DeepSeek Flash cost 23 cents less than Haiku 5.5 on three tests

A team picking a small, cheap model runs the same prompt against two options. Then it watches the bill. The second part is where the comparison tends to go wrong.

Jessica Wachtel at The New Stack published a three-test head-to-head on 22 October 2026. The subjects are DeepSeek V4.1-Flash-1017 and Claude Haiku 5.5. DeepSeek released the update on 17 October. The total bill across the three tests came to $0.07 for DeepSeek and $0.30 for Haiku.

Three tests, one run each

The test ran once per model through OpenRouter. DeepSeek was pinned to its own servers at model ID 1017 and off-peak hours. Haiku was pinned to Anthropic's servers at default settings. A script graded every answer against keys the author had built in advance. A second method checked each key before either model saw it.

Test one asked each model to fix three bugs in a small codebase. The error message pointed at the wrong file for the first bug. DeepSeek got 6 of 6 answers right. Haiku got 5 of 6 and missed one of two bug locations.

Test two was 20 database questions. Both models scored 20 of 20 and answered "unknown" on the questions the data could not answer. Haiku was faster and used fewer tool calls. DeepSeek cost about half as much.

Test three was a triage job over 1,600 fields. DeepSeek got all 1,600 right. Haiku got 1,596. Two misses were severity ratings. One report with 49 affected customers was rated sev5 when the written rule put it at sev4. Two more were routed to the old team for a service that moved to a new team on 1 August. The reports were filed after that date but mentioned an earlier July date in their body.

Cost split by tokenizer

The models count tokens differently. The same big-triage prompt was 153,082 tokens on DeepSeek and 216,811 on Haiku. That gap put Haiku on a higher long-prompt rate. The test's cost for the triage prompt was $0.06 on DeepSeek and $0.28 on Haiku. Haiku charges $0.10 per million input tokens and $0.50 per million output tokens below 100,000 tokens. The rate rises to $0.50 and $2.50 above that cutoff. DeepSeek's off-peak rate is $0.15 input and $0.60 output per million.

Who said what

The accuracy figures are Wachtel's measurement. The pricing is each vendor's published rate at the time of the test. Each model's claimed strengths were taken from the vendor announcements that framed the tests. Anthropic says its Opus models run faster in Fast Mode. Anthropic reports 39.2 on Terminal-Bench 4.0. That is 3.1 points below DeepSeek's reported score. Those are vendor-reported numbers, not a controlled head-to-head.

Three single-run tests do not settle a model choice. Rerun each case a few times and the ordering can move inside the measurement noise. What this comparison does well is state the method. The test, the data and the keys are clear enough for another team to run it on its own stack. DeepSeek V4.1 Flash is an open-weight model. A team that wants to pin pricing can host it. A team that already pays for Claude can run the same three tests against its own data in an afternoon. The useful output is the method a reader can apply. The gap on the triage test came mostly from the tokenizer. The accuracy gap was smaller.

Source

• The New Stack, by Jessica Wachtel: DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: More accurate at a fraction of the price
• DeepSeek V4.1-Flash on Hugging Face: DeepSeek-V4.1-Flash
• Anthropic Haiku 5.5 pricing: claude.com/pricing