Tech • AI • Robotics • Game

ENFR

Google Unleashes Gemini 4 Neon, Microsoft Drops 3 New Models, New RSI, Kling Reaches 8K...

9.5/10
AIAI RevolutionNovember 19, 2026 at 09:48 PM13:03
Audio player

⏱ ≈ 4 min read · 13-min video — 9 min saved

TL;DR

Google, Microsoft, Kling, and HeyGen each launched major AI systems this week, led by Gemini 4 Neon, which posted top benchmark results while undercutting rival pricing.

KEY POINTS

Gemini 4 Neon enters between Hydrogen and Carbon

Google launched Gemini 4 Neon just three weeks after Gemini 4 Carbon, placing it between Hydrogen and Carbon in the lineup. Neon is positioned for coding, office work, and science, with an added X High reasoning mode that Carbon does not have. On the Artificial Analysis Intelligence Index, Neon scored 61, ahead of Carbon at 60 and Hydrogen at 54.

Pricing undercuts premium rivals

Neon is priced at $3 per million input tokens and $15 per million output tokens, below Carbon’s standard $4 in and $20 out. Carbon is temporarily cheaper under a launch promotion at $2 in and $10 out through November 30, creating the unusual situation of a larger model costing less than the mid-tier option. In one reported head-to-head coding task, Neon took about 20 minutes and cost $3.50, while Claude Opus 5.5 took 18 minutes and cost $720.

Benchmark results favor Neon overall

Google’s comparison chart placed Neon ahead of Claude Opus 5.5 and GPT-6.2 Soul overall, with scores of 61, 59, and 57 respectively. Neon posted about 65% on Cursor Bench, 39.0 on SWE-bench style bug fixing, and led in math, science, and vulnerability finding. Its biggest edge came on graph walks, where it scored just over 95%, versus about 80% for Soul and 67% for Opus.

Google showcased unusually large interactive demos

Demonstrations highlighted Neon generating complex software and simulations directly in the browser. Examples included a 10,200-line graphics program for a skiing game with physics running at 120 frames per second, and a spreadsheet-based rendering system using 100,000 cells and more than 500,000 live formulas to create image effects without scripts. Other demos showed autonomous additions such as extra tools, buildings, and autopilot modes beyond the original prompt.

Project Genie expands browser-based world building

Ultra subscribers also gained access to Project Genie, built on Genie 4, Google’s world model. The system lets an agent generate interactive 3D worlds in a browser, remix existing creations from a gallery, and run without specialized graphics hardware. The feature points to broader consumer-facing experimentation beyond text and coding assistants.

Google adds memory to recursive self-improvement

Researchers from Google Cloud AI, Google DeepMind, and Penn State introduced Self-Improvement Memory, or SIM, a notebook-like system that records every attempted improvement to an AI harness, including failures. Each entry stores the reason for a change, expected effects, actual results, and the model and test used. In office-task evaluations, a setup starting near 68 reached about 85 with SIM, compared with 73 without it.

SIM improved transfer across models and tasks

The approach was designed to preserve operational experience when models or tasks change. A harness tuned for Claude Opus 5.5 initially performed worse on Gemini 3.8 Flash than Gemini’s original harness, but with SIM it became the best setup across all four test areas, by as much as 11 points. In 10 identical coding experiments, the SIM-enabled version won nine.

Microsoft researchers had Opus hand-design a neural network

A separate paper from Microsoft Research, Johns Hopkins, Columbia, and NYU described a method called IWSP in which Claude Opus 5.5 built small transformer networks directly in Python instead of training them conventionally. After roughly 39 rounds per run, the resulting systems scored 100% on hidden tests across 13 algorithmic task types and 26 runs. The median network size fell dramatically with stronger models, from about 100,000 parameters for Haiku 5.5 to 166 for Opus 5.5.

Microsoft launched low-cost decision and coding models

Microsoft Decision 1.1 acts as a dispatcher model that selects allowed actions, reports confidence, and can defer uncertain cases. It scored 68 on Artificial Analysis’s decision index, ranked eighth, and improved on the prior release while adding image input. Pricing was set at 4.2 cents per million input tokens, with free output, and the model runs on Google’s Gemma 4, underscoring cross-industry dependence even among direct rivals.

Kling and HeyGen pushed image and avatar realism

Kling 4.0 added a unified image generation and understanding system with native 8K output, support for prompts up to 15,000 tokens, more than 110 color grades, and editable text across seven languages. HeyGen Avatar 6 introduced photo-based talking avatars at 2880 by 1620 resolution with no fixed length limit and selective voice editing in existing recordings. In HeyGen’s internal realism test, viewers chose the AI clip as the real person 42.2% of the time when shown beside actual footage.

CONCLUSION

This week’s releases suggest competition is shifting from raw model size to a mix of price, transfer learning, tool use, and realism. Google currently made the strongest immediate commercial move with Gemini 4 Neon, while Microsoft, Kling, and HeyGen pushed specialized capabilities that could quickly shape enterprise and consumer AI markets.

Ask a question

More from AI