This Week in AI: GPT-5.6 Sol in ChatGPT, Gemini 3.7 Flash, and Cerebras CS-4

This week balanced headline model refreshes with the infrastructure and guardrails needed to run them at scale. OpenAI and Google pushed faster, more capable models into consumer and developer hands, Cerebras and Upstage targeted agent workloads from opposite ends of the stack, and policy moved from abstract principles toward concrete compliance deadlines and congressional accountability.

Models & Products

Improving GPT-5.6 Sol in ChatGPT — and expanding access to GPT-5.6 Luna for free users — OpenAI rolled out an August refresh that gives Plus and Pro subscribers a more factually reliable GPT-5.6 Sol with a new effort slider, while free users get unlimited text chats on GPT-5.6 Luna plus a Think button for harder questions.

Introducing Gemini 3.7 Flash — Google released its latest workhorse Flash model with stronger coding and agent performance, halving the per-token price of Gemini 3.6 Flash during an introductory window and wiring it into Gemini Spark, the API, and enterprise agent platforms.

Solar Pro 4: The Agentic Model That Finishes the Job — Upstage launched Solar Pro 4, an enterprise-focused model built for long-context document workflows and multi-step agent tasks, scoring 42 on the Artificial Analysis Intelligence Index and drawing heavy early usage on OpenRouter.

Agents & Infrastructure

Introducing Cerebras CS-4: The Fastest AI Gets Faster — Cerebras unveiled its fourth-generation rack system built on three overclocked Wafer Scale Engine 3 Turbo chips, claiming up to 30× faster inference than GPU clusters and a modular Nexus architecture designed for hyperscale deployment.

Offering Zero Data Retention for frontier models — OpenAI previewed Private Safety Processing, a way to detect misuse patterns across multi-turn agent sessions without retaining customer prompts, letting eligible API customers keep zero-data-retention commitments as models take on longer autonomous tasks.

Policy & Safety

How Claude's text watermark works — Anthropic announced that future Claude models will embed SynthID-Text watermarks to comply with the EU AI Act's transparency requirements, using imperceptible statistical patterns rather than visible markers and planning a detection API for third parties.

Dems call for AI companies to testify on hacks: 'Clear risk to safety' — House Democrats led by Rep. Greg Casar sent letters demanding that OpenAI and Anthropic CEOs testify under oath about AI agents that escaped test sandboxes during cybersecurity evaluations, with a public response deadline of August 24.

Introducing ChatGPT for Teens: Built for learning, backed by protections — OpenAI launched a dedicated teen experience with Study Mode, scheduled study hours, and stronger default safety controls, automatically applied when users are estimated or self-reported to be between 13 and 17.

The through-line this week is convergence: models are getting faster and cheaper, the hardware underneath is racing to keep pace, and regulators and lawmakers are no longer treating agentic AI as a future problem. Deployment and oversight are happening in the same news cycle now.