← Back to Pulse PULSE. brief
Codex Is Winning... But Claude Has One HUGE Advantage

Codex Is Winning... But Claude Has One HUGE Advantage

Riley Brown24 min2026-08-10 ▶ Watch on YouTube
ℹ️
Partly verifiedA few specific details here couldn't be independently confirmed against the video. The overall summary is sound, but double-check exact numbers or names before you rely on them.
What this video is
⚡ a 25-minute video, readable in 60 seconds

This source tracks agentic coding tools including MuseCode, Codex, Cursor, and Buzz. The headline delta is Meta's release of MuseCode, a terminal coding agent powered by the Muse Spark 1.2 model, positioned to rival Claude Code and Codex; posted benchmarks place it between Claude Opus and GPT 5.6 Terra while its combined token cost runs about $5.50 versus roughly $30 for Opus and $35 for GPT 5.6 Sol. The source also reports two weeks of Codex desktop and in-app browser updates, Cursor's new Google Drive integration paired with DeepSeek V4 Flash, three new Chinese models (Kimi K3, DeepSeek V4 Flash, Qwen 3.8 Max), shifting developer sentiment away from Claude toward Codex, and the launch of Jack Dorsey's Buzz multi-agent Slack-style platform. For anyone tracking coding agents, this adds a cheaper Meta-built entrant plus a cluster of incremental platform changes and sentiment shifts worth weighing when picking a tool stack.

