Codex Replaced Claude for Me… Here’s Why

{ "tldr": [ "OpenAI pulled ahead by simplifying around Codex — Riley and Ross argue Anthropic’s 50+ feature sprint in Q1 2026 created buzz, but OpenAI’s tighter “super app” strategy around Codex now feels more coherent for coding, docs, presentations, spreadsheets, and browser-based work in one place.", "The real platform battle is over the best coding model, not just the nicest UI — Their core thesis is blunt: the best coding model becomes the best general knowledge-work model, because spreadsheets, decks, designs, and code are all just files plus tools.", "OpenClaw mattered because it introduced proactive agents, not because it was polished — Ross’s favorite example is using OpenClaw to triage sponsor emails, research scams, negotiate early pricing, and send a morning report, turning an hour or two of daily creator admin into an automated workflow.", "Cursor was early and right on product, but lacked model economics — They credit Cursor for pioneering the modern agent UI and in-app browser, but say its biggest weakness is not owning a frontier model, which is why the reported XAI/SpaceX-style deal and $10B now / $60B later framing makes strategic sense.", "Prompting is really context management plus domain expertise — Ross’s point is that models are token predictors, so vague prompts produce vague slop; the edge now goes to 6-or-7-out-of-10 generalists who can articulate what good looks like and encode repeatable workflows as skills.", "Knowledge work is about to feel the same shock image generation already delivered — They cite contracts, bookkeeping, decks, reports, and even legal review as the next wave, with Stripe’s agent cards and agentic payments unlocking a future where AI doesn’t just recommend actions but actually buys and executes." ], "breakdown": "### Q1 belonged to Anthropic — until the mood shifted\n\nRiley opens by saying the first four months of 2026 felt like “10 years,” with Anthropic dominating Q1 through a nonstop stream of launches and the OpenClaw frenzy literally creating Mac Mini shortages in Toronto and Canada. Ross agrees Anthropic owned the narrative early, but both say the momentum started to turn after Opus 4.5 and a Karpathy tweet that made skeptical developers finally try these tools seriously.\n\n### Why Codex clicked: one app instead of five half-connected ones\n\nThe big contrast is product philosophy: Anthropic feels fragmented, while OpenAI feels increasingly unified. Riley’s complaint is specific and brutal — Claude Code, Co-Work, Design, Dispatch, Routines, Schedules, and Remote all feel like overlapping tools from different teams, whereas Codex is starting to look like the same capabilities expressed through one interface with different modes.\n\n### Cursor got there first, but frontier labs changed the game\n\nThey give Cursor real flowers here: Composer, the early agents tab, persistent in-app browser, and sandbox testing all showed where the product category was going. But Ross says Cursor’s weakness was always economics — when OpenAI and Anthropic can bundle thousands of dollars of compute into a $200 plan, a tool without its own model stack is fighting uphill, which makes the XAI partnership/acquisition logic obvious.\n\n### OpenClaw’s real breakthrough was agency, memory, and heartbeat\n\nRiley admits he first saw OpenClaw as hypey and scam-adjacent, but changed his mind once he saw the architecture: an agent with full computer access, files that evolve with use, natural-language skills, and especially the 15-minute “heartbeat” that lets it proactively message you. Ross’s sponsor-email workflow is the human proof point — the agent filters scams, researches companies, drafts responses, updates Notion, and hands him only the near-final conversations.\n\n### The future split: reactive super apps vs proactive personal agents\n\nOne of the sharpest distinctions in the conversation is Riley’s two-bucket framework: you either open a super app like Codex when you want to work with an AI, or you text your “claw” when you want something handled in the background. That leads to a bigger point: a truly valuable agent doesn’t just follow instructions, it surprises you with useful initiative — the way a great employee shows up with a report you didn’t know to ask for.\n\n### The hidden skill now is not prompting — it’s onboarding an agent\n\nRoss