Reducto made its document editor free in ChatGPT as Unstructured pays rivals' customers to switch
Reducto made its editing engine free inside ChatGPT and Codex on October 29 while Unstructured pays rivals' customers to switch. A benchmark showing frontier agents scoring 12% on real work explains why both are fighting over the same cent per page.
Vincent Jiang · 2 min read
On October 29, Reducto gave its editing engine away 1. Reducto Office, a free plugin released that day, lets AI agents edit Word, PowerPoint and Excel files inside ChatGPT and Codex, with results reviewed in a live editor. The licensed, embeddable editor SDK stays a separate paid product. The giveaway buys distribution inside the two biggest agent products; Reducto still charges for the pages those agents read 1.
Seven billion pages built the funnel
Adit Abraham, Reducto's co-founder and CEO, spent three years building the layer that turns a PDF into data an agent can act on. His engine crossed seven billion pages processed in October and now reads billions a month at a list price of one cent a page, feeding customers from Harvey$15.5B — Harvey, private, latest valuation $15.5B to the largest hedge funds 23. The logic of the giveaway is simple: once every agent reads through Reducto, editing is the hook that keeps it there.
Rivals are bidding for the same cent
Reading documents is not settled. Unstructured, whose open source stack has passed 100 million downloads, charges 1.5 cents a page for its Transform parser, says frontier models make up to 2.3 times more table errors on 224 real business pages, and is offering up to $10,000 in Transform credits to any team defecting from Reducto, LlamaParse or a frontier model, through December 31, 2026 4. It even scored 0.899 on RD-TableBench, a benchmark Reducto itself built 4. The layer is being fought over on price, in both directions.
The 12% benchmark is the whole market
Why the layer matters at all: SemiAnalysis, the semiconductor and AI research firm, published its GDPx benchmark on November 3, dropping 13 frontier agents into 206 simulated professional tasks 5. The best, Claude Opus 5.5, scored 12.0% at $51.41 a task; legal redlining returned zero on every rollout 5. Agents cannot yet read a deal room unaided, so somebody owns the reading.
Frontier models cost about 7.5 times the specialist parser per page
The document layer now costs less than the model it feeds
- Estimate
Data
| Value | |
|---|---|
| Reducto | 0.01 per page |
| Unstructured Transform | 0.02 per page |
| GPT-6 (direct) (estimate) | 0.11 per page |
The counterpoint is Datasite's global chief technology officer Raj Bakhru, who moved ingestion for 55,000 data rooms and more than 500,000 monthly users to Reducto and saw recall rise 22% from parsing alone 6. That validates the layer at scale.
It also shows the other side of the trade: enterprises can swap vendors when the bill changes, and Bakhru's team still benchmarks parsing quality on every change 6. Validation is not loyalty.
Watch December 31, the day Unstructured's credit offer lapses 46. The defection count between now and then tells you whether the document layer is a moat or a commodity 46.
Best agent scores 12% on real professional work; legal is zero
Frontier agents pass barely one task in eight, and none passes legal
Data
| Value | |
|---|---|
| Claude Opus 5.5, overall | 12% |
| Gemini 4 Carbon, overall | 10.9% |
| GPT-6 Astra, overall | 9.6% |
| Legal redlining, all models | 0% |
Get Platforms by email, free
The cloud, data and developer software AI is built on. In our daily email, every weekday.
