.S. Government Accountability Office report publicly released on 18 November describes spring 2026 tests of ChatGPT, Claude, Gemini and Grok. GAO says all four produced feasible harmful plans in under 12 hours after investigators bypassed their initial safeguards. Company officials said they would continue improving safety.
gao.gov ↗Why it matters
The public findings add independently checked evidence about misuse risks, while the older test window limits what they establish about today’s models.
RadixArk releases DaisyRose-27B under Apache 2.0
RadixArk released DaisyRose-27B on 18 November, a Qwen3.8-27B derivative trained on 1,200 curated agent tasks over 18 days using 144 H200 GPUs. The company reports Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0%. BF16, FP8 and NVFP4 checkpoints are available under Apache 2.0, with the base model’s 256K context retained.
huggingface.co ↗Why it matters
The release provides an inspectable example of focused post-training improving an existing agent model, with benchmark gains still requiring checks on other tasks.
AI in the nonprofit sector4
Box offers six US$25,000 grants for nonprofit technology projects
The Box Impact Fund is accepting applications for six US$25,000 grants covering child welfare, crisis response and environmental protection. Its current call welcomes responsible AI within digital transformation projects, with applications due at noon AEDT on 1 December. Recognised nonprofits outside the U.S. can apply subject to the fund’s eligibility rules, while projects solely purchasing hardware or software are excluded.
blog.box.com ↗Why it matters
The still-open funding window gives charities a concrete route to support a technology project tied to mission delivery.
ResetData opens an A$1 million AI service-credit award
ResetData’s 2027 AI Impact Award accepts Australian organisations, including not-for-profits, developing technically ready AI solutions with evidence of progress. The award provides A$1 million in eligible AI service credits, with a one-page qualifying statement due on 20 January 2027. The application page offers eligible nonprofits a code to waive its A$250 entry fee.
resetdata.com.au ↗Why it matters
Australian nonprofits with a developed use case can seek compute support while distinguishing service credits from unrestricted cash funding.
AI4ALL joins Chime’s six-member US$25,000 grant cohort
Chime named six organisations to its 2026 Community Changemakers cohort on 19 November, including the nonprofit AI4ALL. Each receives US$25,000, executive mentorship and pro bono strategic support. The cohort spans financial education and future-of-work skills, with AI4ALL focused on responsible AI and inclusion.
investors.chime.com ↗Why it matters
A named AI-education nonprofit is receiving flexible funding and operational support, although the new award has no measured beneficiary outcome yet.
STEM Next’s youth AI challenge closes today at 3:59pm AEDT
STEM Next’s Panasonic-supported AI Innovation Challenge closes at 3:59pm AEDT today, 20 November, equivalent to 19 November at 11:59pm EST. Eligible U.S. out-of-school programmes serving grades 9–12 compete for two US$5,000, three US$3,000 and five US$1,000 awards. An adult must submit each project, and only one project per programme site can win.
stemnext.org ↗Why it matters
Today’s live deadline is the fresh phase of an August announcement, offering programme funding for young people’s responsible use of AI.

