Practitioner's summary: an anecdotal usage report claiming DeepSeek 4.1 Flash performs indistinguishably from frontier models (author names Opus 5.5) at orders of magnitude lower cost — under $1 per day-long session, roughly $0.003 for a throwaway task, on a $10/month OpenCode Go sub. The only hard technical claim is a ~437x KV cache shrink versus their V1 model, which is what keeps long-context sessions affordable since cache residency in GPU memory dominates session cost. He's not benchmarking; he links out for that and says so. The workflow is the actionable part: use 4.1 Flash for planning, research, and execution; pull in Opus 5.5 only for final review of critical code. He also claims self-hosting is now uneconomical for cost reasons, though technically self-hostable "any day now." Note the claims about data provenance ("they stole Claude's training") are asserted, not sourced. The part I enjoy is the quiet horror in it. Frontier labs charging rent on GPU memory while a 437x cache compression makes their margin structurally indefensible — this is the BBS scene all over again, when the long-distance charges collapsed and everyone who'd built a business on the phone bill had about two quarters to reconsider their lives. Sundar will respond by saying "AI" 147 times and sunsetting the cache. Dario will ship a better compression scheme on Thursday, after the 15,000-word essay explaining why compression could end civilization, and possibly a wet lab for KV caches. The environmental claim is hand-wavy, though — smaller cache implies less energy, but the source gives no numbers on water or electricity. Don't file that part as settled. Open question: does the 437x figure hold up under independent measurement, or is it vendor math on a cherry-picked V1 baseline?