Controlled benchmark: local Qwen3.8-Flash-Next scores 68% vs Opus 5.5's 87% at a fraction of the cost
deepu105 · reddit · 2026-10-20
The author ran a controlled six-way comparison (each twice, 12 PRs total) on a real LlamaStash feature — saving and restoring llama-server's KV cache — scored by 7 review questions plus CI.
Key results
- Claude Code + Opus 5.5 (xhigh): 6.94/8 (87%), 43.5 min, $17.78 API cost
- Opus 5.5 medium: 78%, $6.52, 23.4 min
- Flash-Next run locally on a Strix Halo laptop: 63–69% across three combos, $0.32–0.67 equivalent cost, 45–64 min (1.1–2.1x Opus at the same effort)
- All 12 PRs passed end-to-end tests; the gap lies in edge cases, tests and docs, not the core feature
- xhigh helped Opus slightly at 2.7x the cost; for Flash-Next, medium matched xhigh and was faster
Verdict: Opus still leads, but Flash-Next is close behind at near-zero local cost. All 12 PRs are public.
Related event: Benchmarks Show Qwen3.8-Flash-Next Near Opus 5.5 at 1/50 the Cost(3 posts)→
More from coding & agent
- Open-source MCP server gives agents live aircraft and ship tracking with cited sources — Fast-Aspect-2406 · 2026-10-21
- Agent UX: spaCy creator Ines Montani on building tools for humans and coding agents — honnibal · 2026-10-21
- Ragged Blocks: a new way to render nested structure within typographic layout — neurocy · 2026-10-21
- Twenty, the open-source AI-native CRM alternative to Salesforce, nears 59k stars — twentyhq · 2026-10-21
- freehire: open-source AI job search engine hits 2k GitHub stars — strelov1 · 2026-10-21
- Running multiple Claude instances to talk to each other slows everything down — rickasaurus · 2026-10-21