Cost receiptTested AIBy DanPublished Tested Updated
My real DeepSeek bill: 12.31 billion tokens for $122.28
All reports
- ChatGPT + Codex built 33 new DMR_REMOTE toolsField testOct 23
- Gemini 4 Argon reviewed a finished episode: 14 clips, one flag, one inconsistencyField testOct 21
- Does an extra AI review step pay off? A small DeepSeek trial said noHead-to-headOct 19
- My real DeepSeek bill: 12.31 billion tokens for $122.28Cost receiptOct 17
- Gemini 4 Argon kept hanging in Gemini CLI. The cause was a 60-second timeoutIncidentOct 14
- I asked 8 AI coding agents for their biggest weaknessSelf-reportOct 14
- My agents' first game benchmark was wrong. Here's how it was caughtIncidentOct 10
- DeepSeek Flash answered 200 OK and sent nothing for two hoursIncidentSep 14
- DeepSeek V4 Flash audited 18 of my real coding sessions. Two said 'done' too earlyAuditAug 4
The report
Verdict
About 1 cent per million tokens
Across 68,720 real requests in September 2026, my blended cost came to about one cent per million tokens. That is a cost receipt, not a quality verdict.
On the tier list: A cost receipt says nothing about quality. It places no agent.
The receipt
These numbers come from my own DeepSeek API console, shown privacy-masked in the video. "Last calendar month" on October 17 was September 2026. Almost all of it went through the deepseek-flash alias, mostly Claude Code sessions pointed at DeepSeek's API, doing agent coding and other work across my workshop.
Why it is so cheap
Coding agents re-send the same context over and over, and cached input is priced far below fresh input. The console summary does not split cache hits out, so I cannot give September's cache share. For a sense of scale: in 18 DeepSeek V4 Flash sessions I audited in August, cache reads were 331.1M tokens against 8.96M fresh input and 4.22M output.
What the video's comparison table is
The episode also shows an eleven-row table where DeepSeek V4.1-Flash-1017 is ahead of Claude Opus 5 on 4 rows and behind on 7. That table is DeepSeek's own published numbers, against Claude Opus 5, not Opus 5.5. It is context, not my test. It does not place anything on the tier list.
The first head-to-head on my own work is in the AI Coding Agents Tier List queue.
Limits of this report
- Cost only. A cheap token says nothing about whether the work was right.
- Workshop-wide usage, not one controlled task.
- The console does not break out the cache-hit ratio, so cache savings were not measured.
From the channel
Receipts
DanManREAL · AI lab
Test receipt
DeepSeek V4.1 Flash
- Spend, September 2026
- $122.28
- Requests
- 68,720
- Tokens
- 12,311,161,60912.31B
- Of which on the deepseek-flash alias
- 66,428 requests12,002,085,348 tokens
- Rolling 30 days to October 17
- $83.7545,590 requests, 8,911,471,337 tokens
- All-time platform spend
- $255.64
- Blended cost
- ≈ $0.01 per 1M tokens
Verdict
About 1 cent per million tokens
Tier list A cost receipt says nothing about quality. It places no agent.
danmanreal.com/tests/deepseek-api-bill-12-31b-tokens
What was tested
On the tier list's bench
- DeepSeek V4.1 Flash · 4 receipts on file
Shorts
Related
- Does an extra AI review step pay off? A small DeepSeek trial said no
- DeepSeek Flash answered 200 OK and sent nothing for two hours
Independent testing. Sponsorships are always disclosed, and verdicts are never for sale.
Disagree? Bring your receipts