Opus 4.8 Drops, Claude Code Bug, Seedance vs Sora 2 — AI Daily May 28

634 messages · 95 active members

634
messages
95
active members
@arielletolome, @mb29266, @nowwatchthisdrive
top contributors

Overview

Claude Opus 4.8 dominated the day, rolling out mid-morning with granular reasoning controls, a 3x cheaper fast mode, a high-effort default, and benchmarks beating GPT-5.5 on engineering tasks. @arielletolome one-shot a converting auto insurance bridge page on her first prompt, but the launch was marred by a serious harness bug: @sibunting traced API 400 errors to Claude Code v2.1.153 persisting thinking blocks with empty text but retained signatures, permanently poisoning sessions on replay during compaction. A fix shipped in Claude Code 2.1.156 the next day. Dan Shipper noted Codex + GPT-5.5 still wins on harness quality, and Claude Design quietly merged its quota with main Claude — one request now burns 8% of daily sessions. Video generation drove the second-biggest thread: @victorbrsss and @samarh90 argued Sora 2 still wins on talking heads (YouTube-trained), but its September 24 deprecation is pushing builders toward Seedance (with storyboarding) and Kling 3. @navuud demoed a fully automated editor that took 4 Seedance clips (30s) and cut them to a captioned 17s finished video with zero manual edits. Veo 3.1 beat Omni on voice authenticity, while @namtalks flagged Veo 3's lip-sync issues. On infrastructure, @ecomrat used Codex computer use to handle 40-50 software admin invites during an acquisition, @rmktg detailed his 2k ads/day Meta pipeline, and @navuud shared a Telegram voice agent built on OpenAI Realtime API + pyrofork with sub-second latency. Side threads covered Oura Ring 5 preorders for personal health dashboards, OpenAI's private MCP servers, Grok CLI / dmux as Claude Code alternatives, and SpaceX's custom C-based training stack for 220K GPUs.

Topics

Claude Opus 4.8 Launches with Fast Mode, High-Effort Default, and Benchmark Wins73 msgs

Opus 4.8 shipped with granular reasoning controls, a 3x cheaper fast mode, ultracode/max-thinking tiers, and benchmarks beating GPT-5.5 on engineering. @arielletolome one-shot a converting bridge page, but credits burn aggressively on high effort. Claude Design quota merged with main Claude — one request hit 8% of daily session. Anthropic teased a higher-intelligence model class in coming weeks.

Claude Code Thinking-Block Bug Poisons Opus 4.8 Sessions21 msgs

@sibunting diagnosed Claude Code v2.1.153 persisting thinking blocks with empty text but retained signatures, causing API 400s on replay and permanently breaking sessions — amplified by 4.8's high-effort default. Nine GH issues confirmed it as harness-side; recommended rollback to 4.7. Anthropic shipped a fix in Claude Code 2.1.156 the next day.

Seedance vs Sora 2 vs Veo: Video Gen Showdown & Automated Editing38 msgs

Sora 2 still wins on talking heads (YouTube-trained) but deprecates September 24, pushing builders to Seedance (with storyboarding) and Kling 3. Veo 3.1 beat Omni on voice authenticity. @navuud demoed an automated pipeline turning 4 Seedance clips (30s) into a captioned 17s finished video with b-roll mapping and dead-space removal — zero manual edits.

Hermes Voice Agents, Grok CLI, and Browser Agent Alternatives42 msgs

@navuud shared a Telegram userbot ('Migi') built on pyrofork + py-tgcalls and OpenAI Realtime API (gpt-realtime-2) for single-hop voice-to-voice with native VAD and barge-in. Hermes Agent v2026.5.28 dropped with crash fixes. @arielletolome highlighted Grok CLI carrying over Claude Code commands and dmux.ai as terminal-native alternatives. @nowwatchthisdrive hunted Claude-for-Chrome replacements — Vercel agent-browser and Camofaux surfaced.

Codex Computer Use, Ad Automation & Personal Health Stacks34 msgs

@ecomrat used Codex computer use to invite IT admins to 40-50 software subscriptions during an acquisition. @rmktg detailed a 2k ads/day Meta pipeline with system user tokens, rate-limit logging, and pattern-based cooldowns. Oura Ring 5 preorders sparked discussion on stitching OAuth biometrics into Claude Code and Hermes for proactive agent nudges — @geilt uses it so AI bots know when to stop bothering him.

Key Takeaways

  • Upgrade Claude Code to 2.1.156 if on Opus 4.8 — fixes thinking-block persistence bug that permanently poisons sessions during compaction.
  • Opus 4.8 fast mode is 3x cheaper but high-effort default burns credits hard; Codex + GPT-5.5 still wins on harness quality per Dan Shipper.
  • Sora 2 deprecates September 24 — migrate talking-head workflows to Seedance (with storyboarding) or Kling 3 now; Veo 3.1 leads on voice authenticity.
  • Codex computer use is production-ready for tedious admin work — handles 40-50 software invites with minimal intervention; great for acquisition integrations.
  • OpenAI Realtime API (gpt-realtime-2) + pyrofork enables sub-second voice-to-voice agents on Telegram with native barge-in and tool calling.

Hot Threads

@sibuntingstarted

Opus 4.8 thinking-block bug breaks Claude Code sessions — root cause analysis and rollback recommendation

13 replies7 participants
@arielletolomestarted

Opus 4.8 one-shot auto insurance bridge page on first prompt, plus Grok CLI / dmux as Claude Code alternatives

12 replies7 participants
@navuudstarted

Telegram voice-call agent on Realtime API + automated Seedance video editor cutting 4 clips to finished 17s output

10 replies4 participants

Linked Items

OpenAI's Sora: Cryptographic Video Authentication in Action

mamuzlgelblzowmikoqr.supabase.co
@arielletolome

Opus 4.8 Launch Imminent as OpenAI Prepares Counter-Move

x.com
@watchmedropship

x.com

x.com
@arielletolome

OpenAI Enables Private MCP Servers for Secure Enterprise Integration

x.com
@weslindquist

Instagram

instagram.com
@realcrischico