DeepSeek vs Haiku pricing debunk: accounting for peak rates and token use, 1017 is 2-2.5x cheaper
teortaxesTex · x · 2026-10-28
Pushing back on the claim that Claude Haiku 5.5 (34.9% vs 28.5% on FrontierCode 2, $0.46 vs $0.52 per task) beats all Chinese models while being cheaper, teortaxesTex argues the comparison is comically unfair: DeepSeek V4.1 Flash 1017 was charged peak pricing ($0.3/$1.2 per M tokens, 3x/2.4x Haiku's, active only 20% of the time), DeepSeek keeps flat pricing across 1M context while Haiku spikes 5x on longer contexts, and Haiku burns far more tokens (33.6M vs 15.3M output tokens). Factoring all three, 1017 is likely 2-2.5x cheaper in real API costs. Both ran on the Stirrup harness. Screenshots include the FrontierCode 2.0 score-vs-cost chart and per-task cost breakdowns.
More from coding & agent
- Parallel launches Decisions API: web-grounded yes/no answers with confidence scores at $0.001/request — heyshrutimishra · 2026-10-28
- Investor Gavin Baker: Grok Bot is the most impactful AI product at his firm, with 17 bots built — elonmusk · 2026-10-28
- Wes Roth's Blood Bench: AI builds and coaches competing fantasy football teams end to end — Wes Roth · 2026-10-28
- Claude reads my social feeds nightly and shows highlights on an Apple TV dashboard — florianvo · 2026-10-28
- Developer slams Apple: each iOS simulator stores its own Siri model, wasting 25.5 GiB — Dimillian · 2026-10-28
- Show HN: RSI-Gear, a Lego-like composable framework for recursive self-improvement — ethan1998 · 2026-10-28