Lead: Google makes the cheap agent model a serious production option
Google has released Gemini 3.7 Flash, a faster workhorse model for coding and agents, at an introductory $0.75 per million input tokens and $3.75 per million output tokens. Google says it beats Gemini 3.6 Flash across coding, web development, document reasoning and business automation, while also improving tool use and planning. The practical shift is less glamorous than a frontier-model launch but more consequential for teams shipping agents: a model that is good enough for production can now be priced for sustained use. Developers can read Google’s Gemini 3.7 Flash announcement.
What else matters
Anthropic adds invisible provenance to Claude text
Anthropic says future Claude models will use cryptographic text watermarking to estimate whether Claude generated a passage. The company says the method adds no visible characters, tokens or cost, and shouldn’t affect quality. The immediate driver is the EU AI Act’s transparency regime; the bigger implication is that provenance becomes a provider-level feature rather than a reader-facing label. Anthropic explains how Claude’s watermark works.
Google opens a compiler path to encrypted AI inference
Google has showcased HEIR, an open-source compiler and toolchain for homomorphic encryption. It converts pretrained models to accept encrypted inputs, allowing a server to compute without seeing the underlying data. The performance cost still matters, but healthcare, finance and other regulated users now have a clearer engineering route from privacy research to deployable experiments. Explore Google’s HEIR announcement.
OpenAI previews a speed-first API tier
OpenAI is previewing Ultrafast, a service tier that runs GPT-5.6 Sol at up to 14 times the speed. That targets latency-sensitive workloads where waiting, rather than raw capability, is the constraint. It also makes infrastructure and pricing choices part of model selection: teams will increasingly choose between a standard endpoint and a premium response-time guarantee. Read OpenAI’s Ultrafast preview.
The AI Act’s transparency clock is now operational
The European Commission says enforcement and new transparency requirements began on 2 August, including obligations for certain systems to disclose AI interaction and for generated or altered content to be marked. This is not a new announcement, but it is the compliance backdrop for Anthropic’s watermarking move and for every provider selling into Europe. Read the Commission’s enforcement notice.
Worth your time
For developers building privacy-sensitive systems, the HEIR project is the most useful technical follow-up: it turns a hard cryptography problem into a compiler and benchmarking workflow.
*Published: 15 August 2026*
