The Vibe Father
Head to head
Not every seat needs a flagship. A lot of agentic work is scouting — fast reads, summaries, triage, cheap first drafts — and for that job you want a small, quick, inexpensive model, not the most expensive one you can afford. This matchup pits two popular scouts against each other: Claude Haiku 4.5 and Gemini 3.5 Flash. The surprise is how far apart they actually land once you read the benchmarks. We run both on our live benchmarks (VCI = SWE-bench 40 / Terminal-Bench 30 / LiveCodeBench 30). Here's the honest read on who deserves the scout seat.
Where Haiku 4.5 wins
It's a clean, fast, cheap Claude scout. Haiku 4.5 runs $1 per million input and $5 per million output, streaming at 96 tokens per second. For quick reads, triage, and cheap drafting inside a Claude-centric workflow, it's a capable little model that does the scout job at a scout price.
A fine small-model repo score. Haiku posts 66.6 on SWE-bench Verified. For a model this size and this cheap, that's respectable — enough to handle routine, well-specified edits where you're not asking it to untangle anything hard. As a first-pass drafter, it earns its keep.
Predictable Claude behavior. If your stack is already tuned around Claude's conventions and prompt style, Haiku slots in as the low-cost tier without you having to relearn how the model responds. That consistency has real workflow value.
Where Gemini 3.5 Flash wins
It's the most complete cheap model in existence. Flash posts 79.3 SWE-bench, 76.2 Terminal-Bench, and 87.6 LiveCodeBench — a full, strong slate across all three axes — at $1.50/$9, streaming at 167 tokens per second with a 1M-token context window. That's not a scout that only reads; it's a genuine all-rounder that happens to be cheap, and it beats Haiku on every published benchmark.
The agentic gap is enormous. This is the number that decides it. Flash posts 76.2 on Terminal-Bench — the benchmark for driving a shell through a multi-step task — versus Haiku's 35.5. That 35.5 is the lowest agentic score anywhere on our board. It tells you exactly what Haiku is not: an autonomous builder. Flash can drive a shell; Haiku, by this measure, largely cannot.
Contest strength too. Flash's 87.6 LiveCodeBench crushes Haiku's 41.2 on algorithmic problems. Across the board, the gap isn't close.
The numbers side by side
| Model | SWE-bench | Terminal-Bench | LiveCodeBench | In / Out per M | tok/s |
|---|---|---|---|---|---|
 Haiku 4.5 | 66.6 | 35.5 | 41.2 | $1 / $5 | 96 |
 Gemini 3.5 Flash | 79.3 | 76.2 | 87.6 | $1.50 / $9 | 167 |
Reading the split correctly
The temptation is to see two cheap models and assume they're interchangeable scouts. The benchmarks say otherwise. On the axis that defines an agentic scout — Terminal-Bench, driving a shell — Flash more than doubles Haiku's score. Haiku's 35.5 is a genuine cliff: it means you should never hand Haiku a multi-step autonomous task and walk away, because the evidence says it will struggle to hold the plot. Use Haiku for what it's good at — fast reads, summaries, triage, cheap drafts you'll review — and never as an unattended builder. Flash, by contrast, can genuinely double as a light builder, which is why it blurs the line between scout and all-rounder.
It's worth being fair to Haiku about what its numbers actually mean. A 66.6 SWE-bench isn't nothing — it's a real repo score, and for well-specified single-file edits Haiku will often produce a correct patch. The problem is specifically autonomy: the wide gap between its 66.6 SWE and its 35.5 Terminal-Bench tells you Haiku can frequently identify the right change but struggles to drive a shell through the many steps of applying, testing, and correcting it unattended. That's the signature of a model built to assist inside a loop you're steering, not to run one on its own. Read that way, Haiku isn't a weak model — it's a model pointed at a narrower job than Flash.
Price barely separates them, which makes the benchmark gap the whole story. Haiku is $1/$5 and Flash is $1.50/$9 — close enough that nobody should pick Haiku to save money if Flash does more. The only reason to reach for Haiku over Flash is workflow fit: a Claude-native stack where consistency across your models has real value. On raw capability per dollar, Flash wins comfortably.
Who should reach for Haiku 4.5
- You're all-in on Claude and want a cheap tier that behaves like the rest of your stack.
- Read-and-summarize scout work is the job — not autonomous shell tasks.
- Routine, well-specified edits where 66.6 SWE-bench is plenty and you're reviewing the output anyway. More in cheapest capable models.
Who should reach for Gemini 3.5 Flash
- You want one cheap model that does it all. The full published slate makes it the most complete budget pick.
- The scout seat sometimes builds. 76.2 Terminal-Bench means it can drive a shell when you need it to.
- Speed and context matter. 167 tok/s and a 1M window remove two real ceilings.
The honest close
Both models can sit in the scout seat, but they're not equals. Gemini 3.5 Flash wins every published benchmark and — crucially — can actually drive a shell, while Haiku 4.5 is best kept to fast reads and cheap drafts inside a Claude workflow. If you're building a routed, multi-model setup where each seat goes to the cheapest model that clears it, Flash is the stronger scout that occasionally doubles as a builder, and Haiku is the read-only helper you keep on a short leash. That per-task routing is exactly what a bring-your-own-key harness like The Vibe Father is for. See best model for each agent role and the full board at /benchmarks.
