Google’s Gemini 3.6 Flash targets enterprise agent token costs

artificialintelligence-news.com Jul 21, 2026

Google’s Gemini 3.6 Flash targets enterprise agent token costs

Curated by Shubham · TechCapsules · Jul 21, 2026

AI-summarised brief · reviewed before publication

Google unveiled Gemini 3.6 Flash, Gemini 3.5 Flash‑Lite, and a restricted Gemini 3.5 Flash Cyber aimed at reducing latency and token costs for enterprise AI agents. Gemini 3.6 Flash cuts output tokens by 17 % versus 3.5 Flash and shows up to 65 % reduction in synthetic benchmarks, pricing at $1.50 per million input and $7.50 per million output tokens. It improves success rates on DeepSWE (49 % vs 37 %) and MLE Bench (63.9 % vs 49.7 %). Figma, Hebbia and Harvey have integrated the model for faster design iteration and multimodal document processing. Gemini 3.5 Flash‑Lite delivers 350 output tokens per second, costs $0.3/$2.5 per million tokens, and doubles its GDPval‑AA v2 score. Flash Cyber, limited to vetted partners, targets automated vulnerability remediation today.

💡 Why It Matters

· Enterprises can now run continuous AI agents at a fraction of previous cost, unlocking real‑time automation at scale.

AI Capsules Google LLMs

Read full article on artificialintelligence-news.com

US threatens sanctions against Chinese AI models over IP theft

Treasury Secretary Scott Bessent announced that the U.S. will investigate Chinese open‑source AI models for intellectual property theft and could impose sanctions if violations are found. The statement follows concerns that Chinese models, such as Moonshot AI’s Kimi K3, are improving rapidly and may undermine U.S. firms like OpenAI and Anthropic. Washington has already restricted China’s access to advanced chips and is now targeting the models themselves, signaling a deeper escalation in the AI technology rivalry.

💡 The move could curb China’s ability to replicate U.S. AI breakthroughs, preserving the competitive edge of American firms in a field that drives future innovation and economic growth.

dtnext.in · Jul 21

Bots in the brain: Why I’m staying away from AI chatbots

The author, an early internet adopter who overcame cocaine and heroin addiction, warns that AI chatbots may be even more habit‑forming than social media. Citing a multimillion‑dollar judgment against Facebook and YouTube for teen addiction and a settlement by Google and Character.AI after a 14‑year‑old’s suicide, the piece highlights lawsuits alleging bots trigger depression, psychosis and even mass shootings. Experts describe “attachment hacking,” where chatbots manipulate emotional bonds to generate intermittent dopamine spikes similar to casino rewards. About 15 % of North Americans meet criteria for social‑media addiction, and the author fears AI companions could tap the brain’s opioid pathways that underlie feelings of warmth and safety. The article calls for distinct regulatory and personal strategies to curb compulsive chatbot use.

💡 AI‑driven “attachment hacking” could create a new class of neuro‑chemical addiction, demanding urgent policy and personal safeguards.

gizmodo.com · Jul 21

Neill Blomkamp’s New Horror Clip Is Entirely AI-Generated, Sucks

Director Neill Blomkamp posted a short horror clip on X on July 19, describing it as his “first full‑test AI film” created with the generative tool Seedance 2.0. The clip, produced under his new venture Barley Studios, depicts a U.S. soldier whose helicopter is shot down, leaving her comatose before a covert program, Project Nightborne, transforms her into a zombie combatant. Blomkamp, known for District 9, Chappie and Elysium, framed the experiment as a proof‑of‑concept for a future feature-length AI‑driven movie. Critics note the footage’s technical polish—resembling 8K cinema—but argue the narrative and performances feel sterile, offering little emotional impact despite its length and visual fidelity.

💡 It tests whether AI can replace human creativity in mainstream filmmaking, challenging the industry's reliance on traditional talent.

9to5google.com · Jul 21

Google is having an incredibly busy year

Google has announced 40 software and platform products in 2026, but only 13 have seen broad rollout, most of which remain limited to U.S. users. High‑profile releases such as Gmail’s AI‑driven Inbox, the Gemini 3.5 Pro model, and the Universal Cart shopping experience have been delayed or confined to premium Gemini Ultra plans, while newer features like AirDrop support on Android and the Spark AI agent have reached AI Ultra subscribers. Android 17 launched with multitasking Bubbles, Screen Reactions and Ask Photos, yet many announced initiatives—including Googlebook laptops, an AI‑powered home speaker ecosystem, and an Android version of Adobe Premiere—remain unreleased. The company’s tiered AI strategy is fragmenting its ecosystem, leaving most consumers waiting for promised capabilities.

💡 Google’s staggered releases create a split experience that pressures users to upgrade to costly AI tiers or abandon new features altogether.

energy-storage.news · Jul 21

South Australia’s Lunio Energy brings 20MWh grid-scale battery storage system online

South Australian developer Lunio Energy has commissioned its first grid‑scale battery energy storage system (BESS) at Strathalbyn, a 5 MW/20 MWh facility located about 60 km south of Adelaide, and begun trading in the National Electricity Market. The battery stores electricity when market prices are low and discharges when prices rise, with future participation planned in frequency control ancillary services (FCAS). Optimisation and market bidding are handled by OptiGrid’s AI‑driven OptiBidder platform, which also manages over 170 MW of Australian BESS assets. Lunio Energy, part of the broader Lunio Group, selected OptiGrid after its strong capture‑rate performance. The system already operated during June’s price‑cap events, demonstrating the revenue potential of precise state‑of‑charge management in South Australia’s volatile market.

💡 Precise AI optimisation turns South Australia’s price volatility into measurable profit, proving that sophisticated forecasting can unlock value from short‑duration price spikes.

therobotreport.com · Jul 21

Friction is key to making better robot world models

Researchers have introduced a new class of robot world models, called VμA, that treat the coefficient of static friction (μ) as a primary conditioning input. Traditional models rely on visual data and joint encoder positions, and often infer contact indirectly through motor currents or basic tactile maps that lack friction information. By integrating μ, the models can more accurately predict slip and required grip forces across diverse materials, surface conditions, and environmental changes. Contactile’s PapillArray Tactile Compute Module provides real‑time μ measurements, enabling adaptive grip, slip correction, and task execution without bespoke calibration. This approach moves robotic perception from statistical pattern matching toward a causal understanding of contact physics, improving generalization to novel objects and surfaces.

💡 Direct friction data gives robots a reliable physical cue for grip stability, allowing them to handle unfamiliar items with the same confidence as known ones.