Fable 5 Killed, Opus 4.8 Regresses, OpenRouter Fusion — AI Daily Jun 13
699 messages · 70 active members
Overview
Topics
Anthropic globally disabled Fable 5 and Mythos 5 following a reported jailbreak disclosure from Amazon devs to the US government, with Pliny's published Fable system prompt cited as the trigger. Theories range from DOD deals requiring ID/PLAID verification, to Amazon protecting AWS compute revenue, to a pre-IPO PR stunt. Builders expect a relaunch this week with KYC requirements and 5x demand from FOMO.
Opus 4.8 was widely panned for half-assing tasks, deleting work, ignoring discussed context, taking shortcuts, and prompting to end sessions early. @jasonakatiff called Fable 'twice as good' and is migrating to Codex 5.5; many are keeping Claude only for copywriting. Opus 4.6 still rules for longform VSL/advertorial work at high effort level.
Multiple builders report GPT-5.5 has fixed the lecturing/communication issues that pushed them to Claude, calling it 'fucking amazing.' Migration tip: symlink .claude → .codex and CLAUDE.md → AGENTS.md. Emerging hot stack: Claude/Opus for PRDs and planning → Codex with goals for autonomous runs → Antigravity CLI for parallel terminal throughput on large codebases, plus Codex subagents on Hermes via Discord.
OpenRouter launched a Fusion API claiming Fable-level intelligence at half the cost by aggregating multiple models. Builders discussed DIY versions: same prompt to multiple models, an aggregator judge merges responses to consensus. Most want real-world tests before trusting the benchmarks, but it's emerging as the cost-effective hedge against frontier model dependence.
@mb29266 shared that 2008-iPhone-quality first frames (blurry, dirty lens, bad angles) produce dramatically more realistic AI UGC via Seedance 2, which can now twist off supplement caps with realistic sound. @zippi101 dropped his personal VSL swipe bible, @realcrischico shared invokable Claude personas (Halbert, Abraham, Kennedy, Sultanic). Emerging pattern: Grok for unfiltered aggressive ideation → Claude 4.6 to execute within proper copy frameworks, with scraped ad libraries as templates.
Key Takeaways
- Anthropic can yank model access mid-session with zero notice — bake fallback CLIs (Codex 5.5, Gemini/Antigravity, Kimi K2.7) into any production workflow and expect KYC on Fable relaunch.
- GPT-5.5 has won back skeptics — fixed the lecturing problem; migrate instantly by symlinking .claude → .codex and CLAUDE.md → AGENTS.md.
- Opus 4.8 is a regression for coding; keep 4.6 for longform copy, push all technical work to Codex with goals + AGENTS.md + superpower plugin.
- For realistic AI UGC, deliberately degrade first frame quality to 2008-iPhone aesthetics — blurry, dirty lens outperforms clean references on Seedance 2.
- Reference default model training for famous copywriters (Halbert, Kennedy, Abraham) rather than stuffing books into context, and feed scraped ad libraries as templates instead of relying on Claude's native creative structure.
Hot Threads
Fable 5 & Mythos shutdown — Amazon snitching, DOD deals, and government intervention theories
Opus 4.8 half-assing work vs Fable 5 one-shotting — migration to Codex
OpenRouter Fusion API beats raw Fable at half cost — real or marketing?