DeepSeek is cheaper per token on input, but not always per task. In my measurements, DeepSeek V4.1 Flash beat Claude Haiku 5.5 only on summarizing a PDF, and lost on translation, ticket classification and SQL generation. The gap comes mostly from reasoning tokens, which are billed as output.
The prices
List prices per 1M tokens (input / output):
- DeepSeek V4.1 Flash: $0.036 / $0.68
- Claude Haiku 5.5: $0.10 / $0.50
- DeepSeek V4 Pro 0813: $0.66 / $1.98
- Claude Sonnet 5.5: $2.00 / $10.00
DeepSeek Flash has the lower input price, but Haiku has the lower output price. Output is where long answers and reasoning add up, so list price alone does not decide the result.
Tokenizers matter too. Anthropic's ratio is 1.39 against OpenAI at 1.00, and DeepSeek is at 1.00. The same English document becomes more tokens on Claude before any reasoning is added.
Per-task cost
Costs below use default reasoning settings.
| Task | DeepSeek V4.1 Flash | Claude Haiku 5.5 | DeepSeek V4 Pro 0813 | Claude Sonnet 5.5 |
|---|---|---|---|---|
| Translate 1,000 words | 0.81¢ | 0.13¢ | 2.4¢ | 2.6¢ |
| Summarize 20-page PDF | 0.18¢ | 0.22¢ | 1.3¢ | 4.5¢ |
| Classify 100 tickets | 0.34¢ | 0.09¢ | 1.1¢ | 0.89¢ |
| Write SQL query | 0.15¢ | 0.08¢ | 0.43¢ | 1.1¢ |
Measured: DeepSeek V4.1 Flash and Claude Haiku 5.5 on all four tasks, and Sonnet on all four tasks. Estimated: DeepSeek V4 Pro 0813 on all four tasks. Each measured run is one run per task at default settings via OpenRouter, so treat small gaps with caution.
| # | Model | Cost per task | Text score | Tokens in / out | Basis | |
|---|---|---|---|---|---|---|
| 1 |  Qwen3.8 FlashAlibaba Qwen | 0.08¢$0.82 / 1,000 | - | 1,386 / 1,302 | measured | |
| 2 |  GPT-6 LunaOpenAI | 0.08¢$0.83 / 1,000 | 1443 | 1,330 / 1,395 | measured best value | |
| 3 |  Llama 4 MaverickMeta | 0.11¢$1.10 / 1,000 | 1327 | 1,375 / 1,292 | estimated | |
| 4 |  Claude Haiku 5.5Anthropic | 0.13¢$1.26 / 1,000 | - | 1,832 / 2,153 | measured | |
| 5 |  DeepSeek V4.1 FlashDeepSeek | 0.18¢$1.78 / 1,000 | 1475 | 1,354 / 11,708 | measured best value | |
| 6 | GPT-5.4 NanoOpenAI | 0.20¢$2.01 / 1,000 | 1401 | 1,330 / 1,395 | estimated | |
| 7 |  GLM 5.3 FlashZ.ai | 0.21¢$2.09 / 1,000 | 1474 | 1,330 / 10,155 | measured | |
| 8 |  MiMo-V2.6-ProXiaomi | 0.57¢$5.65 / 1,000 | 1479 | 1,486 / 5,752 | estimated best value |
Translation: Haiku wins clearly
DeepSeek V4.1 Flash produced 1,417 output tokens plus 10,291 reasoning tokens. Haiku produced 2,153 output tokens and no reasoning. Haiku's visible answer was longer, yet it cost 0.13¢ against 0.81¢ for DeepSeek.
Among the bigger models, V4 Pro (2.4¢, estimated) and Sonnet (2.6¢, measured) are close. Since V4 Pro is an estimate, I would not call that a firm win.
Summarizing a PDF: Flash wins narrowly
This task is mostly input. DeepSeek V4.1 Flash read 12,970 tokens and Haiku read 18,004, which reflects the tokenizer difference. DeepSeek cost 0.18¢ and Haiku 0.22¢, even though DeepSeek spent 1,565 reasoning tokens against Haiku's 118.
At the larger tier, Sonnet costs 4.5¢ and V4 Pro is estimated at 1.3¢. For summarizing, V4 Pro looks cheaper.
Classifying tickets: Haiku is cheaper
DeepSeek V4.1 Flash used 4,462 reasoning tokens and cost 0.34¢. Haiku used 961 and cost 0.09¢.
At the larger tier, Sonnet (0.89¢, measured) is slightly below V4 Pro (1.1¢, estimated). This is the one task where the pricier Claude model is cheaper than its DeepSeek counterpart.
SQL generation: Haiku wins again
SQL answers are short, so reasoning tokens weigh heavily. DeepSeek V4.1 Flash used 1,723 reasoning tokens and cost 0.15¢. Haiku used 810 and cost 0.08¢. At the larger tier, V4 Pro is estimated at 0.43¢ against 1.1¢ for Sonnet.
Quality
On the LMArena Text leaderboard, Claude Sonnet 5.5 scores 1476 and DeepSeek V4.1 Flash scores 1474. I have no LMArena score for Claude Haiku 5.5 or DeepSeek V4 Pro 0813, so I cannot rank those two this way. The cost figures are the firmer part of this comparison. Test quality on your own inputs.
Where the cheaper per-token model stays cheaper
- DeepSeek V4.1 Flash beats Haiku only on the PDF summary, where input volume dominates.
- Haiku is cheaper on translation, classification and SQL, where DeepSeek's reasoning tokens add up.
- DeepSeek V4 Pro 0813 is estimated cheaper than Sonnet on summarization and SQL, close on translation, and more expensive on classification.
If your workload is input-heavy with short answers, DeepSeek is the better bet. If it is reasoning-heavy, the per-token advantage can disappear.
Next steps
Use the calculator with your own token counts. Task breakdowns: translate 1,000 words, summarize a PDF, classify support tickets and write a SQL query. Model pages: DeepSeek V4.1 Flash, Claude Haiku 5.5, DeepSeek V4 Pro 0813 and Claude Sonnet 5.5. See the methodology for how I measure.
FAQ
Is DeepSeek always cheaper than Claude per task?
No. In my measurements, DeepSeek V4.1 Flash was cheaper than Claude Haiku 5.5 only on summarizing a PDF. On translation, classification and SQL generation, Haiku cost less.
Why does DeepSeek cost more on some tasks despite a lower input price?
Reasoning tokens are billed as output. DeepSeek V4.1 Flash used 10,291 reasoning tokens on translation and 4,462 on classification, which outweighed its low input price.
Which results are estimates?
The DeepSeek V4 Pro 0813 results are estimated. DeepSeek V4.1 Flash, Claude Haiku 5.5 and Claude Sonnet 5.5 results are measured.