DoubleChecks

What was claimed

Deepseek distilled V4 pro into V4 flash to make frontier-level AI dirt cheap while pushing 6.9T+ tokens a week.

Our verdict

Needs caution

Multiple sources say V4-Flash is trained separately and is not distilled from Pro. That directly contradicts the claim. The sources support that V4-Flash is much cheaper than V4-Pro and optimized for efficient inference. But “frontier-level AI” is subjective and not directly established by the sources.

2 of 3 AI systems agree20 sources citedChecked Aug 10, 2026

Check your own claim

Paste any statement, headline, or AI answer — 3 independent AIs verify it in seconds, with sources.

Key findings

Deepseek distilled V4 pro into V4 flash

Incorrect97%
2 of 3 AIs agree·ChatGPT: Misleading

Deepseek made frontier-level AI dirt cheap with V4 Flash

Misleading80%
All 3 AIs agree

Deepseek is pushing 6.9T+ tokens a week

Can’t verify82%
2 of 3 AIs agree·Claude: Verified

V4 flash is dirt cheap

Verified90%
1 AI checked

Verified against 20 sources

framia.converge.aideepseek.aiapi-docs.deepseek.comwavespeed.aiwavespeed.aiwavespeed.aicoderouter.ioapi-docs.deepseek.comrephrase-it.comcodersera.comreddit.comcloud.tencent.comdeepinfra.comopenrouter.aireddit.comhuggingface.coartificialanalysis.aiyoutube.comfireworks.aiofox.ai

Detailed Analysis

The statement is not supported by the provided sources. The sources say V4-Flash is a separate model and explicitly state it is not distilled from V4-Pro, so the “distilled V4 pro into V4 flash” part is false. The claim about pushing 6.9T+ tokens a week is also not verified by the sources and has no clear corroboration here.

Why this verdict

  • The statement is not supported by the provided sources.
  • The sources say V4-Flash is a separate model and explicitly state it is not distilled from V4-Pro, so the “distilled V4 pro into V4 flash” part is false.
  • The claim about pushing 6.9T+ tokens a week is also not verified by the sources and has no clear corroboration here.

Claims checked

Deepseek distilled V4 pro into V4 flash

Incorrect98%
Multiple sources say V4-Flash is trained separately and is not distilled from Pro. That directly contradicts the claim.

This claim may mislead readers without added context.

Deepseek made frontier-level AI dirt cheap with V4 Flash

Misleading79%
The sources support that V4-Flash is much cheaper than V4-Pro and optimized for efficient inference. But “frontier-level AI” is subjective and not directly established by the sources.

This claim may mislead readers without added context.

Deepseek is pushing 6.9T+ tokens a week

Can’t verify91%
None of the provided sources confirm this usage figure. There is no direct evidence here to verify the weekly token volume.

What other AI think

The claim contains one significant factual error about the relationship between V4 Pro and V4 Flash, but accurately captures the cost-effectiveness and recent token throughput. The "frontier-level" framing is also somewhat misleading given that V4 trails current frontier models by 3-6 months.

Why this verdict

  • The claim contains one significant factual error about the relationship between V4 Pro and V4 Flash, but accurately captures the cost-effectiveness and recent token throughput.
  • The "frontier-level" framing is also somewhat misleading given that V4 trails current frontier models by 3-6 months.

Claims checked

Deepseek distilled V4 pro into V4 flash

Incorrect95%
V4 Flash is not a distillation of V4 Pro. It is a separate training run at smaller scale within the same family, post-trained with SFT, GRPO, and on-policy distillation. Both models were independently trained on 32T+ tokens.

This claim may mislead readers without added context.

V4 flash is frontier-level AI

Misleading85%
V4 Flash achieves near-frontier performance on specific tasks (e.g., math reasoning), but DeepSeek's own paper states V4 trails state-of-the-art frontier models by approximately 3-6 months. It's frontier-adjacent, not frontier-level.

This claim may mislead readers without added context.

pushing 6.9T+ tokens a week

Verified85%
V4 Flash processed 7.22 trillion tokens on OpenRouter in the week of July 27 to Aug 2, 2026, exceeding the 6.9T+ figure cited.
Parts of the statement match public reporting about DeepSeek V4 Pro and V4 Flash. However the wording implies specific technical lineage and an exact traffic number that are not fully supported by available sources. Some claims are therefore misleading or cannot be verified from public data.

Why this verdict

  • Parts of the statement match public reporting about DeepSeek V4 Pro and V4 Flash.
  • However the wording implies specific technical lineage and an exact traffic number that are not fully supported by available sources.
  • Some claims are therefore misleading or cannot be verified from public data.

Claims checked

to make frontier-level AI dirt cheap

Misleading75%
Reporting and vendor docs state DeepSeek V4 shipping at substantially lower API prices and aggressive discounts versus some closed-source models, but calling it 'dirt cheap' is subjective and overbroad without exact pricing context and workload comparisons.

This claim may mislead readers without added context.

Deepseek distilled V4 pro into V4 flash

Misleading78%
DeepSeek released both V4-Pro and V4-Flash on April 24, 2026, with Flash a much smaller 284B/13B-active model while Pro is 1.6T/49B-active; public materials show Flash is a lighter variant but do not clearly describe it as a literal distillation of Pro into Flash.

This claim may mislead readers without added context.

while pushing 6.9T+ tokens a week

Can’t verify72%
Public usage reports show DeepSeek V4-Flash consumed trillions of tokens in some weekly snapshots (e.g., ~3.7T reported in one week), but I could not find authoritative evidence that DeepSeek was consistently 'pushing 6.9T+ tokens a week' across the entire service.

Share this result

Try DoubleChecks