Skip to main content
Back

Google's Deep Think hits 84.6% on ARC-AGI-2, beats Claude by 1,103

OpenAI's fastest coding model, Ex-GitHub CEO releases open CLI to track agent reasoning, Gemini Skills package released.

February 13, 2026 4 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:

Together with Warp
Sponsor Logo
Summary

Read time: 4 min 43 sec

Top News

▸ Google improves Deep Think, surpassing Claude and GPT-5.2 in reasoning

Warp

▸ Learn how to orchestrate cloud agents to scale agent productivity

Top News

▸ OpenAI rolls out low-latency Codex model for real-time coding

Anyscale

▸ Join the interactive session to deploy multi-GPU LLM training systems

Top Repo

▸ Ex-GitHub CEO releases open-source CLI to version AI agent sessions

Signals

▸ Karpathy-backed Simile raises $100M to simulate human populations with AI
▸ OpenAI retires GPT-4.1, GPT-4.1 mini, and o4-mini from ChatGPT
▸ MiniMax unveils M2.5, an open-source model scoring 80.2% on SWE-Bench
▸ Google adds official Skills package for Gemini API developers
▸ Anthropic reports 4% of GitHub commits from Claude Code, raises $30B
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 upgrades Gemini 3 Deep Think, sets Codeforces record, outpaces Claude Opus 4.6 by 1,103 points
43,582 Likes
Grok 4 Fast Benchmark

Google releases a major upgrade to Gemini 3 Deep Think, its specialized reasoning mode for science, research, and engineering. The system now posts 84.6% on ARC-AGI-2, verified by the ARC Prize Foundation, marking its strongest public reasoning result.

Traditional reasoning modes often follow one chain of thought and get stuck defending early mistakes. Deep Think instead explores several solution paths at once, compares them, and selects the most consistent result.

Google makes the update available in the Gemini app for AI Ultra subscribers and opens limited early API access.

Key features and results:

  • 84.6% on ARC-AGI-2 (abstract reasoning benchmark)

  • 48.4% on Humanity’s Last Exam without tools

  • 3455 Elo on Codeforces competitive programming

  • Gold-level results on 2025 Physics and Chemistry Olympiads

  • Sketch-to-3D pipeline that generates printable object files

How to use it:

  • Enable Deep Think mode in the Gemini app (AI Ultra)

Start exploring
Presented by Warp
Transform Your Eng Org With Agent-First Development [Report]
alpha_signal_image_1

Your team's cranking out code with AI agents. Fast at first. Then it plateaus hard.

Without orchestration, gains don't compound throughout the organization and leadership has no visibility, way to measure impact, or confirm it meets security standards. 

Introducing Oz: the orchestration platform for cloud agents.

Download the report to learn:

  • Why 75% of companies fail at building their own agentic systems 

  • How teams save hours per engineer per day by using agent automations

  • What makes 60%+ of PRs agent-generated actually achievable

Download now
partner with us
Top News
OpenAI releases GPT-5.3-Codex-Spark, a low-latency research preview model for real-time code generation
5,028 Likes
Grok 4 Fast Benchmark

You write code, wait for output, then adjust. That loop often slows you down. OpenAI introduces GPT-5.3-Codex-Spark to remove that delay and push real-time coding inside Codex.

It delivers over 1000 tokens per second on Cerebras Wafer Scale Engine 3 hardware and focuses on instant interaction instead of long autonomous runs.

OpenAI does more than speed up the model. It reduces full pipeline latency with a persistent WebSocket connection, cuts roundtrip overhead by 80%, lowers per-token overhead by 30%, and reduces time-to-first-token by 50%.

You see output almost immediately after you send a request.

Codex-Spark targets rapid edits, logic fixes, and interface tweaks. It uses a 128k context window and text-only input. On Terminal-Bench 2.0, it scores 58.4% accuracy, below GPT-5.3-Codex at 77.3%, but completes tasks much faster.

Key details:

  • 1000+ tokens per second inference speed

  • 128k context window

  • Separate preview rate limits

  • Access in Codex app, CLI, and VS Code

You use it through ChatGPT Pro in Codex tools for fast iteration workflows.

Try codex-spark
Presented by Anyscale
Scaling LLM Fine-Tuning with FSDP, DeepSpeed, and Ray
alpha_signal_image_2

Hitting memory limits when fine-tuning LLMs is a sign you're ready to scale.

This session covers distributed fine-tuning with FSDP, DeepSpeed, and Ray, the open-source framework behind Cursor, xAI, and Apple.

You'll walk through:

  • ZeRO stages and mixed precision config

  • Distributed training jobs with PyTorch

  • Checkpointing and failure recovery

Limited seats to keep it interactive.

Register now
partner with us
Top Repo
Ex-GitHub CEO Thomas Dohmke launches Entire and open-sources a CLI that records full AI agent sessions inside Git
4,285 Likes
Grok 4 Fast Benchmark

When AI agents write most of your code, Git shows what changed but not why. Thomas Dohmke, former CEO of GitHub, launches Entire to address that gap and releases an open-source CLI as the first step. The company raises $60M at a reported $300M valuation and focuses on one core problem: agent reasoning disappears after each session.

Entire CLI introduces Checkpoints, a Git-aware primitive that stores agent session context as structured metadata linked to commit SHAs. Instead of saving only diffs, you also store the transcript and decision trail behind the code.

What it captures:

  • Full session transcripts and prompts

  • Model responses and tool calls

  • Files modified in the session

  • Token usage per commit

  • Metadata stored in entire/checkpoints/v1

You install it with a curl script and run entire enable inside your repository. It supports Claude Code and Gemini CLI at launch. The repository remains unchanged; the CLI writes context as versioned data alongside your code history.

Install cli
Signals
1 Simile raises $100M to build an AI foundation model that simulates society using agents modeled on real humans 14,357 Likes
2 OpenAI deprecates GPT-5, GPT-4o, and GPT-4.1 in ChatGPT on Feb 13 at 10am PT 5,826 Likes
3 MiniMax introduces M2.5, an open-source RL-trained model built for coding, search, tool use, and office tasks 6,531 Likes
4 Google publishes official API Skills with CLI installation support for Gemini API and SDK workflows 1,836 Likes
5 Anthropic reaches $380B valuation as Claude Code drives $2.5B revenue 15,201 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