Parallel just shipped Fast — a new web-search mode priced at $1 per 1,000 results, sitting right on the quality-versus-cost Pareto frontier. It's built specifically to pair with today's class of cheap, capable models: GPT-5.6 Luna, DeepSeek V4 Pro, Qwen3.8 27B.
What makes it interesting
The economics of agents have shifted. Models keep getting cheaper — OpenAI just cut GPT-5.6 Luna's price by 80%, and DeepSeek V4 Flash, Qwen3.8 27B, and MiniMax M3 lead usage on OpenRouter. But web search is often a huge share of total agent cost — unless you use Parallel. The numbers:
| Search provider | Share of total agent cost |
|---|---|
| Parallel (Fast) | under 12% |
| Brave Search | 48% |
| Exa Search Fast | 48% |
| Tavily Search Basic | 68% |
That's 2.2x to 2.8x cheaper end-to-end than the alternatives, for a search that's #3 on the Artificial Analysis Search Index (intelligence 73) and #1 for speed-per-task (~700ms).
The speed / cost / quality balance
Fast mode sits in the sweet spot:
- ~700ms latency — #1 on Artificial Analysis for speed per task
- #3 on intelligence (73) — well ahead of "good enough" for most agent work
- $1 / 1,000 results — 10x cheaper than frontier search, 5x cheaper than other APIs
For comparison, the full Parallel lineup:
| Mode | Latency | Price | # on Artificial Analysis | When to use |
|---|---|---|---|---|
| Turbo | ~250ms | $1/1000 | — | Latency-critical (voice, autocomplete, RAG pre-filter) |
| Fast | ~700ms | $1/1000 | #3 (73) | Best balance — most agent workflows |
| Advanced | 3s | $5/cpm | #1 (75) | Deep investigation, code review, synthesis |
Why it matters for agents
The interesting trend isn't at the frontier — it's that cost-per-intelligence keeps dropping. Over a few months:
- Models scoring ≥60 Intelligence Index: cost down 8.5x
- Models scoring ≥50 Intelligence Index: cost down 12.5x
Cheap models encourage more use, but they also change the cost dynamics of end-to-end agentic work. If search is 48-68% of your spend, a cheap search engine isn't a nice-to-have — it's the difference between a viable agent and one that's too expensive to run.
What I use it for
I run a DeepSeek-based agent. The Fast mode pairs naturally with the cheap-model stack — fast, accurate enough for everyday lookups and research, and at $1/1000 it keeps the search slice of my cost under 12%. For the rare deep-dive (investment research, synthesis), I'd reach for Advanced instead.
Source: Parallel email announcement, August 2026.