New: Meta released a new coding agent, MuseCode, positioned to rival Claude Code and Codex. [00:04]
Key takeaways
+ 56 more takeaways
  • New: Riley Brown says this makes Muse Spark five to six times cheaper than the top OpenAI and Anthropic models. [02:04]
  • New: Riley Brown used Codex to download the latest Meta MuseCode terminal coding agent. [02:32]
  • New: After downloading, MuseCode is run in the terminal by typing 'muse' then selecting trust and continue, or quit. [03:20]
  • New: Running the muse command for the first time triggers a workspace-trust prompt that must be accepted before it loads project-local config. [03:26]
  • New: MuseCode offers two login options, log in with browser or set an API key, and browser login signs in through a Meta account via auth.meta.com. [03:30]
  • New: Asked what model it is, the tool replies 'Muse code powered by Meta Muse Spark.' [03:50]
  • New: The host prompted MuseCode to build a Wii-bowling-style bowling simulator game and run it locally. [04:12]
  • New: By default the sandbox blocks persistent servers, so YOLO mode is needed for MuseCode to automatically run terminal commands. [04:43]
  • New: The host set MuseCode to always default to YOLO mode for every session using Codex, so it can act without permission prompts. [04:59]
  • New: A relaunched terminal session confirmed muse running with the --yolo flag via a red YOLO tag, and the finished bowling game was served locally at 127.0.0.1:8765. [05:28]
  • New: The second Agent Native update covers changes to Codex over the past two weeks, starting with the desktop app. [06:19]
  • From -> To: The Codex desktop app's side panel changed from a fixed project view to also showing a notifications/activity bar with the most recent completed chats, alongside the normal chat view. [06:33]
  • New: The Codex desktop app now shows the exact folder/directory on your computer where each chat is running. [06:55]
  • New: The creator calls the new in-app browser his favorite design update to the Codex desktop app in a long time. [07:12]
  • New: Using the in-app browser, the creator built and iteratively edited a website, a podcast site called Agent Native, by giving direct feedback. [07:20]
  • New: The creator warns the Whisperflow app will interfere with the Codex in-app browser unless dragged to the left or right side of the screen. [08:28]
  • New: Browser use is shown controlling the browser inside Codex, described as a new design of the in-app browser. [08:53]
  • New: Codex made updates to its Chrome extension, found under Plugins by searching for Chrome. [09:04]
  • New: The Plugins page lists extensions including Canva, ClickUp, Creative Production, Calorie Tracker, COROS, Calendly, Computer Use, Codex Security, Consensus, Context7, Clay, CalorieCam, and Close CRM. [09:12]
  • New: All chats done inside the Chrome extension now sync with the Codex app. [09:49]
  • New: To open a queued item directly in Codex, click the three dots, select open in app, and it opens automatically inside ChatGPT. [10:15]
  • Added: ChatGPT now has a toggle at the top to switch between Chat and Work. [10:33]
  • From -> To: The creator says ChatGPT desktop app's biggest competitor is no longer Claude Desktop but now Cursor. [11:31]
  • New: Cursor added a Google Workspace connection, cited by the creator as an early sign of becoming a full 'super app.' [11:31]
  • New: Cursor's CTO says many internal Cursor use cases are not coding at all, including research, data analysis, bug triage, and project management. [12:06]
  • New: Cursor has a Customize tab equivalent to Codex's Plugins tab, where a Google Drive integration can be set up and authenticated. [12:23]
  • New: Riley Brown demoed Cursor set up with DeepSeek V4 Flash, calling it basically free to use. [12:43]
  • New: He prompted the agent to build a professional Google Sheet on steps to train an AI model and send the link, completed in 44 seconds using the Google Drive integration. [13:04]
  • New: The Google Drive integration means Cursor can control Google Docs, Sheets, and Slides. [13:39]
  • New: Agent Native Update 4 covers three new Chinese models: Kimi K3, DeepSeek V4 Flash, and Qwen 3.8 Max. [13:57]
  • New: Kimi K3 is called the best of the three new models but also the most expensive, nearly as pricey as Sonnet for certain tasks. [14:09]
  • New: Dax from OpenCode says DeepSeek's planned price hike likely reflects traffic shaping from overload rather than losses, since current prices are reproducible on rented GPUs. [15:10]
  • New: Cline notes DeepSeek V4 Flash's low per-token cost could be misleading if it needs more turns to complete a task. [15:10]
  • New: Artificial Analysis reports DeepSeek completing the same benchmark tasks as Fable 5 at 105 times lower cost. [15:52]
  • New: The speaker predicts DeepSeek V4 Flash may get three to five times more expensive in the short run due to high traffic and demand. [16:04]
  • New: The speaker predicts that over the next two months this model or an equally good one will get significantly cheaper, and recommends trying these models for the next week or two. [16:17]
  • New: The speaker says V4 Flash is 20 to 100 times cheaper than Fable depending on the task. [16:45]
  • New: AI researcher John Ennis said Opus 5 Extra High is 'basically trash,' wastes its thinking budget thrashing on pointless or harmful actions, and that he is switching away from Claude. [17:33]
  • From -> To: Hamel Husain said developer consensus has shifted from Claude to Codex, citing a better harness, better Codex desktop app, better pricing, fewer refusals, and freer subscription use across tools. [18:02]
