Claude Haiku 5.5 vs GPT-6 Luna vs DeepSeek V4.1 Flash: A Research-Agent Test
I gave three new small models the same research job, and Claude Haiku 5.5 finished first: 35 seconds, against 45 seconds for DeepSeek V4.1 Flash and 4 minutes 33 seconds for GPT-6 Luna. Its answer wasn't the most accurate of the three, though. DeepSeek scored a little higher. So this is a post about a trade-off, plus an honest list of what one small test can and can't tell you.
Time to finish the same research task
The test
Each model ran as a research agent on one task: find and compare three AI meeting scheduling tools that integrate with Google Calendar. All three got the exact same prompt.
They ran in the same agent setup with the same tools, including web search, and each model got one run.
| Model | Reasoning setting | Time to finish | Quality score (provisional) |
|---|---|---|---|
| Claude Haiku 5.5 | Default (medium) | 35 s | 66% |
| DeepSeek V4.1 Flash | Off | 45 s | 69% |
| GPT-6 Luna | Default (medium) | 4 m 33 s (273 s) | 64% |
"Default" means medium for both Haiku 5.5 and Luna, according to Anthropic's model guide and OpenAI's GPT-6 Luna model page.
I scored each output myself, which is why I call the quality scores provisional: one reviewer, one run.
Claude Haiku 5.5 vs GPT-6 Luna vs DeepSeek: the results
Speed first. Claude Haiku 5.5, which Anthropic released on October 7, 2026, took about 22% less time than DeepSeek and ran roughly 7.8 times faster than Luna (35 seconds versus 273).
Quality tells a different story. DeepSeek V4.1 Flash, released September 10, 2026, delivered slightly more complete research and scored 69%. Haiku came in at 66% with a structured comparison that covered pricing, features, Google Calendar compatibility, limitations and a final recommendation. GPT-6 Luna, which OpenAI calls its "most efficient model for focused, high-volume tasks," gave careful explanations and scored 64%, but it took more than four minutes to get there.
So Haiku was not the most accurate model in this test. What it did was land within three points of the top score in under a minute.
Why speed matters for an AI agent
An agent doesn't answer once. It searches, reads a page, decides what to do next, calls another tool and repeats, and a slow model pays its delay on every one of those steps while someone waits for the result.
Anthropic's own guidance points the same way. Its model guide says choosing a model means "balancing capabilities, speed, and cost," and it suggests an efficiency-first approach: start with Haiku 5.5, test your use case, and "upgrade only if necessary for specific capability gaps." It lists that approach as a fit for applications with tight latency requirements. That's vendor advice about its own product, not proof of my result, but it matches what I saw.
What this test doesn't prove
- It's one task and one run per model. A coding agent or a long document review could rank these models differently.
- The settings weren't identical. DeepSeek ran with reasoning off, which you'd expect to help its speed, and Haiku still finished faster in this test.
- The quality scores are provisional, and I gave them myself.
- It isn't a ranking of general intelligence.
- Providers update models, so these numbers can change.
My pick
For a basic research agent, Claude Haiku 5.5 is my preferred model of the three. It finished faster than DeepSeek V4.1 Flash, far faster than GPT-6 Luna, and its output was close enough in quality that I'd rather have the speed. DeepSeek was a bit more thorough, and Luna explained itself carefully; neither matched Haiku's time.
If you're building a responsive research assistant, extra thinking time that doesn't produce a noticeably better answer is just waiting. A fairer follow-up would use matched settings, several tasks and more than one run per model. You can see the agents I've built so far on my projects page.