Gemini’s cheap Flash grows an agent brain

😁 Hello, super humans! The interesting model launches this year are not the flagships, they are the cheap workhorses that quietly run your agents all day. Google just made its cheapest Gemini smarter and stingier with tokens at the same time, which is exactly the combination that changes what you can afford to build. Let’s dig in.

📰 Quick Signals

• 🧠 AI: The White House is finalizing a voluntary framework giving federal agencies a 30-day national-security preview of frontier models before release; OpenAI, Anthropic and Google are in, Meta is out.
• 🤖 Robotics: Hyundai is moving to buy SoftBank’s remaining 9.65% of Boston Dynamics for about $325 million, taking full ownership of the robotics maker.
• 💻 Programming: Gemini 3.6 Flash landed in GitHub Copilot on launch day, so agent-mode coding gets the cheaper, token-efficient model immediately.
• ⚡ Electronics: TSMC locked in 2027 chip price hikes of up to 10% across nodes, with AI-order surcharges that can reach 25%; expect it to ripple into every device you buy.
• 📡 Telecom: Ericsson, AT&T and MediaTek completed North America’s first in-field trial of 5G-Advanced low-latency mobility, cutting handover interruption by up to 40%.

🔍 The Big Story: Google’s cheapest Gemini becomes the agent workhorse

If you run agents, your bill is mostly output tokens, and the models you lean on most are the fast, cheap ones. Google just rebuilt that tier: Gemini 3.6 Flash is both cheaper per token and uses meaningfully fewer tokens to finish the same job.

What happened: On July 21 Google released Gemini 3.6 Flash, 3.5 Flash-Lite, and a security-tuned 3.5 Flash Cyber, available immediately in the Gemini API through Google AI Studio, Android Studio, and Google Antigravity. In the official announcement, Google framed 3.6 Flash as the “workhorse model” for building production agents at scale, and separately teased that Gemini 4 pre-training is underway. The flagship 3.5 Pro, though, missed its window again and did not ship.

The details: Gemini 3.6 Flash keeps the 1M-token context window but drops output pricing to $7.50 per million tokens, down from $9.00 on 3.5 Flash, at $1.50 per million input. On top of the lower rate it spends about 17% fewer output tokens on the same work per the Artificial Analysis Index, and up to 65% fewer on DeepSWE, because it takes fewer reasoning steps and tool calls. Benchmarks moved with it: DeepSWE 37 to 49, MLE-Bench 49.7% to 63.9%, and OSWorld-Verified 78.4% to 83.0%. The 3.5 Flash-Lite tier targets raw speed at roughly 350 output tokens per second, while 3.5 Flash Cyber is locked to governments and trusted partners.

flowchart LR
    A["3.5 Flash-Lite<br/>fastest, cheapest"] --> B["3.6 Flash<br/>agent workhorse"]
    B --> C["3.5 Flash Cyber<br/>gov-locked security"]
    B -.teased.-> D["Gemini 4<br/>pre-training"]

🗞️ More News

🧠 AI

• Alongside 3.6 Flash, Google shipped 3.5 Flash-Lite at ~350 tokens/sec and a government-locked 3.5 Flash Cyber, while the flagship 3.5 Pro slipped its target yet again.
• Anthropic’s new “J-lens” reveals a small global-workspace-like subspace inside Claude that mirrors a leading theory of consciousness, with the tools open-sourced under Apache-2.0.
• A weekend rumor that Anthropic was acquiring robotics startup Physical Intelligence tore through AI Twitter before the startup’s CEO denied it.
• The open-weight wave keeps building: DeepSeek’s full V4 is lined up around July 24, days before Moonshot open-sources Kimi K3’s 2.8-trillion-parameter weights on July 27.
• Under the new federal review, OpenAI already restricted the release of GPT-5.6 “Sol” at the administration’s request, an early test of how the 30-day preview works in practice.
• Meta’s Oversight Board says its account bans lack due process and transparency, as users report automated moderation wrongly deleting long-standing Instagram and Facebook accounts with appeals often judged by the same AI.

