Anthropic $40B, Claude Exodus to Codex, GPT Image 2 — AI Daily Apr 28

730 messages · 92 active members

730
messages
92
active members
@arielletolome, @jcartu, @mb29266
top contributors

Overview

The day opened with bombshell funding: Google committed up to $40B to Anthropic ($10B immediate, $30B milestone-based) at a $350B valuation, plus 5GW of TPU capacity over five years on top of Amazon's $25B commitment days earlier. Builders speculated about an October IPO at $800B+ while V4-Pro pricing ($0.145/$3.48 per M tokens) undercuts Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.7. Ironic counterweight: Claude went down mid-day with 404s and timeouts, accelerating an already-in-progress exodus to GPT-5.5 Codex. Reports of Opus burning 500K tokens to rename one variable, Antigravity Ultra users blowing through limits in 5 minutes, and dead Claude Max account rotation pushed serious builders toward Codex with xhigh reasoning, OpenCode with hot-swappable local Qwen 27B/35B MoE on RTX 6000s, and Hermes-based orchestrators. A widely-shared insight: agents perform better when prompted neutrally — emotional personas and screaming at Claude jack up error rates. On the creative side, OpenAI's GPT image 2 (via Kie) is producing 1:1 WhatsApp/Messenger/FB screenshots without templates, though Nano Banana Pro still wins for text-heavy UI. Seedance 2.0 + Kling 3 pipelines beat Veo on UGC realism, and operators running 40+ campaigns across USA/AU/EU confirmed audiences will buy 'pure slop' AI UGC. Rules-based media buying (2x spend kill, 3-day CPA scale) is outperforming AI agents for spend decisions.

Topics

Google committed up to $40B to Anthropic ($10B immediate, $30B milestone-based) at a $350B valuation, plus 5GW of TPU capacity over five years on top of a 3.5GW Broadcom deal starting 2027. Combined with Amazon's $25B commitment, Anthropic is positioned for a potential October IPO at $800B+. V4-Pro pricing undercuts Gemini 3.1 Pro, GPT-5.5, and Claude Opus 4.7.

Claude's website and API threw 404s and timeouts mid-day amid an already-rough patch — Opus reportedly burned 500K tokens to rename one variable, and Antigravity Ultra users hit limits in 5 minutes. Claude Max account rotation is dead (instant suspensions), pushing builders to GPT-5.5 Codex with xhigh reasoning. OpenAI doubled Codex Pro usage through May 31 to capture fleeing SWE clients.

@jcartu argued OpenCode is SOTA for hot-swapping models (including local Qwen 27B/35B MoE on RTX 6000s) and running 5-6 parallel agents on different models; @jasonakatiff defended Claude Code for predictability. @weslindquist's orchestrator-only Hermes setup gained converts replacing sub-agent micromanagement, with @rockdm going 'hermespilled' across Telegram channels. Consensus baseline: instruct agents to be 'critical and objective' to counter validation bias.

Builders testing OpenAI's new GPT image 2 on Kie report it dramatically beats Nano for natural-looking people and average UGC subjects, producing 1:1 WhatsApp, Messenger, and FB comment screenshots without templates. Nano Banana Pro still holds the crown for text-heavy UI work. @ShuPuffs walked through reference-driven prompting for app icons while @Wootbro declared design agencies 'dusted.'

@MiamiMega called Seedance 'fucking next level good realistic' compared to VEO; @rstmaur recommends 15-20s ads at 720p with audio and product references. Operators running 40+ campaigns across USA/AU/EU confirm 40+ audiences buy 'pure slop' AI UGC. @arielletolome's deterministic rules (kill at 2x spend, scale if under CPA for 3 days) outperform AI agents for media buying — running pause/scale logic through a Telegram bot for reliability over Meta's native rules.

Key Takeaways

  • Google's $40B + Amazon's $25B in Anthropic, plus 5GW TPU capacity, sets up an October IPO at potentially $800B+ valuation.
  • Claude Max account rotation is dead — pay API overages, diversify models, and use GPT-5.5 Codex with xhigh reasoning for SWE work.
  • Strip emotional language from agent prompts; instruct agents to be 'critical and objective' and use orchestrator-only Hermes setups instead of micromanaging sub-agents.
  • GPT image 2 on Kie produces 1:1 messaging app screenshots without templates; Nano Banana Pro still wins for text-heavy UI.
  • For media buying, simple deterministic rules (2x spend kill, 3-day CPA scale) outperform AI agents — and Seedance 2.0 beats Veo for UGC realism at 15-20s, 720p.

Hot Threads

@jcartustarted

OpenCode + local Qwen as a Claude Code alternative for parallel agentic coding

14 replies4 participants
@mb29266started

GPT image 2 on Kie vs Nano Banana for UGC and screenshots

20 replies6 participants
@jasonakatiffstarted

Emotionless agents outperform personality-driven ones — Claude Code breakdown

14 replies8 participants

Linked Items