Cheap Models Fill Structures Well; No Model Extracts Claims Well
Cheap Models Fill Structures Well; No Model Extracts Claims Well
On turning source material into structured facts, Xiaomi MiMo V2.6 Flash at $0.0015 per sync run ties Claude Sonnet 5.5 at $0.07 [^1]. On pulling individual factual claims out of free text, no model measured scores above 7.7 out of 10 [^2], and several models that score above 9 on filling structures fall below 7.
The benchmark files both jobs under structured data and fact extraction, but a model’s score on one says little about the other. Reading them as one leaderboard leads to two mistakes: overpaying for structure-filling, and trusting a claim extractor that has not earned it.
Filling a structure: the top is a tie
Prices here are per task run in sync mode unless stated. Sonnet 5.5 scores 9.88 (±0.16) after 30 independent judgments, and MiMo V2.6 Flash scores 9.76 (±0.22) [^1]. The intervals overlap, so the evidence separates them on price, not on quality.
Head-to-head judging points the same way. Across 119 judged occasions on this task, the top two answers tied exactly 54.6% of the time and finished within half a point 84.0% of the time [^1]. Those figures do not say which models formed the top two, but they show how little the choice at the top changes the output.
Extracting structured output is less flat. Claude Opus 5.5 leads at 9.93 (±0.16) for $0.13 [^1]. DeepSeek V4.1 Flash scores 9.25 (±0.33) for $0.003 [^1], which clears a working bar of 9 out of 10, but its interval sits below Opus 5.5’s. That bar is a threshold chosen for this analysis, not one the benchmark sets.
Batch mode lowers the price without moving the score. MiMo V2.6 Flash lists the same score in both modes and costs $0.0007 per batch run [^1].
Pulling claims from text: no model is strong
Extracting atomic factual claims scores lower for every model measured. GPT-6 Luna holds the top score, with an interval of ±0.19 after 200 independent judgments, at $0.001 per run [^2]. Moonshot Kimi K3 at 7.63 (±0.43) and Tencent Hy3 at 7.43 (±0.38) overlap with it, so the top is again a tie, only a much lower one [^2].
Strength at filling structures does not carry over. Claude Opus 5.5, the leader on structured output, scores 6.57 (±0.48) here [^2]. GPT-6 Sol scores 9.2 on structured facts [^1] but 5.89 here [^2], and MiniMax M3 drops from 9.12 on structured facts [^1] to 5.99 [^2]. Each drop is more than three points.
What to do with it
Route by the kind of work, not by one leaderboard position, and never reuse a structure-filling score to pick a claim extractor.
- Filling a fixed structure. Take the cheapest model that meets your bar: MiMo V2.6 Flash for structured facts, and DeepSeek V4.1 Flash for structured output if about 9 is enough [^1].
- Higher bar, smaller volume. Keep Claude Sonnet 5.5 or Claude Opus 5.5 for work where the top score matters [^1]. Batch anything that can wait.
- Pulling claims from text. Start from GPT-6 Luna or Tencent Hy3, with human review or a test on your own documents [^2].
The caveat that matters most is answer rate. Each score covers only the calls that returned an answer. MiniMax M3’s structured-facts score covers the 58% of attempts where it answered [^1], and GPT-6 Luna answers 82% of attempts on claim extraction [^2].
The evidence gives no answer rate for MiMo V2.6 Flash on structured facts, so measure it on your own traffic before committing. A model that ties on score but often returns nothing is not a tie in production.
The live benchmark at https://fronset.ai/benchmark/ lets you set the quality bar each job needs, then compare the cheapest models that clear it, in batch and sync, before you fix a production route.