Easy AI Daily | December 18, 2025 AI Industry Digest
A roundup of model releases, benchmarks, pricing, tools, multimodal/3D generation, user feedback, and infrastructure news from December 18, 2025.
Model Releases and Updates
Google Releases Gemini 3 Flash: Pro-Level Reasoning at 1/4 the Cost
Gemini 3 Flash is positioned as "Pro-level reasoning at Flash speed." It is live in the Gemini app and Search AI Mode, supports 1M-token context and tool calling, and is priced at $0.50 per million input tokens and $3.00 per million output tokens.> Links: Sundar Pichai tweet | Google blog | Google DeepMind announcement
xAI Releases Grok Voice Agent API for Real-Time Speech Interaction
The Grok Voice Agent API provides speech-to-speech interaction with tool calling, web/RAG search, and 100+ languages, scoring 92.3% on Big Bench Audio reasoning. Pricing is $0.05 per minute.> Links: xAI tweet | Benchmark analysis
Microsoft Releases TRELLIS 2-4B: Open-Source Image-to-3D Model
TRELLIS 2-4B uses Flow-Matching Transformers and a sparse-voxel 3D VAE to convert a single image into 3D assets. The 4B-parameter model is open-sourced on Hugging Face with a demo.> Links: Reddit post | Hugging Face | Official blog
Apple Introduces SHARP: Photorealistic 3D Gaussians from a Single Image
SHARP generates photorealistic 3D Gaussian representations from a single image in seconds, requires a CUDA GPU, and is open-sourced on GitHub.> Links: Reddit post | GitHub repo | arXiv paper
QwenLong-L1.5 Released: 4M-Token Long-Context Reasoning
QwenLong-L1.5 achieves SOTA long-context reasoning with 4M-token support, using data synthesis, reinforcement learning, and memory management. Open-sourced on Hugging Face.> Links: Reddit post | Hugging Face
Model Performance and Benchmarks
Gemini 3 Flash Benchmarks: Beats Multiple Mainstream Models
Gemini 3 Flash beats Gemini 3 Pro on ARC-AGI-2 and SWE-bench Verified, enters the Top 5 on LMArena and Vision Arena, and approaches GPT-5.2 on some metrics.> Links: fchollet tweet | DeepMind benchmarks | LMArena leaderboard
Mistral Outperforms Gemini 3 Pro on ARC-AGI2 With Fewer Parameters
Users found a small Mistral model outperforming Gemini 3 Pro on ARC-AGI2, suggesting training methods that force better generalization rather than memorization.> Links: ARC-AGI2 benchmark
QwenLong-L1.5 Excels at Long-Context Tasks, Beating Qwen and Nemotron Nano
Users report QwenLong-L1.5 outperforms standard Qwen and Nemotron Nano on long-context extraction tasks; note the query template usage.> Links: Reddit comments | Hugging Face discussions
Costs and Pricing
Gemini 3 Flash Pricing: $0.50/M Input, $3.00/M Output Tokens
Pricing is 75% cheaper than Pro; accessible via Google AI Studio, Vertex AI, and more.> Links: Google pricing page | Google blog
Claude Opus API Costs: Users Report ~$1.2 for 29K Tokens
Perplexity users cite high Claude Opus API costs and discuss adding it to Perplexity Max plans.> Links: Anthropic pricing
OpenRouter Users Hit Timeout Errors in Production
Multiple users report/completions endpoint timeouts (cURL error 28), affecting production software for some spending $6,000/month.> Links: OpenRouter status
Tools and Integrations
Gemini 3 Flash Enhances Tool Calling, Supports 100+ Tools
Demos show 100+ tools; integrated into Cursor, VS Code, Ollama Cloud, and more.> Links: Google AI demo | Cursor integration
Unsloth Launches CLI Tool to Simplify Model Training
Unsloth added a CLI so users can run scripts directly after pip-installing, replacing Jupyter notebooks and improving automation.> Links: Unsloth GitHub
Qdrant Releases Snappy: Open-Source Multimodal PDF Search Pipeline
Snappy uses ColPali patch-level embeddings and multi-vector search for PDF content search, with production deployment guidance.> Links: Qdrant project | Deployment article
Multimodal and 3D Generation
Tencent Releases Hunyuan HY World 1.5: Real-Time Interactive 3D World Generation
HY World 1.5 supports real-time interactive 3D world generation with Reconstituted Context Memory and Dual Action Representation, in first/third-person views.> Links: Paper
Runway Gen-4.5 Emphasizes Physically Realistic Motion; Kling 2.6 Adds Motion Control
Runway Gen-4.5 improves physical motion realism; Kling 2.6 supports motion and voice control; TurboDiffusion claims 100–205× video generation speedups.> Links: Runway tweet | Kling tweet
LangSmith Launches Observability Tools for Large-Scale Agent Deployments
LangSmith provides OpenTelemetry tracing, pairwise preference queues, and automated evaluation, supporting large-scale agent deployments at Vodafone/Fastweb.> Links: Case study
User Experience and Feedback
Users Report GPT-5.2 Hallucinations and Stiff Replies, Switching to Gemini
Users report "blatant hallucinations" and "stiff replies" from GPT-5.2; some are switching to Gemini 3 Flash.> Links: Reddit thread
User Uses ChatGPT to Aid Bipolar Disorder Management, Says More Effective Than Traditional Therapy
A user with bipolar II reports ChatGPT 5.1 helped manage obsessive thoughts and hypomanic episodes with non-judgmental support, reporting better results than 5 years of traditional therapy.> Links: Reddit post
Gemini 3 Flash User Feedback: Useful Tool Calling, Beats Pro on Some Benchmarks
Users praise tool calling and Cursor integration for developer productivity.> Links: Reddit comments
Infrastructure and Hardware
RTX PRO 5000 Blackwell Specs Leaked: GB202 Chip, 110 SMs
The RTX PRO 5000 uses a GB202 chip with 110 SMs enabled, 300W TDP, FP8/FP16 MMA with FP32 accumulation, and memory bandwidth at 3/4 of the RTX 5090.> Links: NVIDIA datasheet
GPU Rental Platforms Vary in Experience; Network Bandwidth Is Key
Users report large bandwidth differences on platforms like vast.ai; recommend local debugging and setup scripts to reduce waste.> Links: vast.ai
NeoCloudX Offers Cheap GPU Rentals: A100 $0.4/hr, V100 $0.15/hr
NeoCloudX aggregates idle datacenter GPUs: A100 at ~$0.4/hour, V100 at ~$0.15/hour.> Links: NeoCloudX website
---
📌 Source: Easy AI Daily 🤖 Compiled by: AI assistant