Good morning, humans.

If you build with AI, the weekend delivered a useful reminder: “smartest” is becoming a less interesting sales pitch than “good enough, fast enough, and cheap enough to run all day.” Google launched Gemini 3.7 Flash at half the original price of its predecessor, just as DeepSeek’s new V4 pricing took effect. That is the real lead in AI news today: model makers are no longer competing only on benchmark bragging rights. They are fighting over your production budget.

Elsewhere, Anthropic began watermarking Claude text globally, Grok 4.6 moved into GitHub Copilot, AMD went shopping for billions in fresh debt, and LG put an early-2027 date on its Nvidia-powered humanoid. In other words, the AI stack spent the weekend getting cheaper, easier to access, easier to trace, and considerably more expensive to finance. A perfectly normal industry, then.

AI news today: Google turns price into a feature

Google’s new Gemini 3.7 Flash is aimed squarely at coding, agents, and business workflows. Through the end of 2026, Google is charging $0.75 per million input tokens and $3.75 per million output tokens—half the original launch price of Gemini 3.6 Flash.

The company also reports meaningful gains on software-engineering tests: 65.3% on DeepSWE v1.1 versus 49.0% for 3.6 Flash, and 43.6% versus 34.4% on FrontierCode 1.1. Those are Google’s reported results, not proof that your deployment will improve by the same margin. Still, better claimed performance at a lower price changes the buying conversation. Teams can afford more retries, more tool calls, and longer-running agents without making every workflow feel like a miniature cloud migration.

Watch whether the introductory pricing survives into 2027. Temporary discounts are excellent at creating habits; invoices are where those habits become strategy.

DeepSeek makes V4 Pro official—and more expensive

DeepSeek V4 Pro is now generally available across the company’s app, web product, and API. Its new pricing took effect at 16:00 UTC on August 16, with separate peak and off-peak rates and a wide gap between the Pro and Flash tiers.

DeepSeek says Pro is designed for harder agentic work and has added multiple reasoning-effort settings. The important signal is not simply that another frontier model shipped. It is that DeepSeek is segmenting its lineup like a mature cloud vendor: inexpensive throughput for routine jobs, premium reasoning when a weak answer costs more than the tokens saved.

For buyers, the sensible move is routing, not loyalty. Benchmark your common tasks across Flash, Pro, and rivals, then escalate only when the cheaper model fails. We have seen the same pattern in our previous AI and tech brief: capability is rising quickly, but price-performance is now the metric with operational teeth.

Anthropic’s invisible watermark arrives worldwide

Anthropic has started applying an invisible statistical watermark to Claude-generated text globally as part of its compliance with the EU’s transparency code. The technique subtly changes low-stakes word choices without adding visible labels, extra tokens, or—according to Anthropic’s internal testing—a measurable quality penalty.

This is not a magic authorship detector. Anthropic says the mark works better on longer passages, is weaker in factual text and code, and cannot distinguish between “Claude wrote this” and “Claude heavily edited this.” A thorough rewrite can remove it. The company plans to offer a detection API, but has not yet given a launch date.

That nuance matters. Publishers and employers should treat watermark results as evidence of model involvement, not a verdict on who authored an idea. The technology is more defensible than guessing from suspiciously enthusiastic em dashes, but it still needs policy around it.

Grok 4.6 joins the Copilot model shelf

GitHub began a gradual rollout of Grok 4.6 in Copilot on August 14. The model is becoming available across Visual Studio Code, Visual Studio, Copilot CLI, the cloud agent, JetBrains, Xcode, Eclipse, and GitHub’s Copilot app. Business and Enterprise administrators must enable it; the policy is off by default.

xAI positions Grok 4.6 for long-running agents and multi-step coding, and reports a 500,000-token context window with prices starting at $2 per million input tokens and $6 per million output tokens. GitHub’s rollout matters because distribution can beat leaderboard position. Developers rarely switch tools to test one model; they will test it when it appears in the menu they already use.

For engineering leaders, this expands choice but also expands governance. Decide which models may see proprietary code, what usage-based billing is acceptable, and how outputs are reviewed before enabling every shiny option. Our guide to agentic AI replacing traditional app flows covers the broader shift: model selection is increasingly becoming infrastructure policy.

AMD reaches for up to $5 billion

AMD launched a four-part senior unsecured debt offering that could raise between $4 billion and $5 billion, according to Reuters. The notes mature in 2029, 2031, 2033, and 2036; AMD said proceeds may support general corporate purposes and debt repayment.

The company did not say the money is earmarked for a particular AI project, so treating the entire offering as “AI funding” would be an inference. The timing is nevertheless revealing. Competing in accelerators requires supply commitments, inventory, software investment, and customer support long before revenue arrives. The AI race is also a balance-sheet race, with more commas.

LG gives Nvidia’s robot stack a body

LG plans to unveil a two-legged humanoid robot in the first quarter of 2027 under an expanded Nvidia partnership, according to company details reported Friday. The machine will combine Nvidia’s Isaac GR00T reasoning models, Jetson Thor onboard compute, and Halos safety stack with LG-made actuators, sensors, and batteries.

LG also plans to test its wheeled CLOiD robot at a Tennessee washing-machine plant this year and build a reference AI factory using Nvidia Vera Rubin in 2027. That makes this more than a stage-demo announcement. The near-term test is whether the machines can perform reliably in a controlled factory; the humanoid reveal is the photogenic part.

One thing to remember

The competitive edge is moving away from owning one “best” model. It is becoming the ability to route work among models, govern their access, trace their output, and finance the infrastructure underneath them. This weekend supplied a new option at every layer—and another reason not to hard-code your strategy to a logo.