DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: Accuracy, Speed and Cost Compared

Oct 22, 2026 867 views
Digital abstract art of layered pink and orange prisms and curled paper shapes against deep teal and green tones

DeepSeek V4.1-Flash-1017 and Claude Haiku 5.5 landed close on accuracy and total runtime in a small hands-on comparison. The standout difference was cost: DeepSeek’s three runs totaled $0.07, compared with Haiku’s $0.30, with most of the gap coming from a long triage prompt that crossed Haiku’s higher pricing tier.

DeepSeek released V4.1-Flash-1017 on October 17 as an update to its Flash model. The change log cites shorter task runtime and output length, alongside 11 benchmark results, including a 42.3 score on Terminal-Bench 4.0.

The model is available through DeepSeek’s API and services such as OpenRouter. Its open weights can also be downloaded from Hugging Face for self-hosting. These tests used OpenRouter requests routed to DeepSeek’s own servers and billed at DeepSeek’s rates.

Claude Haiku 5.5, which Anthropic introduced on October 7, is positioned for high-volume work such as classification, summarization and database queries. Anthropic reports a 39.2 score on Terminal-Bench 4.0, 3.1 points below DeepSeek’s published result. Those scores come from the vendors; they aren’t results from a controlled comparison.

Pricing depends on prompt size and caching

For prompts up to 100,000 tokens, Haiku costs $0.10 per million input tokens and $0.50 per million output tokens. Above that threshold, its rates rise to $0.50 and $2.50. DeepSeek’s off-peak prices are $0.15 per million uncached input tokens, $0.003 per million cached input tokens, and $0.60 per million output tokens; peak pricing doubles those rates.

That makes Haiku less expensive for short, uncached prompts, while DeepSeek can cost less on long inputs. As tokenization, caching and provider routing affect actual charges, list prices alone don’t predict every workload’s bill.

To compare practical performance, I ran three tasks once per model through OpenRouter, with requests pinned to each company’s servers and default settings. DeepSeek was set to model ID 1017, and all runs took place off-peak. A script graded the outputs against answer keys prepared and independently checked before testing.

  • Agentic bug fix: A Python job queue with three interacting defects and one faulty test. Each model could inspect files, make changes and run tests, with a limit of 25 turns.
  • Database queries: Twenty questions about a mock CI database, answered in SQL. The data included soft deletes, inconsistent casing, retried builds, time zones and currency conversion. Two questions had no answer in the dataset.
  • Triage: Classification of 400 developer error reports by category, severity, service and team, for 1,600 scored fields. The prompt counted as 153,082 input tokens on DeepSeek and 216,811 on Haiku, putting Haiku above its 100,000-token price threshold.

Bug fixing and SQL favor different tradeoffs

DeepSeek passed all six checks in the bug fixing task; Haiku passed five. Both found the first defect despite a misleading error location, fixed the third, avoided regression failures and left the incorrect test untouched. Haiku missed one hidden check on the second bug.

Haiku completed the work in eight calls and 21.6 seconds. DeepSeek needed 12 calls and 49.6 seconds, but its run cost less: $0.009 against Haiku’s $0.012. DeepSeek’s automatic caching covered most repeated input. Anthropic prompt caching wasn’t enabled, since it requires opting in, so turning it on could change the cost comparison.

  • DeepSeek: 190,139 input tokens, 7,877 output tokens, $0.009, 49.6 seconds.
  • Haiku: 100,461 input tokens, 3,255 output tokens, $0.012, 21.6 seconds.

The database task produced a tie: both models answered all 20 questions correctly, including recognizing that two answers couldn’t be determined from the available data. Haiku used fewer calls and finished sooner, while DeepSeek cost about half as much.

  • DeepSeek: 12 calls, 75,151 input tokens, 4,608 output tokens, $0.004, 34.9 seconds.
  • Haiku: 7 calls, 55,304 input tokens, 5,917 output tokens, $0.009, 28.4 seconds.

Triage exposes small accuracy gaps and a large bill gap

DeepSeek classified all 1,600 fields correctly. Haiku missed four. Two errors were severity ratings: the guide called for sev4 when fewer than 50 users were affected or the error rate was below 1%. Haiku assigned sev3 to a report affecting 49 customers and another with a 0.8% error rate.

The other two errors involved team routing. A service had moved to a new team on August 1, and reports were to be routed according to the date in their headers. Haiku sent two post-August 1 reports to the former team. Both included an earlier July date in their body, though the outputs don’t reveal why Haiku selected the old team.

The same text generated different input counts because the models use different tokenizers: 153,082 for DeepSeek and 216,811 for Haiku. Haiku’s higher count put the request in its more expensive tier. The DeepSeek run cost $0.058 and took 188.3 seconds; Haiku cost $0.282 and took 228.9 seconds.

  • DeepSeek: 153,082 input tokens, 58,923 output tokens, $0.058, 188.3 seconds.
  • Haiku: 216,811 input tokens, 69,354 output tokens, $0.282, 228.9 seconds.

Close results, with workload-specific caveats

TestDeepSeek V4.1-Flash-1017Claude Haiku 5.5
Agentic bug fix6 of 65 of 6
Database queries20 of 2020 of 20
Triage fields1,600 of 1,6001,596 of 1,600
Total tokens in / out418,372 / 71,408372,576 / 78,526
Total cost$0.07$0.30
Total model time272.8 s278.9 s

Across these runs, DeepSeek had a small accuracy edge and finished 6.1 seconds sooner overall. Haiku was faster in both tool-based tasks; DeepSeek was faster on the large triage job. The cost difference was much larger than either the accuracy or runtime gap, driven mainly by Haiku’s long-prompt tier.

My read: DeepSeek looks like the stronger fit for these specific tests, given its perfect scores and lower bill under the settings used. That’s a narrow result, though. Each model ran only once per task, and Haiku’s prompt caching was off. Teams choosing between them should test both against their own workloads, especially where prompt length, caching and routing rules shape cost or accuracy.

The post DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: More accurate at a fraction of the price appeared first on The New Stack.

Source: Jessica Wachtel · thenewstack.io

Comments

Sign in to comment.
No comments yet. Be the first to comment.

Related Articles

DeepSeek V4.1-Flash-1017 vs. Claude Haiku 5.5: More accur...