logo Gemini 3.7 Flash Gains on Coding, Price Doubles in 2027

Gemini 3.7 Flash Gains on Coding, Price Doubles in 2027

Dosa AI Tools 3 min read nvidia funding models agents
In this brief · 8 sections

This brief covers AI news from 2026-08-22 UTC.

Google’s pricing clock starts ticking, Liquid AI speeds up decoding, and Rillet raises $100M in two days.

Gemini 3.7 Flash: Big Benchmarks, Price Doubles in 2027

What happened: Google’s Gemini 3.7 Flash, released August 13, shows major coding gains: DeepSWE up to 65.3%, AutomationBench to 30.4%. Intro pricing of $0.75/$3.75 per million tokens doubles on January 1, 2027.

Why it matters: A 62% input cost cut versus Claude Sonnet 5 until year-end, but teams must plan for the rate hike or lock in savings now.

Source: byteiota

Nvidia Pays $6B for Poolside AI Tech and Talent

What happened: Nvidia struck a $6 billion licensing agreement with Poolside AI, committed $1 billion in fresh capital, and extended job offers to 109 engineers. The deal values Poolside at $12 billion pre-money.

Why it matters: Gives Nvidia broad access to Poolside’s model-development software without absorbing the company, a new template for chipmakers to secure AI software leverage.

Source: WebProNews

DeepSeek Launches V4-Flash Vision Model with Free Files API

What happened: DeepSeek released DeepSeek-V4-Flash-Vision-Exp on August 21, adding multimodal support. A new Files API lets developers reuse uploaded images for free via file_id references.

Why it matters: Cuts repeated image upload costs for multimodal agent workflows, and DeepSeek claims agent performance near Opus-4.8 on visual tasks.

Source: Meyka

Liquid AI’s DSpark Draft Models Speed Decoding Up to 3.18x

What happened: Liquid AI released DSpark draft checkpoints for three LFM2.5 models, adding speculative decoding paths. Up to 3.18x faster on H100, 2.87x on M4 Max, with identical outputs.

Why it matters: Self-hosted developers get faster inference without accuracy loss, with day-one llama.cpp and SGLang support.

Source: The News 92

NVIDIA AVO Hits 100% on ARC-AGI-3 Benchmark

What happened: NVIDIA’s Agentic Variation Operators (AVO) architecture achieved a 100.00 RHAE score on ARC-AGI-3, completing all 183 levels. It also beat FlashAttention-4 by up to 10.5% in GPU-kernel optimization.

Why it matters: Shows system design, not just model capability, can unlock frontier-level long-horizon agent performance, a blueprint for building robust agent harnesses.

Source: NVIDIA Technical Blog

Rillet Raises $100M at $1B Valuation in 48 Hours

What happened: AI accounting startup Rillet raised $100 million at a $1 billion valuation, becoming a unicorn in 48 hours. It has 600 customers and doubled annualized revenue in the last quarter.

Why it matters: Signals strong investor appetite for AI-native vertical SaaS that displaces legacy ERP systems like Oracle and NetSuite.

Source: TechCrunch

Civil Society Groups Urge FTC Probe into AI Book Destruction

What happened: Civil society organizations sent a letter to the FTC asking for an antitrust investigation into practices like Anthropic’s effort to ‘destructively scan all the books in the world’, calling it a potential unfair method of competition under Section 5.

Why it matters: Could reshape how AI companies acquire training data, potentially raising costs for startups and affecting data sourcing strategies.

Source: MLex

Sources

Share

Discuss with AI

© 2026 dosa.dev