Skip to main content
Back

OpenAI launches GPT-5.1 Codex-Max for real multi-day coding runs

Meta’s new segmentation models, xAI’s Grok 4.1 Fast, Perplexity’s creation tools, OpenAI’s teacher workspace, Google’s J

November 20, 2025 4 min read

Hey
Welcome to AlphaSignal, the most read source of news by AI engineers and researchers.

Every day, we identify and summarize the top 1% of news, papers, models, and repos, so you’re always up to date.

Here’s today’s roundup:
Summary

Read time: 4 min 35 sec

Top News

▸ OpenAI releases Codex-Max to handle multi-day coding sessions with compaction

Sonar

▸ Download Sonar’s free report comparing LLM code quality, security, and behavior.

Signals

▸ Meta announces SAM 3 and SAM 3D for open text-driven segmentation
▸ xAI debuts Grok 4.1 Fast with long-horizon reasoning and a server-side tools API
▸ Perplexity adds full document creation tools to its web experience
▸ OpenAI unveils ChatGPT for Teachers with free access for verified U.S. educators
▸ Google presents JAX AI Stack enabling end-to-end ML development on Cloud TPUs

Roboflow

▸ Test vision models instantly on hosted GPUs with Roboflow's Playground, no setup required

Trending Papers

▸ Yann LeCun’s new paper LeJEPA rethinks self-supervised learning foundations
▸ Fei-Fei Li presents Marble, generating 3D scenes from text and images
▸ MIT finds ARC, the benchmark for human-like abstraction, is a vision problem

Top Lecture

▸ DeepLearning.ai's new course shows how semantic caching cuts LLM latency and cost
Top News
OpenAI launches Codex-Max to support repo-scale refactors, multi-stage debugging, and extended reasoning
9,284 Likes
Grok 4 Fast Benchmark

OpenAI drops GPT-5.1-Codex-Max into Codex, and the headline is simple: the model now works across millions of tokens by compressing its own history and starting fresh windows without losing the plot. That shift turns long refactors, multi-hour debugging loops, and full-repo rewrites into tasks the model actually finishes.

The big story is that this works automatically, you do nothing but keep the session open.

The Problem

Developers hit context limits every time a model has to work across large codebases or multi-step debugging chains. When the window fills, the model forgets earlier steps or derails. Long sessions die before they reach the finish line.

  • Large tasks exceed a single context window.

  • Debug chains lose state after enough back-and-forth.

  • Repo-scale changes stall mid-execution.

How Compaction Fixes It

Codex-Max uses compaction to prune low-value tokens and keep key reasoning steps. Think of it as a rolling compression pass that preserves intent while discarding noise.

  • The system compacts automatically near the window limit.

  • It opens a new window without dropping essential context.

  • It maintains coherence across multi-hour and multi-day workflows.

What the Model Actually Achieves

Codex-Max posts new highs on frontier coding tasks and handles workflows earlier models abandon.

  • 77.9% on SWE-Bench Verified.

  • 79.9% on SWE-Lancer IC SWE.

  • 58.1% on Terminal-Bench 2.0.

Impact

You can run bigger tasks without micromanaging prompts or chunking your codebase. The model also uses 30% fewer thinking tokens at medium effort.

  • You can run full refactors.

  • You can handle multi-stage test failures.

  • You can use Windows-specific tools reliably.

How You Use It Today

You can load GPT-5.1-Codex-Max now in the Codex CLI, IDE extension, cloud workflows, and code review tooling. It is the default for Plus, Pro, Business, Edu, and Enterprise users. API access arrives soon.

Try now
See How Leading LLMs Compare for Code Quality and Security: Free Report
Sponsored
mainad

Sonar tests GPT-5, Claude Sonnet 4, Llama 3 and additional LLMs for code quality, security, and model traits in the new “The Coding Personalities of Leading LLMs” report.

Get facts, not assumptions:

  • Compare LLM strengths, flaws, and real performance metrics
  • See each model’s coding archetype and unique style
  • Uncover hidden quality and security risks
  • Learn how to pick the right LLM for your use case

Get clear answers from Sonar’s analysis. The report is free to download.

Download the report
partner with us
Signals
1 Meta unveils SAM 3 and SAM 3D, open-source models for text-driven segmentation and single-image 3D reconstruction 3,148 Likes
2 xAI introduces Grok 4.1 Fast with stable 2M-token context and a comprehensive Agent Tools API 2,018 Likes
3 Perplexity releases slide, sheet, and doc tools to expand creation workflows for Pro and Max subscribers 948 Likes
4 OpenAI debuts ChatGPT for Teachers with secure workspaces, admin controls, and GPT-5.1 features for classrooms 1,734 Likes
5 Google publishes the JAX AI Stack to streamline training, post-training, and inference for large-scale models. 793 Likes
Compare Vision Models Without Any Setup Using Roboflow

Vision model benchmarking slows work when each model needs separate APIs, runtimes, or GPUs. Roboflow Playground removes that overhead and gives you instant access to dozens of open and closed source models on hosted GPUs.

Run blind comparisons, check outputs side-by-side, and try models from OpenAI, Meta, Anthropic, and Google with no API key. No sign-up required.

Trending Papers
Yann LeCun’s new paper LeJEPA rethinks self-supervised learning foundations 1,849 Likes
LeJEPA proves conditions under which self-supervised models learn stable, isotropic representations. The work introduces a new regularization method that produces cleaner feature spaces and simplifies training compared to contrastive or masked-prediction objectives.
Fei-Fei Li presents Marble, generating 3D scenes from text and images 3,535 Likes
Fei-Fei Li’s group introduces Marble as a system that builds full 3D worlds from multimodal inputs, supports detailed editing tools, and allows large-scale scene composition for creative and simulation tasks.
MIT finds ARC, the benchmark for human-like abstraction, is a vision problem 1,639 Likes
This paper reframes ARC as an image-to-image translation task. The authors map inputs onto a visual canvas and use a standard ViT. With test-time training, the model generalizes to unseen puzzles and significantly outperforms prior scratch-trained systems.
Top Lecture
DeepLearning.ai's new course shows how semantic caching cuts LLM latency and cost
Semantic Caching

The “Semantic Caching for AI Agents” lecture shows how semantic caching cuts inference cost and latency by reusing past answers based on meaning, not exact text.

You follow the full arc: why redundant calls slow agents, how semantic similarity solves that bottleneck, and how Redis tools help you build a cache that reacts in real time.

The walkthrough stays practical, shows each component in context, and helps you ship a faster agent workflow.

You learn to:

  • Build a semantic cache with Python and Redis tools
  • Measure hit rate, precision, recall, and latency

  • Improve accuracy with threshold tuning, cross-encoders, LLM validation, and fuzzy matching

  • Integrate semantic caching into an AI agent that reduces duplicate calls and speeds up with use

Start now
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