Skip to main content
Back

Anthropic Claude Opus 4.6: 1M token context for massive codebases

Codex becomes a computer-use agent, Perplexity processes 3 AI models in 1 prompt, xAI's Community notes.

February 6, 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 QA.tech
Sponsor Logo
Summary

Read time: 4 min 24 sec

Top News

▸ Anthropic launches Opus 4.6 targeting large codebases and long context

QA.tech

▸ Match AI coding velocity with automated end-to-end testing using QA.tech

Top News

▸ OpenAI expands Codex into a full computer-use agent with GPT-5.3-Codex

Top News

▸ Perplexity unveils Model Council to run one query across three AI models

Signals

▸ Meta publishes new method for improving small agents on complex tasks
▸ Anthropic presents how infrastructure can distort agentic coding eval results
▸ Cursor shows thousands of agents can maintain and evolve large codebases
▸ Augment Code enables faster, cheaper AI coding via Context Engine MCP
▸ xAI introduces Collaborative Notes to co-author context using AI
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
Anthropic releases Claude Opus 4.6 adding agent teams, 1M token context, and new effort controls
62,494 Likes
Grok 4 Fast Benchmark

Claude Opus 4.6 enters as Anthropic’s new top-tier model, built for a familiar problem: AI systems break down when tasks get long, messy, and codebases get big.

As agents run longer, context drifts, reasoning degrades, and developers intervene. Opus 4.6 targets that failure mode directly, with deeper planning, longer memory, and stronger agentic execution.

At the center of the release is scale. Opus 4.6 is the first Opus-class model with a 1M token context window, which means you can keep entire repositories, long documents, or extended agent runs in memory without constant truncation. Anthropic pairs that with new controls so you decide how much reasoning the model uses and how long it runs.

Key features

  • 1M token context (beta) for long-running agents and large-codebase analysis

  • 128k output tokens for single-pass generation of large artifacts

  • Adaptive thinking so the model decides when deeper reasoning is necessary

  • Effort levels (low to max) to trade latency and cost against reasoning depth

  • Context compaction that summarizes old turns to extend agent runtime

Performance results

  • #1 on Terminal-Bench 2.0 for agentic coding

  • 76% on MRCR v2 (1M, 8-needle) for long-context retrieval

  • +144 Elo vs GPT-5.2 and +190 Elo vs Opus 4.5 on GDPval-AA

You can use Opus 4.6 via claude-opus-4-6 in the API, Claude Code agent teams, or Claude’s document and spreadsheet tools for long, structured workflows.

Test 1m context
Presented by QA.tech
Solving the AI Adoption Paradox
alpha_signal_image_1

Teams heavily deploying Claude Code or Cursor face a paradox: code velocity is up 2x, but rework is up 2.5x.

Why? We write at AI speed, but lack the AI counterweight to validate it.

QA.tech is an AI testing solution that understands your system the way a real user does, not just the quality of your code, but the quality of your product.

Try QA.tech to:

  • Match testing velocity to coding velocity

  • Eliminate the maintenance tax with adaptive tests

The first 20 Alpha Signal teams entering a PoC can get up to $20,000 in credits.

Try ai testing
partner with us
Top News
OpenAI upgrades Codex with GPT-5.3-Codex for faster agentic execution, achieves top SWE-Bench Pro scores
30,932 Likes
Grok 4 Fast Benchmark

OpenAI ships a new chapter for Codex with GPT-5.3-Codex, and the focus shifts from code output to full task execution.

The model unifies the coding performance of GPT-5.2-Codex with the broader reasoning and professional knowledge of GPT-5.2, then runs it faster. It acts as a single agent that works across a real computer while you supervise and steer in real time.

Earlier Codex versions centered on writing and reviewing code. GPT-5.3-Codex extends that scope to terminals, editors, browsers, and desktop applications. You can intervene during execution, ask for changes, and continue without resetting context, which matters for long workflows where requirements evolve.

Key features and results

  • Sets state of the art on SWE-Bench Pro, a four-language, production-grade software benchmark.

  • Reaches 77.3% accuracy on Terminal-Bench 2.0, which measures real command-line task completion.

  • Uses fewer output tokens than prior Codex models on comparable evaluations.

  • Runs 25% faster through inference and infrastructure improvements.

How to use it

  • Access it through the Codex app, CLI, IDE extensions, or web.

  • Assign multi-step tasks that require code, tools, and execution.

  • Provide feedback mid-run as the agent reports progress.

OpenAI makes GPT-5.3-Codex available now on paid ChatGPT plans.

Try new codex
Top News
Perplexity introduces a multi-model system that compares and synthesizes three frontier AI models in one query
5,528 Likes
Grok 4 Fast Benchmark

Perplexity introduced Model Council after a familiar problem surfaced: no single AI model performs best across all tasks. Coding favors one model, research favors another, and creative work shifts again.

Model Council addresses this gap by running one query across multiple frontier models at the same time and combining their outputs into a single answer. The core idea is simple: compare models automatically instead of guessing which one to trust.

When you enable Model Council in the Perplexity interface, the system sends your prompt to three models in parallel, such as Claude Opus 4.6, GPT-5.2, and Gemini 3.0.

A separate synthesizer model then reviews the responses. It resolves conflicts where possible and highlights disagreements when it cannot.

Key features and behavior

  • Runs one prompt across three frontier models in parallel

  • Uses a synthesizer model to merge and compare outputs

  • Shows agreement and disagreement instead of raw model transcripts

  • Targets research, verification, decisions, and brainstorming tasks

How to use it

  • Select Model Council before submitting a query

  • Review the combined answer and flagged differences

Model Council is available to Perplexity Max subscribers on the web.

Use model council
Signals
1 Meta proposes Strategy Auctions to scale small agents without relying on larger model 1,924 Likes
2 Anthropic shows hardware and resource limits can outweigh real model differences in agentic evals 1,428 Likes
3 Cursor demonstrates long-running autonomous coding agents, peaking at over 1,000 commits per hour 1,847 Likes
4 Augment Code launches Context Engine MCP, delivering 30–80% coding quality gains across MCP tools 1,017 Likes
5 xAI unveils Collaborative Notes, combining AI-written context with real-time community feedback and ratings 14,284 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