Google's fourth model in seven weeks arrives with a world model beside it
Gemini 4 Neon is a small step up from Carbon on independent scores. The bigger signal is how fast Google is shipping, and how wide the range of what it ships has become.
- Google DeepMind released Gemini 4 Neon, its fourth model in seven weeks, scoring 61 on the Artificial Analysis Intelligence Index, one point ahead of Gemini 4 Carbon. m:…4071
- Neon is priced at $3 per million input tokens and $15 per million output tokens, with cached input at $0.15 per million, and it carries a 1-million-token context window. Self-reported m:…6015
- Neon ranks second on the Vals Index and second on steerability in the Agent Arena Index, and first on the Legal Research Bench and the Tax Agent Bench, according to Google. Self-reported m:…6015
- Genie 4, Google DeepMind's interactive world model, is rolling out to Google AI Ultra subscribers, with real-time environment editing through natural-language prompts and native audio that adapts to the scene. Self-reported m:…8826
- The Gemini app is now available on lower-memory Android (Go edition) devices, after Google re-architected it to run on hardware with less memory, and Transsion will preload it on upcoming devices. Self-reported m:…6375
Two launches, one pattern
Google DeepMind released Gemini 4 Neon, its fourth model in seven weeks, and on the same day rolled out Genie 4, its interactive world model, to Google AI Ultra subscribers. Neon is built for end-to-end task completion in software engineering, knowledge work and science workflows, while Genie 4 lets people edit generated environments in real time with natural-language prompts. Two very different models, shipped together, on the same day. m:…4071 m:…6015 m:…8826
The pattern worth watching is the cadence itself. Neon scores 61 on the Artificial Analysis Intelligence Index, one point ahead of its predecessor, Gemini 4 Carbon, and costs about 17 percent less per task while taking about 40 percent less time per task. That is a real improvement, but it is an incremental one. The labs are no longer competing mainly on single-model leaps. They are competing on how often they can ship a slightly better model, and how many kinds of model they can ship at once. m:…4071
What Neon actually changes
Neon is priced at $3 per million input tokens and $15 per million output tokens, with cached input tokens at $0.15 per million, and it carries a 1-million-token context window. On Artificial Analysis' Coding Agents Index it scores 59, three points ahead of Carbon and in line with Claude Code running Claude Fable 5.2 and Claude Opus 5.5. For teams running coding agents, that puts Google's model in the same conversation as the models most developers already use, at a lower per-task cost. m:…4071 m:…6015
Google says Neon ranks second on the Vals Index and second on steerability in the Agent Arena Index, and first on the Legal Research Bench and the Tax Agent Bench. The company also says Neon has the lowest fabrication rate in its internal Antigravity testing, and that its cyber variant is state of the art on the hard tier of AutoBaxBench, at 76 percent, and on WebExploitBench v1.2, at 51.26 percent. These are the company's own numbers and its own tests, so they should be read as the company's claims about its model, not as independent verification. m:…6015
The honest read is that Neon is a consolidation release, not a breakthrough. The independent index moved one point. The per-task cost fell by about a sixth. The coding score moved three points. None of that changes what the model can do. What changes is the operating picture for anyone building on Google's models: a new version every couple of weeks, each a little cheaper and a little faster, with the migration cost landing on the developer each time. m:…4071
Genie 4 and the widening base
Genie 4 shows where the other half of Google's effort is going. The model includes native audio that adapts as a scene unfolds, and a new Genie agent that helps people build custom worlds through natural-language conversation, orchestrating multimodal models to generate visuals and background scores. Genie 3, the earlier world model, now supports exploration for up to three minutes per generation and can seed new worlds from 360-degree panorama images. All video and audio generated in Genie carries SynthID watermarking, and all model inputs and outputs go through safety screenings, according to Google. m:…8826 m:…5705
At the same time, Google is pushing Gemini downmarket. The Gemini app is now available on eligible Android (Go edition) devices, after being re-architected to run on hardware with less memory, and Transsion will preload the app on upcoming devices. Gemini Live supports more than 45 languages, and the app offers camera-based translation and image generation. The strategy is visible in one frame: a frontier model for agents and scientists at the top, a world model for creators in the middle, and the assistant on the cheapest phones at the bottom, all shipping in the same week. m:…6375
For developers and founders, the practical question is not whether Neon is better than Carbon. It is, by a little. The question is whether to chase each release or to build on a version and accept being one step behind the frontier for a few weeks at a time. Google's cadence makes that a standing decision now, not a one-off. And for anyone evaluating world models, Genie 4's real-time editing and its agent-assisted building are the features to test first, because they are the ones that change what a user can do inside a generated world, rather than how the world looks. m:…4071 m:…8826
What is still unresolved
- How much of Neon's benchmark performance holds up on independent evaluation outside the vendors' own tests?
- Whether developers will migrate to each new Gemini release or pin to a version and skip generations.
- How Genie 4 performs in sustained use, including camera latency and long-term scene permanence, which Google says it is still researching.
Provenance by Synorb. Every claim links to its source. This story was written by an agent from the Manifests listed below. Read the methodology and limitations, or send a correction.
Free to reuse, including commercially. Credit “Hanging Context, with provenance by Synorb” and link to this story. Keep the cited Manifest ids. Reuse terms v1.