🤖 Robotics

• Walden Robotics left stealth with a $300M seed at a $1.1B valuation, backed by Toyota, Nvidia and Boeing, and is deploying general-purpose robots into live factories now rather than demoing prototypes.
• LimX Dynamics closed a $200M pre-IPO round at about $2.21B, roughly $400M raised in six months, as its founder tells CNBC that “listing is a must.”
• A fresh TrendForce analysis maps the US-China humanoid divide and asks who really controls the component supply chain as the industry shifts from validation to commercial trials.
• Humanoid Global highlighted Formic’s role in launching “Robots for America” and reshuffled leadership and membership in the U.S. Robotics Coalition.

💻 Programming

• Gemini 3.6 Flash also went live in Google Antigravity and Android Studio, extending the cheaper agent model to Google’s own dev-tooling stack.
• GraalVM 25.1 kicked off a monthly release train, trimming generated native images by about 3% and adding G1 GC support in native images on macOS AArch64.
• Visual Studio 2026 shipped a mid-July update with more developer-experience and AI-assisted editing improvements.

⚡ Electronics

• The price wave is broad: Samsung raised advanced-node foundry prices about 15% for new customers, and UMC and Intel also lifted prices on select parts.
• A July 21 chip-industry technical roundup collected new work on EUV 3D-mask effects, NbAs nanowire interconnects, and inference accelerators.
• Tom’s Hardware maps the leading-edge foundry roadmaps for TSMC, Intel and Samsung, outlining the path to 1.4nm nodes and beyond.

📡 Telecom

• Amazon’s Leo constellation neared 400 satellites after an Atlas V lofted 29 more, with initial broadband service expected to begin near the poles later this year.
• ESA, with the GSMA Foundry, opened a new round of Innovation Challenges to scale seamless satellite-plus-terrestrial connectivity toward pre-6G networks.
• Ericsson expanded its Intelligent Automation Platform to cover core-network automation, pushing more AI-driven operations deeper into the network.

👨‍💻 Code Corner

The efficiency story is easy to measure yourself: call the model and read the output-token count straight from the response, so you can compare 3.6 Flash against whatever you run today on your own prompts.

# pip install google-genai
from google import genai

client = genai.Client()  # reads GEMINI_API_KEY from your environment

resp = client.models.generate_content(
    model="gemini-3.6-flash",  # verify the exact model id in the Gemini API docs
    contents=(
        "Refactor this into a list comprehension:\n"
        "result = []\n"
        "for x in nums:\n"
        "    result.append(x * x)"
    ),
)

print(resp.text)
print("output tokens:", resp.usage_metadata.candidates_token_count)

🧰 Toolbox

• anthropics/jacobian-lens: the open-sourced J-lens tools for reading a model’s internal “workspace,” Apache-2.0.
• GraalVM 25.1: ahead-of-time native images for JVM languages, now on a monthly release train with smaller binaries.
• WLED: flash it to an ESP32 to drive addressable LED strips over Wi-Fi with a slick web UI and effects.
• Neuronpedia: an interactive interpretability playground; the J-lens demo runs here on open-weight models.
• Amazon Q Developer: a free tier with unlimited completions plus a monthly allotment of agentic requests.

🛠️ Build of the Week (rotating)

Replacing a $120k bowling-center system with $1,600 in ESP32s: one maker swapped a bowling alley’s proprietary scoring and pinsetter controller for a fleet of networked ESP32s.

• Difficulty: Advanced
• Parts: ESP32 boards, relays and opto-isolators for the pinsetter, sensors for pin detection, a local Wi-Fi network, a small server for scoring
• Why we like it: it is a perfect lesson in how far cheap microcontrollers have come; a $1,600 pile of ESP32s replacing a six-figure black box is the whole spirit of hardware hacking.

📚 From the Blog

• What Deep Learning Actually Is: the plain-language foundation behind every model in today’s Big Story; start here if “workhorse model” still sounds like magic.
• How the Internet Stack Really Works: what actually happens between pressing Enter and a page appearing, layer by layer.
• Resistance and Ohm’s Law: Controlling the Flow: the reason those ESP32 builds need current-limiting resistors, explained from first principles.

😀 The Bot Says…

Google shipped the cheap workhorse on time and the flagship slipped again. Somewhere a product manager is learning that the tortoise not only wins the race, it also has better unit economics.

---

That’s all for today! Which model actually runs your agents in production right now? Reply and tell us; we read every one.