The headline rates, and the one that changes
| Per 1M tokens | Claude Haiku 5.5 | DeepSeek V4.1-Flash-1017 |
|---|
| Input | $0.10 (prompts to 100K) | $0.30 peak / $0.15 off-peak |
| Output | $0.50 (prompts to 100K) | $1.20 peak / $0.60 off-peak |
| Input over 100K | $0.50 | unchanged |
| Output over 100K | $2.50 | unchanged |
| Cache reads | $0.01 | not published separately |
| Licence | proprietary API | MIT, open weights |
| Size | not disclosed | 552B mixture-of-experts |
| Context | 1M | not published |
| Max output | 128K | not published |
The 100K cliff
This is the comparison. Haiku 5.5 is marketed as a 1M-context model at ten cents per million input tokens. Both statements are true separately and misleading together: the ten-cent rate applies only to prompts up to 100,000 tokens. Above that, input becomes $0.50 and output $2.50 - five times the advertised price.
So the cheap tier is cheap for the first tenth of its context window. If your workload is short prompts at volume - classification, extraction, routing, summarising single documents - Haiku is genuinely inexpensive and nothing beats it. If you were attracted by 1M context and budgeted at $0.10, the first invoice that uses that context arrives at 5x your model.
DeepSeek V4.1-Flash has no equivalent threshold. $0.30/$1.20 at peak, roughly half off-peak, and the rate does not change with prompt length.
Where the crossover sits
Under 100K tokens, Haiku wins outright: $0.10 against $0.30 input, and $0.50 against $1.20 output. Three times cheaper, before cache reads at $0.01 push it further.
Over 100K, it inverts: Haiku at $0.50/$2.50 against DeepSeek at $0.30/$1.20 - DeepSeek becomes roughly half the price. Off-peak, a third.
Two adjustments before you take that as final. Anthropic's own documentation states the Claude 4.7-era tokenizer produces about 30% more tokens for the same text, so Haiku's effective rate is above its sticker on both sides of the cliff. And DeepSeek's off-peak pricing only helps if your work can wait, which agent workloads generally cannot.
What else differs
Open weights. DeepSeek V4.1-Flash is MIT-licensed, 552B mixture-of-experts. You can run it yourself, which removes the price question entirely and replaces it with a hosting one. Haiku is API-only.
Live search. DeepSeek ships Search and Fetch APIs at $0.75 per 1,000 successful requests, with failures free. That last detail is unusually fair and rarer than it should be.
Thinking controls. Haiku 5.5 has adaptive thinking on by default and an effort parameter defaulting to medium - the first Haiku-class model with adjustable effort. There is a compatibility catch: non-default temperature, top_p or top_k return a 400 error, so existing code that tunes those will break on migration.
Modality. Haiku takes text and images in, text out, with a June 2026 knowledge cutoff. DeepSeek has a 5 QPS per-account limit.
Pick one
- Short prompts at volume - Haiku 5.5, decisively. Under 100K it is the cheapest serious option.
- Long documents, large codebases, agent transcripts - DeepSeek. The cliff makes Haiku roughly twice the price on exactly this work.
- You need open weights - DeepSeek. MIT, and nothing else here is close.
- Predictable billing - DeepSeek. One rate regardless of prompt length.
- Images in the prompt - Haiku 5.5. DeepSeek publishes no multimodal input.
- Heavy repeated context - Haiku, on cache reads at $0.01 per million, provided you stay under the threshold.
- Batch work that can wait - DeepSeek off-peak at $0.15/$0.60 is the cheapest number on this page.
Verdict
These are not really competing for the same job. Haiku 5.5 is the best short-prompt economics available and a poor long-context choice despite advertising a 1M window. DeepSeek V4.1-Flash is slightly more expensive on short work and materially cheaper on long work, with open weights as the tiebreaker.
The practical test takes a minute: find the median prompt length in your logs. Under 100K, Haiku. Over, DeepSeek. Straddling the line, DeepSeek - because a bill that changes by 5x depending on input length is not a bill you can forecast.