Gemini 3.8 Flash Pricing, Copilot Exclusions, and Faster Speculative Decoding

This week’s usable AI changes for individual developers sit in three places: Google’s Gemini 3.8 Flash as a priced reasoning and coding model, GitHub Copilot content exclusions that now apply in the standalone app and CLI, and NVIDIA research that turns speculative decoding into design guidelines and training workflows. Each item is a model, product feature, or research result with a direct effect on API cost, agent context, or generation latency. [S8] [S15] [S3]

Gemini 3.8 Flash posts coding and reasoning prices

Google launched Gemini 3.8 Flash, an updated reasoning and coding model priced at $0.75 per million input tokens and $3.75 per million output tokens that improves performance on software engineering benchmarks like DeepSWE v1.1. Google also described performance gains across long-horizon software engineering, complex multi-step reasoning, and agentic workflows while maintaining a low price point. [S8]

For a developer or creator paying an API bill, the concrete shift is that those listed token rates now sit next to claimed gains on software-engineering and agentic workloads. That is a cost and capability change for people building coding agents or multi-step analytical workflows. The sources do not list a free tier, rate limits, or extra discounts beyond the stated $0.75 input and $3.75 output prices per million tokens. [S8]

Copilot content exclusions go GA in app and CLI

GitHub made content exclusion policies generally available across the Copilot app and Copilot CLI. Copilot now enforces enterprise and repository content exclusion policies in both interfaces, preventing AI agents from reading or processing designated files and code repositories as context. [S15]

Individual professionals on confidential or regulated projects can therefore use Copilot coding assistants while keeping designated sensitive files out of AI context. The same exclusion rules now apply whether the assistant runs in the standalone app or the command-line interface. The sources describe exclusions for designated files and repositories used as context; they do not document how those rules interact with every Copilot feature beyond the app and CLI. [S15]

Speculative decoding guidelines for faster generation

NVIDIA published research and implementation guidelines showing how speculative decoding co-design accelerates LLM inference without reducing output accuracy. It established five design guidelines for speculative decoding and released SPEED-Bench benchmarks alongside training workflows for draft mechanisms like EAGLE-3, DFlash, and DSpark in NVIDIA Model-Optimizer. [S3]

AI developers who ship interactive LLM applications can use those guidelines when selecting draft lengths and mechanisms. NVIDIA frames the payoff as increased token generation speed and lower latency for interactive LLM applications, with co-design meant to keep output accuracy intact. SPEED-Bench and the EAGLE-3, DFlash, and DSpark workflows are the artifacts to inspect; the research does not assign a single speedup percentage that applies to every model. [S3]

What to watch next

The three stories change spend, privacy boundaries, and latency for people who write and run software themselves. Gemini 3.8 Flash attaches $0.75 input and $3.75 output prices per million tokens to a coding-oriented reasoning model with reported DeepSWE v1.1 gains; Copilot exclusions now stop designated files from becoming agent context in the app and CLI; and NVIDIA’s five guidelines, SPEED-Bench, and Model-Optimizer workflows target faster speculative decoding. Treat the published prices, exclusion switches, and draft-mechanism docs as the documented change—the sources do not document additional hidden discounts, new Copilot products, or guaranteed latency numbers for every stack. [S8] [S15] [S3]