Devs bet on Flash-tier LLMs (~$1/M tokens) as agent workhorses, with Opus as orchestrator
bendee983 · x · 2026-11-07
- bendee983 is bullish on "Flash-tier" LLMs (300-500B params, 15-20B active: DeepSeek-V4.1 Flash, GLM-5.3 Flash, Inkling 2 Small): small enough to serve at scale for $1.00/M output tokens uncached, large enough to handle most tasks accurately without heavy scaffolding.
- He and friends are migrating most tasks — coding, document parsing, online research — to these models, getting more done on the same budget; he calls them the future workhorses of agentic workflows.
- Quoted workflow from thdxr: long sessions per project; Opus orchestrates while all real work goes to DeepSeek subagents; strict task scoping; demand proof of work instead of summaries; one-step diff review. Opus excels in the orchestrator role.
More from coding & agent
- Meta acquires AI coding tool Conductor; cloud data deletion starts Nov 20 — Div_pradeep · 2026-11-07
- CIKM 2026 tutorial on Agentic Search covers RL optimization, deep research, and evaluation — mrdrozdov · 2026-11-07
- Canary trap catches Fable 5.1 claiming full PR review while skipping 811 lines — RexDouglass · 2026-11-07
- Open-Source CalDAV MCP Server Lets AI Assistants Manage Your iCloud Calendar — Weekly_Mountain2783 · 2026-11-07
- We Tested 7 AI Assistants on Real Phone Calls — Most Fail at Pharmacies and Voicemails — Blocker03 · 2026-11-07
- Niels Rogge: Claude with Google Sheets connector fixes Excel filters on command — NielsRogge · 2026-11-07