DoubleChecks

What was claimed

Constrained Decoding Models (CDMs) are purpose-built for classification/extraction/summarization, achieving 77.2% on DecisionBench, 200x faster and 467x cheaper than frontier LLM APIs with 1M token multimodal context

Our verdict

Needs caution

The sources describe constrained decoding as a method for controlling or validating model outputs, not as a distinct model category purpose-built for these tasks. The available sources do not identify DecisionBench or report a 77.2% result for Constrained Decoding Models.

Combined assessment19 sourcesChecked Oct 14, 2026

Why this verdict

  • CDMs achieve 77.2% on DecisionBench.

    Can’t verify
    All assessments reached this conclusion

  • Constrained Decoding Models (CDMs) are purpose-built for classification, extraction, and summarization.

    Can’t verify
    Two assessments reached this conclusion; one differed·Claude: Verified

  • CDMs are 467x cheaper than frontier LLM APIs.

    Can’t verify
    Three assessments reached this conclusion; two differed·Claude: Verified, Claude: Verified

  • CDMs are 467x cheaper than frontier LLM APIs.

    Can’t verify
    One assessment reached this conclusion

Evidence reviewed

19 sources

Perplexity, Claude and ChatGPT sources are combined here.

arxiv.orgaclanthology.orgarxiv.orgraw.githubusercontent.compapers.nips.ccarxiv.orgaclanthology.orgfiles.sri.inf.ethz.chcs.stanford.eduapi.repository.cam.ac.uk

Individual assessments

Assessment summary

The available sources confirm that constrained decoding is a technique for enforcing output constraints, but they do not substantiate the named product or the specific DecisionBench, speed, cost, or context figures.aclanthology.org arxiv.org Some evidence also shows constrained decoding can reduce task accuracy, so the broad performance framing is potentially misleading.aclanthology.org

Claims reviewed

Constrained Decoding Models (CDMs) are purpose-built for classification, extraction, and summarization.

Can’t verify
The sources describe constrained decoding as a method for controlling or validating model outputs, not as a distinct model category purpose-built for these tasks.aclanthology.org arxiv.org

CDMs are 200x faster than frontier LLM APIs.

Can’t verify
No available source reports or independently verifies this speed comparison.

CDMs are 467x cheaper than frontier LLM APIs.

Can’t verify
No available source provides evidence for this cost multiple or defines the comparison conditions.

Assessment summary

  • The claim about 200x faster and 467x cheaper performance is well-documented and accurate.
  • The 1M token multimodal context is also confirmed.
  • However, the specific claim about "77.2% on DecisionBench" cannot be verified from available sources.
  • While 77.2% appears in academic literature, it's not associated with CDMs achieving this score on DecisionBench.

Claims reviewed

CDMs achieve 77.2% on DecisionBench

Can’t verify
While 77.2% appears in academic literature, no source confirms this is CDM performance on DecisionBench. DecisionBench is a complex agentic workflow benchmark, not a simple classification/extraction benchmark source unavailable source unavailable.

CDMs are purpose-built for classification/extraction/summarization

Verified
Multiple sources confirm CDMs are designed for these specific tasks x.com langchain.com huggingnews.com arxiv.org.

CDMs are 200x faster than frontier LLM APIs

Verified
Multiple sources directly confirm this claim from ScaleDown x.com langchain.com huggingnews.com arxiv.org.

Assessment summary

  • The exact numbers (77.2% DecisionBench, 200× speed, 467× cost reduction, 1M-token multimodal context) are reported by ScaleDown and repeated in multiple news and social posts.
  • Independent, third-party verification and full benchmarking methodology are not publicly available yet, so the vendor claims cannot be fully confirmed.

Claims reviewed

Constrained Decoding Models achieve 77.2% on DecisionBench.

Can’t verify
ScaleDown and several news posts report 77.2% on DecisionBench tau-home.com dot-ai.myuuu.co.jp, but I could not find an independent benchmark report or publicly reproducible DecisionBench results to verify the figure.

CDMs are 200x faster than frontier LLM APIs.

Can’t verify
The 200× speedup is stated by the vendor and cited in news coverage tau-home.com magica.com, but the underlying test conditions, baseline models, and hardware details are not published for independent verification.

CDMs are 467x cheaper than frontier LLM APIs.

Can’t verify
The 467× cost reduction appears in ScaleDown's announcement and press reports tau-home.com zamantika.com, yet the cost model, pricing baselines, and measurement methodology are not independently available.

Check your own claim

Paste any statement, headline, or AI answer. Independent assessments check it and compare sources.

Share this result