SpaceXAI drops Grok 4.5, GPT-Live-1 goes full duplex - Edition 014

Read time: 3 mins • 100% Jargon-Free

Edition Summary

Today: SpaceXAI drops Grok 4.5 as an Opus class coder priced at $2 per million input, OpenAI's GPT-Live-1 lets ChatGPT listen and speak at once, and Google Search Console AI reports just went wider so you can see how AI Overviews treat your pages.

Oh So AI Issue #014 | Friday, 10 July 2026

💬 Daily Quote

The task you keep pushing to tomorrow is usually the one that would move everything today.

📌 TL;DR

Today: SpaceXAI drops Grok 4.5 as an Opus class coder priced at $2 per million input, OpenAI’s GPT-Live-1 lets ChatGPT listen and speak at once, and Google Search Console AI reports just went wider so you can see how AI Overviews treat your pages.

🏆 LLM Leaderboard #AI #Models

Context. The most popular and smartest models this week.

1. Claude Opus 4.8 Score 61.4

2. GPT-5.5 Score 60.2

3. GLM-5.2 Score 51.0

4. DeepSeek V4 Flash #1 Popularity

5. Tencent Hy3 Preview #2 Popularity

6. Claude Opus 4.7 #3 Popularity

7. Nemotron 3 Super #4 Popularity

8. DeepSeek V3.2 #5 Popularity

Two different charts, one story. Opus 4.8 is still the smartest model on the market but sits outside the top five by volume. The everyday traffic has moved to DeepSeek V4 Flash and Tencent Hy3, because cheap and fast wins the volume war. Chinese open weights now move close to half the tokens on OpenRouter. Watch Grok 4.5, which launched Wednesday and will likely enter both charts next week.

Source. OpenRouter Rankings + Benchmarks

📰 Oh, So AI did this #AI #News

1. Grok 4.5 goes public as an Opus class coder for a fraction of the tokens. #AI #News

SpaceXAI shipped Grok 4.5 to grok.com and Cursor on Wednesday, priced at $2 per million input and $6 per million output. On SWE-Bench Pro, it solves tasks in about 16,000 tokens against Opus 4.8’s 67,000, per TechCrunch. That is a 4x gap on real coding work.

Why you care: If your team pays for coding agents, you can plug Grok 4.5 into Cursor today and cut your token bill without dropping to a weaker model.

Read: TechCrunch

2. GPT-Live-1 lets ChatGPT listen and speak in the same breath. #AI #News

OpenAI rolled out GPT-Live-1 to paid users and GPT-Live-1 mini to free users on Wednesday. It runs full duplex, so it can nod, interrupt, and hold pauses like a person, per MacRumors. Streamed text still shows up beside the voice.

Why you care: Your voice sessions with ChatGPT stop feeling like walkie-talkie turns and start feeling like real chats, which changes how you brainstorm while walking, driving, or cooking.

Read: MacRumors

3. Anthropic locks in 3.5 gigawatts of Google and Broadcom compute. #AI #News

The new multi-year deal brings on multi-gigawatt TPU capacity starting in 2027, roughly 4.5x the current Anthropic fleet, per Yahoo Finance. Claude run rate is now $30 billion, up from $9 billion at the end of 2025. Over 1,000 customers spend more than $1 million a year on Claude.

Why you care: Anthropic just guaranteed the compute for years of Claude upgrades, which means the tool you build workflows on today is not going away or getting slower.

Read: Anthropic

4. GitHub Copilot adds Kimi K2.7 Code as its first open-weight option. #AI #News

Kimi K2.7 Code is now generally available in Copilot Business and Enterprise, per the GitHub Changelog. It sits in the model picker beside GPT and Claude, at a lower per-token cost.

Why you care: If your company has a Copilot licence, you can route cheaper work like renaming files or writing tests to a model that costs less, and keep expensive models for hard problems.

Read: GitHub Changelog

5. Copilot Vision goes fully live inside GitHub. #AI #News

As of July 1, you can drop images and PDFs into Copilot chat and it will reason about them beside your code, per GitHub. Design mockups, terminal screenshots, and PDF specs go straight into the same conversation.

Why you care: You can paste a Figma screenshot into your IDE and ask Copilot to build the layout, without leaving the editor.

Read: GitHub Changelog

6. Google Search Console opens AI performance reports to more sites. #AI #News

The reports now show how your pages perform inside AI Overviews and AI Mode, with impressions by page, country, and device, per Search Engine Land. Click data is still missing.

Why you care: You can finally see if your pages are being read inside AI answers, even when the click never lands on your site, so you know what content the AI actually pulls.

Read: Search Engine Land

Enjoying this breakdown?

Join 4,200+ others getting these simple translations every Tuesday and Friday.

