Skip to main content
Back

Gemini 3 Flash: 50% cheaper, 3x faster Pro

On xAI's Grok Voice API, Stanford on AI in 2026, ChatGPT Apps, GitHub's Action Billing, Ai2's Open Video Model.

December 18, 2025 3 min read

Hey,

Your daily briefing is ready. You can finally take a break from the AI firehose.

Our algos spent the night splitting signal from noise and pulled the top news, models, papers, and repos.

Here’s the must-read:

Summary

Read time: 3 min 46 sec

Top News

▸ Google announces Gemini 3 Flash with faster, cheaper Pro-level reasoning

Coresignal

▸ Search B2B web data instantly using plain language with Coresignal

Top News

▸ xAI makes Grok’s mobile and Tesla voice system available to all via new API

Top Paper

▸ Stanford predicts 2026 shifts from AI hype to task-level outcome measurement

Signals

▸ OpenAI opens ChatGPT app submissions with review, publishing, and discovery tools
▸ GitHub postpones self-hosted Actions billing, but cuts hosted runner prices 39%
▸ Ai2 introduces SAGE, an open-source RL-trained agent for long-video reasoning
▸ Anthropic engineers improve terminal stability with an 85% flicker reduction
▸ ElevenLabs adds WhatsApp support for unified voice and chat agents
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Top News
Google launches Gemini 3 Flash, matching Pro-level reasoning while cutting inference cost by almost 50%
20,531 Likes
Grok 4 Fast Benchmark

Google expands the Gemini 3 family with Gemini 3 Flash, a model built for one clear goal: keep frontier-level reasoning fast enough for real systems.

Gemini 3 Pro proved Google could push reasoning, multimodality, and agents forward, but many production systems need speed first. Gemini 3 Flash closes that gap by keeping most of Pro’s intelligence while running far faster and cheaper.

The key idea is simple. The model adjusts how much internal reasoning it uses based on task difficulty. Easy requests finish quickly. Hard tasks still get deep reasoning.

This design makes Flash suitable for agents, coding loops, and multimodal apps that need rapid feedback.

Key features and performance

  • Scores 33.7% on Humanity’s Last Exam without tools

  • Reaches 90.4% on GPQA Diamond and 95.2% on AIME 2025

  • Achieves 78.0% on SWE-bench Verified, surpassing Gemini 3 Pro

  • Supports text, image, video, and audio with 1M-token context

  • Runs about 3× faster than Gemini 2.5 Pro at lower token usage

You can use Gemini 3 Flash through the Gemini API, Gemini CLI, Google Antigravity, or Vertex AI.

Build with flash
Query Public Web Data Without Writing Code
Sponsored
alpha_signal_image_1

Access to public web data often stops at technical barriers. Query languages, custom scripts, and internal tooling slow teams down before analysis even starts.

Coresignal introduces an AI-powered Data Assistant that removes those steps. You type a plain-language request and retrieve structured B2B data immediately. Search across millions of company, employee, and job records without schema knowledge or setup work.

The workflow stays simple:

  • Search using natural language
  • Access full Coresignal B2B datasets
  • Download results instantly
  • No technical support or data engineering required

The Data Assistant focuses on speed and accessibility. It replaces manual query building with direct intent-based search and immediate export.

Coresignal opens early access to this feature now. You can try the AI Data Assistant at no cost during the initial rollout.

Try it free
partner with us
Top News
xAI debuts a real-time Grok Voice Agent API with sub-second latency and flat per-minute pricing
5,299 Likes
Grok 4 Fast Benchmark

xAI opens the same voice system that powers Grok in mobile apps and Tesla vehicles to external developers through the Grok Voice Agent API.

Until now, low-latency voice agents required stitching speech models, streaming logic, and tool calls into fragile pipelines. Grok Voice addresses that gap by exposing a single real-time API that already runs at production scale.

The key insight comes from control over the full voice stack. xAI trains its own voice activity detection, tokenizer, and audio models, which cuts latency and simplifies coordination between speech, reasoning, and tools.

This design leads to a concrete result: Grok delivers first audio responses in under 1 second and scores 95% on Big Bench Audio.

You can use the API to build voice agents that talk, reason, and act in real time.

Key features and results

  • Real-time audio sessions with speech input, speech output, and tool calls

  • Flat pricing at $0.05 per minute of connection time

  • Built-in web search, X search, and custom function execution

  • Automatic language detection with mid-conversation language switching

  • Multiple expressive voices such as Ara, Eve, Leo, Rex, and Sal

To use it, you start a real-time session, select a voice, define tools, and send audio. The API follows the OpenAI Realtime specification and works with the xAI LiveKit plugin, so existing real-time stacks adapt with minimal changes.

Explore the api
Top Paper
Stanford says 2026 replaces AI hype with continuous, task-level metrics that track real outcomes
1,244 Likes
Grok 4 Fast Benchmark

Stanford HAI publishes its 2026 AI outlook after years of rapid model releases and uneven deployments.

The institute frames the next phase as a reset, where teams stop asking what AI can do and start measuring what it actually delivers. The standout shift is the move to task-level, monthly evaluation, which replaces broad claims with data you can track.

For years, teams asked whether models can write, diagnose, or reason. Stanford faculty argue that question no longer helps engineers shipping systems.

They instead push for answers tied to cost, accuracy, and real workflows, where many deployments still fail outside coding and call centers.

The insight comes from multiple domains converging at once. Economists, doctors, and lawyers all report the same pattern: AI works well in narrow tasks, and weakly elsewhere, unless teams measure outcomes precisely.

Key shifts you should track

  • Monthly AI dashboards use payroll and usage data to measure task-level productivity

  • Self-supervised medical models train without expert labels, cutting data costs

  • Legal AI benchmarks test multi-document reasoning, not single prompt writing

  • Evaluation focuses on accuracy, cost, and turnaround inside real workflows

Read more
Signals
1 OpenAI launches developer app submissions and introduces a ChatGPT directory for browsing published apps 1,457 Likes
2 GitHub pauses self-hosted billing changes, introduces lower hosted runner prices, and clarifies platform charges 3,112 Likes
3 Ai2 presents SAGE as an open-source RL video agent that outperforms baselines on long clips 1,132 Likes
4 Anthropic engineers rewrite Claude Code’s terminal renderer, reducing offscreen flicker by 85% 2,314 Likes
5 ElevenLabs introduces WhatsApp support, letting one agent handle voice and chat across all channels 1,472 Likes
At Alpha Signal, our mission is to build a sharp, engaged community focused on AI, machine learning, and cutting-edge language models, helping over 200,000 developers stay informed and ahead. We’re passionate about curating the best in AI, from top research and trending technical blogs to expert insights and tailored job opportunities. We keep you connected to the breakthroughs and discussions that matter, so you can stay in the loop without endless searching. We also work closely with partners who value the future of AI, including employers and advertisers who want to reach an audience as passionate about AI as we are.

Our partnerships are based on shared values of ethics, responsibility, and a commitment to building a better world through technology.Privacy is a priority at Alpha Signal. Our Privacy Policy clearly explains how we collect, store, and use your personal and non-personal information. By using our website, you accept these terms, which you can review on our website. This policy applies across all Alpha Signal pages, outlining your rights and how to contact us if you want to adjust the use of your information. We’re based in the United States. By using our site, you agree to be governed by U.S. laws.
Looking to promote your company, product, service, or event to 250,000+ AI developers?
Work with us