More accurate and 31×–228× cheaper: grounding, benchmarked
We asked four models fifteen live-data questions three ways — alone, with web search, and grounded through Pipeworx. Grounding made every model more accurate and cheaper at the same time. The receipts are public.
There’s an assumption baked into most AI budgets: better answers cost more. Bigger model, higher token rate, better result. For reasoning and writing, that’s often true. For questions about the live world — what a company just filed, what a rate is today, what an agency approved this week — we kept seeing something different in our own usage, so we measured it.
We took fifteen questions with verifiable, current answers — the kind agents get asked every day: a figure from the latest SEC filing, a central-bank rate, a drug approval, an earthquake, a market price. Ground truth came from the primary sources themselves, captured the same day. Then we asked four models — Claude Fable 5, Claude Opus 5, GPT-5, and DeepSeek V4 Flash — each question three ways: the model alone, the model with a web-search tool, and the model connected to Pipeworx over MCP.
The full run, per-question receipts included, is at pipeworx.io/benchmark. Here’s the shape of it.
Alone, frontier models got 1–2 of 15
Not because they’re bad models — because the answers aren’t in the weights. Most attempts were honest refusals: “I don’t have real-time access,” “check EDGAR.” Honest, but not an answer. If your product needs live facts, model scale is not the lever.
Web search helped — at a price
With a search tool, scores rose to 6–10 of 15. Real improvement. It also cost up to $17.98 for fifteen questions on the most expensive arm, because search stuffs pages into an expensive context window and hopes the right line is in there somewhere.
Grounded, every model got better and cheaper at once
Connected to Pipeworx, models scored 12–13 of 15 — above their own web-search runs — while spending a fraction of the money. Claude Fable 5 grounded was 16× cheaper than Claude Fable 5 with web search, and more accurate. The gateway routes the question to the right primary source, makes the call, and hands the model a compact, cited record instead of a pile of pages.
And the result that reframes the budget conversation: grounded, the cheapest model matched the most expensive one — 13/15 for $0.0789 against 13/15 for $1.15. Once the facts come from the source, the model’s job shrinks to reading and judgement, and a small fast model reads records just fine. That’s the 31×–228× in the title: the cost gap between grounded answers and search-assembled ones, with our margin included at the least favorable tier.
The part that decays
Four of the fifteen true answers changed overnight during the run — a new filing landed, a rate moved. Only the grounded arms noticed. A benchmark score is a snapshot; freshness is a property. Answers built on live records stay right after the blog post about them is published.
What we’d tell a builder
Pick your model for what models are good at — reasoning, judgement, writing. Pick grounding for facts, and stop paying frontier token rates for recall that no model can do. A routed, cited answer through the gateway costs about half a cent, and every response carries _meta.source, _meta.fetched_at, and a pipeworx:// citation URI your agent can show its user as evidence.
Caveats, because we’d rather you trust the next number too
Fifteen questions is enough to see cost gaps of this size; it is not a precise accuracy claim. The first experiment’s arms were re-graded blind by an independent judge with labels stripped and order shuffled — 91.5% agreement — and the MCP arms haven’t had that treatment yet. GPT-5’s grounded run hit a tool-loop cap on three questions, so its denominator is 12. The question set deliberately measures the freshness gap, not general intelligence. All of it, including the raw rows, is on the benchmark page.
See it yourself
The free model on pipeworx.io/try is the same grounded setup that scored 13 of 15. Ask it something that changed this week — no signup, and check the citation that comes back. Or point your agent at the gateway and ask through your own stack.