🪄 Oh, So AI can do that?! #AI #Tools

1. GPT-Live-1 API for developers. #AI #Tools

OpenAI opened the same full-duplex voice model it just shipped in ChatGPT to app builders. Text streams beside audio in one call, so your app can show captions while it talks.

Example. A founder can drop this into their app so a demo assistant can talk over its own answer the moment the user cuts in.

Why you care: You can add a voice layer to your product without wiring up a separate speech pipeline.

Try it: OpenAI

2. Grok 4.5 inside Cursor. #AI #Tools

SpaceXAI’s new Opus class coder is available in Cursor on every plan from Wednesday, at $2 per million input and $6 per million output.

Example. A solo dev can point Grok 4.5 at a full repo and refactor a module for a fraction of the Opus token cost.

Why you care: Your Cursor bill drops for the same quality of code review or refactor.

Try it: SpaceXAI

3. Kimi K2.7 Code inside GitHub Copilot. #AI #Tools

GitHub added the open-weight Kimi K2.7 Code as a picker option in Copilot Business and Enterprise. First open-weight model in the picker.

Example. A dev team can send boilerplate and test writing to Kimi and keep design work on Claude or GPT.

Why you care: You get to split cheap and expensive work across models inside the same tool.

Try it: GitHub Changelog

4. Copilot CLI runs inside GitHub Actions. #AI #Tools

You can now run the Copilot CLI in Actions using the built-in GITHUB_TOKEN. No more long-lived personal access tokens for agent jobs.

Example. A team can spin up a nightly Copilot job that reviews PRs without storing a PAT anywhere.

Why you care: One less credential to rotate and one less security risk to babysit.

Try it: GitHub Changelog

5. Codex agent provider inside JetBrains IDEs. #AI #Tools

GitHub added Codex as an agent provider in JetBrains, in public preview. It joins Claude and the built-in Copilot agent inside the same IDE.

Example. IntelliJ users can now switch between Codex, Claude, and Copilot without leaving the editor.

Why you care: Pick the agent that matches the task, inside the same window.

Try it: GitHub Changelog

⚡ Oh, So I can do this #SEO #WebDev

1. Google Search Console AI performance reports go wider. #SEO

More sites now get impression data for AI Overviews and AI Mode, per Search Engine Land. Click data is still absent, so treat it as a share-of-voice metric for AI answers.

Why you care: You can see which pages are winning inside AI answers, which is where a growing share of your search traffic now lives.

Read: Search Engine Land

2. Google Search Profiles now report analytics. #SEO

The new profile pages have their own performance data inside Search Console, and Google is testing a platform property tooltip next, per Search Engine Roundtable.

Why you care: Personal brand pages that live inside Google Search now show real data, so you can measure them like any other page.

Read: Search Engine Roundtable

3. Next.js 16.2 ships a security backport. #WebDev

Vercel patched multiple advisories across Server Components, App Router, Middleware, and image handling. It also fixes cache tag handling, Turbopack quirks, and FormData behaviour.

Why you care: A one-line version bump closes several CVEs. Schedule the patch before the weekend.

Read: Releasebot

4. Vercel Flags SDK gets a 10x speedup. #WebDev

The July Vercel drop also trims Workflow chat payloads to shrink logs, and exposes token usage and finish reason on workflow stream calls.

Why you care: A/B tests and feature gates now add less overhead to your page load.

Read: Vercel Changelog

🎣 Top Hook Ideas

Hook 1. Grok 4.5 solves the same coding task Opus does using 4x fewer tokens. Everyone is quiet about what that means for AI budgets in Q3. #AI

Hook 2. ChatGPT can now nod, pause, and cut in mid-sentence like a real friend. Voice work will not sound the same again. #AI

Hook 3. Anthropic just bought enough compute to run 4.5x more Claude by 2027. That is the real signal hidden inside this week’s AI news. #AI

Hook 4. GitHub Copilot now runs an open-weight Chinese model beside Claude. This is the moment where “the model” stops being one model. #Tools

Hook 5. Google just told you which of your pages get pulled into AI answers. Most brands still have no idea what theirs are. #SEO

WhatsApp Community Message

🪄 Oh, So AI can do that?!

Tool. GPT-Live-1 by OpenAI

What it does. Talks with you in real time. Listens while it speaks. Cuts in when you cut in.

Why it’s cool. It stops feeling like a voice command and starts feeling like a call. Great for hands-free brainstorming while you drive, walk, or fold laundry.

Try it. https://chatgpt.com

Did 'Oh, So AI' help you understand something new today?

Forward this to a friend who is overwhelmed by AI.

Curated with ❤️ ByHarshal

Creative Director orchestrating AI workflows for founders' teams. Writing about productivity, design, and AI systems.

You made it to the end! 🎉

Subscribers got this days ago. Drop your email or join our WhatsApp community to get the next one first.

Previous Edition