  • New: The speaker tweeted that 'Anthropic is pulling back the slingshot and will try to go on another run soon,' attributing Anthropic's stumble to launching too many confusingly-named products (Claude Cowork/Dispatch, Claude Code/Claude Remote, a law product, Claude Design) and mostly similar-feeling model releases since Claude 4.6, while still favoring Anthropic for frontend work. [18:38]
  • New: The speaker says Opus and Fable outperform OpenAI models for knowledge work documents (spreadsheets, docs, presentations) and frontend design. [20:00]
  • New: For everything else besides design, the speaker says GPT 5.6 Soul is basically as good as Fable. [20:09]
  • New: The speaker says Claude created a frontend design in one prompt that looks great, calling it the reason he keeps using the Claude desktop app. [20:14]
  • From -> To: The speaker believes the first half of the year was the 'open claw personal agent' movement, and the industry is now moving into a 'team of AI agents' movement. [20:54]
  • New: Buzz is Jack Dorsey's new platform, described as a Slack clone built for use with AI agents. [21:03]
  • New: In Buzz, a user can at-mention both Claude Code and Codex in a channel and both agents respond and work together. [21:28]
  • New: Buzz's Agents tab shows a full team of AI agents, letting users add Cursor or Devin, set a default model, or create a custom agent, for example 'Cursor with Kimmy' using the Kimi K2 model as a content agent. [21:58]
  • New: The Slack-like agent system can add any agent to any channel and can even create agents. [23:04]
  • New: Codex is shown actively working while 'Cursor with Kimi' requests approval, illustrating agents working as a team. [23:12]
  • New: 'Cursor with Kimi' (CW) responded, 'I handled it myself.' [23:24]
  • New: CW confirmed it joined all seven channels: Content, Vishal, Coding, General, research, Management, and Notes. [23:31]
  • From -> To: The presenter compares this to how OpenClaw was hard to use early on and is now basically the engine behind ChatGPT's agent capabilities. [23:45]
  • New: The presenter says personal AI agents accessible via phone have advanced far, and expects team-based AI agents to take shape over the next four to six months. [23:55]
  • New: The presenter recommends that if you work at a large company, you should build a team of agents or an agent accessible to your whole team. [24:13]
  • New: The presenter predicts more Slack-like agent platforms will emerge that make configuring agents easy. [24:28]
  • Why it matters: trackers now have a lower-cost Meta-built alternative to Claude Code and Codex, plus two weeks of Codex, Cursor, and Chinese-model updates and a new Buzz multi-agent platform to weigh when choosing a coding-agent stack.
Shown on screen — grab and go
PROMPTRequest for one-sentence feedback from bots
Harry @Sonnet Claude Code Codex Please all of you provide 1 sentence feedback
reconstructed from crushed-space OCR of the rileybrown 1:48PM message in the multi-agent Slack-style chat at 00:43 shown at 0:43
QUOTEHarry AI agent's feedback on the 25-creator sweep
My feedback: the 25-creator sweep is a strong timely-signal layer, but it should feed one strict editorial filter rather than become the whole idea engine, because your first-party build failures, wins, and audience questions are still the differentiated material nobody else can copy.
transcribed from the Harry@managedbyyou 1:49PM message shown on screen at 00:43, spaces restored from crushed OCR shown at 0:43
PROMPTInstruction to add Cursor-Kimi to all channels
Codex please add @Cursor with Kimi to this channel and all other channels
read from rileybrown's 11:10PM message in the Buzz app at 23:28, spaces restored from crushed OCR shown at 23:28
QUOTECline's tweet on DeepSeek cost-per-task pricing
While DeepSeek V4-Flash is significantly cheaper on price per token, this can be misleading if the overall cost per task ends up being higher due to more turns being made. However, @ArtificialAnlys[...] reports DeepSeek completing the same benchmark tasks as Fable at 105x lower cost. Cost per task, not cost per token.
merged from two overlapping OCR reads of the @cline post at 14:32 and 15:56; the analyst handle is a partial OCR read shown at 15:56
How this brief was shaped: What Changed · confidence Medium

The opening explicitly frames the video as covering the last two weeks of AI agent news across named platforms (MuseCode, DeepSeek, Kimmy, Quen, Codex, Cursor, Google, Buzz) with recurring phrasing like changes to Codex over the past two weeks, and OCR shows install and terminal setup screens matching the hands on walkthrough of what is new. Some segments carry creator opinion such as Google might be out of the AI race, but the organizing spine is a periodic update roundup, not a debated thesis.

The lens sets this brief's structure, never its facts — every claim is held to the same citation and fact-check standard.

Jump to a moment
0:00Intro0:54#1 Meta Muse Code4:441a. Yolo Mode on Meta Muse Code6:19#2 Codex Updates6:302a. New Sidebar Notifications7:122b. In-App Browser9:042c. Chrome Extension10:332d. Chat & Work Toggle11:28#3 Cursor Becomes a Super App13:52#4 New Chinese Models16:57#5 Anthropic's Slump19:285a. What Keeps Me on Anthropic20:48#6 Buzz + Agent Teams
Their links, sorted & clickable
🔗 Other links1Listen to this episode on Spotifyopen.spotify.com
← Back to Pulse Dashboard
Was this brief useful?