Google Releases Gemini 3.8 Flash With Smarter Reasoning at Same Price as 3.7

·1 views

Google launched Gemini 3.8 Flash on September 2, 2026, just three weeks after Gemini 3.7 Flash, positioning it as an incremental upgrade built on its predecessor. The newer model introduces smaller reasoning steps, self-verification of outputs, and iterative tool calls, which push its Artificial Analysis Intelligence Index score from 56 to 59. Pricing, context window, and output speed remain virtually identical between the two versions, with both set at $0.75 per million input tokens and $3.75 per million output tokens through December 31, 2026. However, the enhanced reasoning process means individual tasks may consume more tokens, raising the effective cost depending on workload. Google has confirmed that Gemini 3.7 Flash remains fully supported with no deprecation date announced, so migration is not urgent.

Read the full story at DEV Community

This is an AI-generated summary. ShortSingh links to the original source for the complete article.

Discussion (0)

Open LLM Gateway offers self-hosted, Apache 2.0 LLM control plane with no user cap

A developer has released Open LLM Gateway, a fully open-source, self-hostable LLM control plane licensed under Apache 2.0, after finding existing tools like LiteLLM imposed seat-based licensing restrictions beyond five users. The project was born out of a real operational problem: teams using multiple LLM providers across Google, Azure, and AWS had no unified way to track API key ownership, token usage, or per-product spending. The gateway supports proxying for OpenAI and Anthropic APIs, encrypted key management, cost and token analytics, and an admin portal with role-based access and optional Google login. It is deployable on Cloudflare Workers and Vercel, with the full admin UI included in the open-source release — not just the proxy layer. The project is positioned for organizations that need a self-hosted LLM gateway without enterprise licensing walls, though features like spend limits and guardrails are not yet available.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Developer builds browser-only PDF compressor that never uploads your files

A developer added a client-side PDF compression tool to ToolHub, designed so that no file data ever leaves the user's browser. Each PDF page is rendered as an image using a pdf.js pipeline, recompressed via a canvas-based quality slider, and reassembled into a new PDF using pdf-lib. Testing on a 4.7MB image-heavy PDF achieved a 72% size reduction, with output verified using pypdf to confirm image integrity. The approach comes with a noted tradeoff: because pages are converted to images, text in the compressed output is no longer selectable or searchable. The tool is recommended for image-heavy or scanned PDFs and openly discloses this limitation to users upfront.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Google's TimesFM 3.0 Offers Zero-Shot Time Series Forecasting Without Training

Google Research has released TimesFM 3.0, a pretrained foundation model that forecasts time series data without requiring any model training or hyperparameter tuning by developers. Built on a decoder-only transformer architecture similar to GPT, the model processes time series data in 32-step patches rather than individual points, making it efficient over long histories. It was pretrained on over a trillion time points spanning retail, finance, web traffic, and energy sectors, enabling zero-shot forecasting across unseen datasets. The model natively outputs nine quantile estimates at every forecast step, providing built-in prediction intervals without additional configuration. Developers can install and use TimesFM via a simple pip command, passing raw NumPy arrays to receive forecasts, including the latest improvements introduced in the August 2026 version 3.0 release.

0 comments Read more at DEV Community

ProgrammingDEV Community ·

Talent Is Pattern Recognition, Not Memory Capacity, Argues DEV Community Post

A widely discussed piece on DEV Community argues that memory and talent are fundamentally different cognitive abilities, with memory serving merely as storage while talent involves recognizing patterns, building abstractions, and transferring knowledge across domains. The article uses a graph analogy to explain that a truly skilled thinker has fewer but better-connected ideas, rather than many isolated facts. It draws a parallel to machine learning, noting that models — and people — who only memorize training data tend to fail when faced with unfamiliar problems, a phenomenon known as poor generalization. The author contends that real-world problem-solving demands the ability to map new inputs to underlying principles, not just retrieve cached answers. Ultimately, the piece concludes that the ability to discover connections between ideas, not the volume of information retained, is the stronger indicator of mastery.

0 comments Read more at DEV Community