大模型网摘"Open Models Do 90% of Agentic Work at 1/20 the Cost": Where the Claim Comes From and When It Holds
Around November 3, several similarly titled threads appeared on Hacker News making the same claim: open-weight models can handle about 90% of agentic work at roughly one-twentieth the cost of closed flagship models. The number triggered the familiar developer-community argument — half posting bills in agreement, half posting failure stories in rebuttal. This piece picks no side; it takes the claim apart: where it comes from, what "90%" covers, and when it holds.
Where the claim comes from
As far as we could verify, the claim currently circulates mainly through community discussion (multiple HN threads, developer threads on X); there is no rigorous benchmark report behind it with a consistent methodology — no defined task set, no unified success criteria, no published cost worksheet. In other words, it is an experiential judgment, not a measurement. Comparable judgments have precedent: October's report that DeepSeek's API volume reached three times OpenAI's (see our write-up) already suggested, from the demand side, that cheap open models carry real production load. The market has voted with its feet once.
Unpacking "90%": which parts hold
Split "agentic work" into subtasks and open-model coverage is uneven. The part that holds: routing and dispatch, structured extraction, format conversion, code changes with clear acceptance criteria, retrieval-augmented QA, batch processing — tasks whose success is machine-verifiable, where tuned open models perform steadily. DeepSeek's V4.1-Flash-1017, for instance, is explicitly tuned for agentic coding (see our tech-report explainer); Kimi K3.1 Pro ships 3T-scale open weights with 1M context (our hands-on notes), matching flagships on raw window. The doubtful part: long-horizon planning (staying on rails across dozens of tool calls), complex cross-system orchestration, and open-ended tasks where the model must decide when to stop — the rebuttal posts' failure stories cluster here. If "90%" counts subtasks by frequency, the hold-fast segment dominates call volume, which is exactly why the claim spreads.
The cost math: how you get to 1/20
A price anchor helps: deepseek-flash lists $0.15 in / $0.60 out per 1M tokens off-peak, cache-hit at $0.003; mainstream closed flagships sit around $1 input and $3–5 output (e.g., GPT-6.1 Terra at $1/$5 — see our comparison). Divide sticker prices and the output gap is roughly 5–8x; reaching "1/20" requires stacking three more conditions: batch jobs off-peak, cache hits shrinking input cost toward zero, and self-hosting the open weights (leaving only power and hardware depreciation as marginal cost). Teams with all three in place aren't exaggerating; users on list-price APIs see closer to 1/5–1/10.
A self-check you can actually run
Don't trust anyone's ninety percent — including this piece. Take your last 100 real agent calls, split them into machine-verifiable vs. human-review-required, replay each group on an open tier (deepseek-flash, Kimi, Qwen, GLM — pick one) and on a closed flagship, and record success rate and per-call cost. Two weeks gets you your own two numbers — and your task mix's "90%" may be 70%, or 95%. The numbers will shift quickly with each release, but the conclusion is stable: cost-sensitive agentic workloads deserve this bake-off once a quarter.
Note: this is an unofficial multi-model news site; content is compiled from publicly available sources, and prices follow the latest official pages.