The AI log, frontier · image, video, world · open weights
18 Nov 2026
Gemini 4 Neon, Genie 4 live, Rumble-2-Flash, DaisyRose-27B
Google ships Gemini 4 Neon and opens Genie 4 to Ultra users, while two Apache 2.0 releases cover a 0.5B speech model and an agent-tuned 27B.
By Ali Asif, published
What shipped
- Google DeepMind
Introducing Gemini 4 Neon
A frontier model priced at $3 and $15 per million input and output tokens with a 1 million token context window, reporting 61 on the Artificial Analysis Intelligence Index and 65.2% on Terminal-Bench 4.1.
- Google DeepMind
Genie 4 is now available
Genie 4 moves from preview to eligible Google AI Ultra subscribers, with real-time prompt edits, native audio, a Project Genie agent and a community gallery.
- Reson8
Rumble-2-Flash: an open 0.5B speech model for 40+ languages
An Apache 2.0 speech recognition model that reports a 6.96% average error rate on the Open ASR multilingual track and over 7,000 RTFx on full batches.
- RadixArk
DaisyRose-27B: Pushing the Frontier of Post-Training Efficiency
An Apache 2.0 reinforcement-learning tune of Qwen3.8-27B that reports Terminal-Bench 4.1 pass@1 rising from 5.6% to 12.0% after 18 days on 144 H200 GPUs.
Four items today, all from pages I opened and read. I have not run any of them, so every number below belongs to the publisher. Google has the big day, and two open releases are worth a look next to it.
Gemini 4 Neon
Google DeepMind describes Neon as frontier-level intelligence with fast end-to-end task completion across software engineering, knowledge work and science. It reports a score of 61 on the Artificial Analysis Intelligence Index at its xhigh thinking level, which it places at number 3 overall, just behind larger and slower models. On the Vals Index it reports number 2 at 70.9%, and it lists number 1 on the Legal Research Bench and the Tax Agent Bench.
The pricing is the part I noticed. Input costs $3 per million tokens, cached input $0.15, output $15, with a 1 million token context window. It is rolling out in the Gemini API in Google AI Studio, in Gemini for business, in Google Antigravity for paid Pro and Ultra subscribers, and in the Gemini app for Pro and Ultra.
On coding, Google reports 65.2% on Terminal-Bench 4.1 and 46.2% on FrontierCode 2.0 at the default high setting, and says Neon leads the Artificial Analysis Coding Agent Index at 58.6. On SWE-sweep it reports 12.8% in the mini-SWE setup and 39.0% inside Antigravity, which tells me how much the surrounding tooling matters. Google also says Neon ranks number 2 on steerability on Arena's new Agent Arena Index and records its lowest fabrication rate yet in internal Antigravity testing. That last claim is internal, so I would treat it as a hint, not a result.
Genie 4 reaches Ultra subscribers
Genie 4 was a waitlist preview when I covered it in October. Now DeepMind is rolling it out to eligible Google AI Ultra subscribers inside Project Genie. You can type a new prompt mid-session to change the weather, add a character or redirect the action, and the world model pairs the visuals with real-time native audio.
Genie 3 stays in the model picker with updates: sessions of up to 3 minutes, up from 60 seconds, linked worlds, and seeding from 360-degree panorama images. The new Project Genie agent builds a world from a conversation, and a community gallery lets people publish and remix. Google AI Pro members can join a waitlist.
Rumble-2-Flash
Reson8 open sourced a 0.5B speech recognition model under Apache 2.0, covering 40+ languages with automatic language detection. It uses the Qwen3-ASR architecture, so it runs in Transformers and vLLM without custom code. The team distilled its larger Rumble-2 model into it with on-policy distillation, where the student transcribes audio itself and the teacher grades each token.
The reported numbers: a 6.96% average error rate on the Open ASR Leaderboard multilingual track, ahead of Whisper large-v3 and Meta's omniASR-LLM 7B, and 6.6% on the English track. Throughput passes 7,000 RTFx on full batches, and on a 302-hour call-centre benchmark it is on par with NVIDIA's 30B Nemotron 3.5 Transcribe.
DaisyRose-27B
RadixArk released its first model, built on Qwen3.8-27B and trained with reinforcement learning using Miles. It used 1,200 curated long-horizon agent tasks over 18 days on 144 H200 GPUs, with sessions up to 1 million tokens. Terminal-Bench 4.1 pass@1 goes from 5.6% to 12.0%, and the post lists gains on Terminal Bench 2.1, DeepSWE, SWE Bench Pro, HLE, IFBench and GPQA Diamond. Checkpoints in BF16, FP8 and NVFP4 are Apache 2.0, and the architecture and 256K context stay the same as the base.
What I'd watch next
Independent reruns of Neon's coding numbers outside Antigravity. How Genie 4 behaves for people who are not on Ultra. Whether Rumble-2-Flash holds up on my own noisy audio. And whether others copy RadixArk's task-curation recipe.
Sources
Topics: Google DeepMind, Gemini, World Models, Agents