Google’s New “Workhorse” Model Just Made AI Agents Cheaper to Run

Part 1: AI Agents

Google’s New “Workhorse” Model Just Made AI Agents Cheaper to Run

What happened. Google shipped Gemini 3.7 Flash on Thursday — just three weeks after its predecessor, 3.6 Flash. It’s a “workhorse” model tuned specifically for software coding and autonomous agent workflows, and Google says it beat Anthropic’s Claude Sonnet 5 and OpenAI’s GPT-5.6 across nine benchmarks. The headline numbers: FrontierCode 1.1 jumped from 34.4% to 43.6%, DeepSWE from 49% to 65.3%, and AutomationBench nearly doubled from 17% to 30.4%. Just as important, it’s priced at half the cost of 3.6 Flash — $0.75 per million input tokens and $3.75 per million output tokens, through the end of the year.

Why it matters. This is the economics of agents changing in real time. Every point of improvement on DeepSWE and AutomationBench is a task you can hand to an agent instead of a senior engineer — and a halving of token cost is a halving of the bill for every agent loop you’re already running. Google’s product lead Tulsee Doshi framed it bluntly: the model “thinks more diligently, putting in more effort into multi-step planning and tool calls.” That’s the exact behavior that separates an agent that completes a workflow from one that stalls halfway through. If you’re building or budgeting for agentic workflows, this is a signal that the cost-per-task curve is still falling fast.

What to watch. Two things. First, the pricing is explicitly promotional (“through end of 2026”) — so lock in assumptions now, not later. Second, Google keeps shipping Flash models while the long-promised “Gemini 3.5 Pro” remains missing, and Ars Technica notes coding performance has been a weak spot amid a talent exodus. The model is live in the API, AI Studio, and Gemini Enterprise, but regular chatbot users stay on 3.6 Flash — so this is a developer/agent story, not a consumer one.

Source: Ars Technica · SiliconANGLE

Quick Hits

Cisco says agentic AI is fueling a “Networking Supercycle.” Cisco logged record $17.3B quarterly revenue (up 18% YoY) and $9.3B in AI infrastructure orders from hyperscalers, with CEO Chuck Robbins explicitly naming “the accelerating adoption of agentic AI” as the driver. The takeaway: agents aren’t just a software story — they’re now reshaping hardware, data-center, and networking budgets. Source

DeepSeek’s V4 Pro went live — and it’s a cybersecurity specialist, not a generalist. The model’s vendor-reported scores show explosive gains on CyberGym (52.7 → 83.3) and DSBench-Hard (31.1 → 67.2), but it struggles on general benchmarks, and independent verification is still pending. The real signal: DeepSeek is leaning into a specific, defensible niche while signaling a “significant” API price hike ahead. Source

Part 2: AI News

Databricks raises $5B at a $190B valuation. The data-and-AI platform closed the round led by Coatue, Blackstone, MGX, T. Rowe Price, and Sixth Street Growth, just six months after a $134B valuation — a signal that late-stage capital is still flooding into the picks-and-shovels layer of the AI stack. Source

China’s Z.ai targets OpenAI and Anthropic on coding with GLM-5.3. The Beijing-based firm (world’s first publicly listed LLM maker) says its new model closes the gap with Anthropic’s Fable 5 on coding, with open weights due in two weeks — part of a broader pattern of Chinese labs hitting near-frontier performance at lower cost. Source

Mindgard raises $30M to secure AI models and applications. The AI-cybersecurity specialist landed a Series A led by Album VC to meet surging demand as enterprises grapple with prompt-injection, data-poisoning, and model-theft threats — a reminder that every agent you deploy is a new attack surface. Source