Skip to main content
Back

Anthropic Claude hits 52x code speedup, writes 80% of codebase

Claude hits 52x code speedup, OpenAI ships smarter memory, Codex builds iOS apps in browser, NVIDIA drops Nemotron Ultra

June 5, 2026 6 min read

Hey,

AI is now writing the code that makes AI better. Anthropic's numbers today aren't subtle: 52x code speedup, 80% of their own codebase authored by Claude, task horizons doubling every four months.

This isn't a benchmark flex. It's a structural shift. The loop is closing.

OpenAI quietly doubling ChatGPT memory capacity the same week feels less like a feature drop and more like everyone racing to build the scaffolding before the ceiling disappears.
Together with · partner_image
Author photo
Today's Author
Lior Alexander. Founder of AlphaSignal and former ML Engineer. Previously built ML systems at Iguazio, Guesty, Enphase, Mila.
Summary

Read time: 6 min 38 sec

Top News

▸ Anthropic says Claude now achieves 52x code speedups, outpacing humans 64% of the time

AssemblyAI

▸ Ship a voice agent this afternoon.

Top News

▸ OpenAI ships smarter ChatGPT memory with 2x capacity for Plus and Pro users

Viktor

▸ One AI employee. Finance, engineering, growth.

Top Repo

▸ OpenAI's Codex plugin lets you build and test iOS apps without leaving the browser

Signals

1. Google releases open 2.4B model for real-time on-device music generation
2. Presented by Tiger Data: CERN's Large Hadron Collider generates more data than most Al pipelines. Learn how they store and analyze it at scale
3. NVIDIA releases Cosmos 3, an open platform for building physical AI models
4. New diffusion model training trick cuts sampling steps while staying on the correct noise path
5. Google's Gemma 4 12B model gets a GGUF build you can run locally
6. NVIDIA releases Nemotron Ultra, the smartest US open-weights model at 400 tokens per second
Top News
Anthropic says Claude now achieves 52x code speedups, outpacing humans 64% of the time
26,026 Likes
Anthropic says Claude now achieves 52x code speedu

Anthropic just dropped a report that's hard to ignore. Their latest model, Mythos Preview, is showing signs that AI might soon be building the next version of itself, and the numbers back it up.

Here's what the data actually shows:

  • Anthropic runs the same test every release: give Claude code that trains a small AI model, ask it to speed it up. Claude Opus 4 hit a 3x speedup. Mythos Preview hit 52x.
  • Over 80% of code merged into Anthropic's production codebase is now written by Claude, not humans.
  • When shown research sessions where humans took a wrong turn, Mythos Preview suggested a better next step 64% of the time, up from 51% six months prior.
  • The length of tasks AI can handle solo is doubling every four months. Opus 3 managed 4-minute tasks. Opus 4.6 now handles 12-hour tasks independently.

Anthropic is clear: recursive self-improvement, meaning AI designing its own successor without humans, is not inevitable yet. But it could arrive sooner than most are ready for.

Read more
forward →
Presented by AssemblyAI
Ship a voice agent this afternoon.

STT, LLM, TTS, and orchestration used to mean four vendors, four dashboards, four invoices, and four places to look when something breaks.

AssemblyAI's Voice Agent API collapses the entire pipeline into a single WebSocket. One connection, one bill, one log stream. $4.50/hr covers everything.

Built on Universal-3 Pro. The most accurate speech-to-text model in the industry.

Get started now
partner with us →
Top News
OpenAI ships smarter ChatGPT memory with 2x capacity for Plus and Pro users
10,870 Likes
OpenAI ships smarter ChatGPT memory with 2x capaci

ChatGPT just got a smarter memory system, and it changes how you work with it every day.

The old system was basically a sticky note. You had to say "remember this" and it stored a fact. Simple, but it went stale fast. It depended heavily on you giving clear instructions, and saved memories could become outdated over time.

The new system, called Dreaming V3, works in the background automatically. It synthesizes useful context from past chats so ChatGPT can carry forward preferences, constraints, and patterns without waiting for you to say "remember this."

The numbers back it up:

  • Factual recall jumped to 82.8%, up from 41.5% in 2024.
  • Preference adherence hit 71.3%, and staying current over time reached 75.1%.
  • A new memory summary page lets you review, add, or update what ChatGPT knows about you.

Memory capacity is also doubling for Plus and Pro users in the US. Update your iOS or Android app to get it now.

Read more
forward →
Presented by Viktor
One AI employee. Finance, engineering, growth.

Last week one team had Viktor close the books, open 14 pull requests, draft a board update, and deploy three landing pages.

One AI employee across finance, engineering, and growth, working from inside Slack and Teams. Hire him on day zero and run leaner from there.

Get started
partner with us →
Top Repo
OpenAI's Codex plugin lets you build and test iOS apps without leaving the browser
1,349 Likes
OpenAI's Codex plugin lets you build and test iOS

iOS development has always meant constant context-switching: write code, rebuild, check Xcode, repeat. The new Build iOS Apps plugin for Codex closes that loop.

You can now view and test your iOS app in a browser, open SwiftUI previews (think live UI snapshots of your components), and hot reload edits, meaning your changes appear instantly without a full rebuild, all without leaving Codex.

Here is what the plugin unlocks:

  • View a live simulator stream directly in the Codex browser
  • Open SwiftUI previews to visually verify UI components as Codex edits them
  • Hot reload changes so Codex sees results instantly and keeps iterating
  • Debug, capture screenshots, and run UI automation without touching Xcode GUI

Install it with npx codex-marketplace add openai/plugins/plugins/build-ios-apps --plugin. Under the hood, serve-sim streams the simulator to the browser and SnapshotPreviews extracts SwiftUI previews, both open source.

Read more
forward →
Signals
1
Google releases open 2.4B model for real-time on-device music generation
547 Likes
2
CERN's Large Hadron Collider generates more data than most Al pipelines. Learn how they store and analyze it at scale
Presented by Tiger Data
3
NVIDIA releases Cosmos 3, an open platform for building physical AI models
8,902 Stars
4
New diffusion model training trick cuts sampling steps while staying on the correct noise path
104 Likes
5
Google's Gemma 4 12B model gets a GGUF build you can run locally
6,657 Downloads
6
NVIDIA releases Nemotron Ultra, the smartest US open-weights model at 400 tokens per second
530 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