Delta Report · Stacked Podcast · Aug 3 2026 · 28 min

Qwen 3.8 Max just commoditized Fable-class intelligence

What this video adds on top of what the vault already knew, and what it changes for your stack.

NEWThe headline numbers

$2 / $6per M in / out

2.4Ttotal params, MoE

95Bactive params

1Mcontext window

$0.17cache read / M

~7 daysto open weights

Alibaba dropped Qwen 3.8 Max on Aug 3. Hosts claim it ties or beats Fable 5 on PaperBench, Agent's Last Exam, OSWorld Verified, ERQA, Terminal Bench 2.1, and sits within ~1% on Cowork Bench. At Haiku-level pricing: Fable is $10/$25, so roughly 5x cheaper in, 4x cheaper out, and cheaper than Sonnet 5. Cache write $0.25, cache read $0.17 per M.

Open weights are confirmed for release within about a week, with a likely quantized cut and interest in a ~27B sibling for laptop-local use. Runs today via Qwen Studio (desktop-app style) and Qwen Cloud API, and drops into the Claude Code harness easily.

SHIFTWhat this changes vs the vault

Open-weight SOTA (Jul 24) is now stale early. The standing pick was Qwen3.5-35B-A3B with a 90-day revisit on Oct 24. Qwen 3.8's open-weight drop next week likely resets that decision two months ahead of schedule, including the LMS RTX 5090 + Qwen3.5-27B sizing note if a 3.8-generation 27B lands.
Model routing stack gets a new tier. The frugal doctrine routes grunt to Haiku/Sonnet-class. A Fable-adjacent model at Haiku pricing collapses the price-quality tradeoff for API grunt work, worth a bake-off before any wholesale routing change. Hermes routing (ChatGPT OAuth first) is unaffected until someone hosts it cheap.
Aug 3 "ChatGPT replacement" SOTA strengthens. That report already had Qwen winning on breadth; a frontier-tier flagship on top makes the Qwen ecosystem call look more right, not less.
Benchmark fatigue caveat from the hosts themselves: they say charts have lost predictive meaning and they'll trust 3-4 days of real use. Their live one-shot website test had Qwen 3.8 visibly beating Fable 5, one anecdote, not proof.

NEWSecondary signals

Hugging Face hack goes legal. The rumored GPT-6 driven Hugging Face breach is now a policy fight: the 1986 anti-hacking statute requires human intent, which breaks when an AI does the hacking. Likely landing zone: vicarious liability, you own what your model does (the dangerous-dog analogy). Precedent here shapes all agent accountability.
PLA distillation claim: China's military allegedly used GPT-3.5 distillation outputs in drone surveillance networks since ~2022. Context for why US labs run anti-distillation measures and why Chinese labs price at a loss to harvest usage data.
State-funded loss-leader thesis: the pricing is strategic, pulling users off Western models partly to collect training data.

ACTMoves worth considering

Watch for the open-weight drop (~Aug 10) and re-run the open-weight model selection then, not Oct 24.
Bake off Qwen 3.8 Max vs current grunt tiers on one real Angel workload before touching routing.
For LMS local-LLM proposals, hold quotes that name Qwen3.5 until the 3.8 family lands.

NOISESkippable

Roughly a third of the runtime: a Djokovic elevator-bodyguard story in Montenegro, Q&A banter, sponsorship jokes, Asimov recommendations. No signal lost by skipping.