Can Levanto Sage 2 extract fields as well as GPT-6 Astra?

On our small synthetic bench it got 96.9% of 228 field values right on clean documents and 88.4% on messier ones. That says nothing certain about your documents and fields, which is why the audit compares it with GPT-6 Astra’s own output on your samples.

Which numbers here are measured and which are estimates?

Measured: Levanto Sage 2’s tokens per document and its field accuracy, from our bench run on 2026-11-03. Estimated: GPT-6 Astra’s tokens per call (700 input, 120 output, 300 reasoning). Prices are list prices checked 2026-11-03, without batch or cached-input discounts.

Should I just write a regex?

If your documents follow a fixed template, yes, try that first. On our templated bench documents hand-written regexes got 100% of fields. A model earns its cost when wording and layout vary.

What happens to fields the small model is unsure about?

In a hybrid you keep confident fields and ask GPT-6 Astra only for the rest. These estimates charge a full GPT-6 Astra call for every document not kept, so they are on the conservative side.

Extraction estimate · prices checked 2026-11-03

GPT-6 Astra vs Levanto Sage 2 for field extraction

List prices

Model	$ per M input	$ per M output	Source
GPT-6 Astra	$10	$50	OpenAI pricing
GPT-6 Luna	$0.1	$0.5	OpenAI pricing
Levanto Sage 2	$0.05	$1	Levanto list price

Checked 2026-11-03. Same table the public prices endpoint serves.

What we measured and what we assumed

Measured: Levanto Sage 2 used 329 input and 607 output tokens per document on average, over 36 synthetic lead, ticket and invoice documents (228 field values), in our bench run on 2026-11-03. Its output includes per-field probabilities, which is why it is larger than the JSON alone.
Assumed for GPT-6 Astra: 700 input tokens (a ~300-token document plus the audit’s default 400-token prompt with field definitions) and 120 visible output tokens for the JSON.
Our estimate: 300 reasoning tokens at GPT-6 Astra’s API default effort (medium), billed as output. Extraction gets 2.5× our yes/no estimate.

That gives an estimated $28.00 per 1,000 documents on GPT-6 Astra, against $0.624 on Levanto Sage 2.

How accurate was Levanto Sage 2 on our bench?

Set	Documents	Field values right	Lead / ticket / invoice
Clean templates	36	96.9% of 228	98.8% / 90% / 100%
Messier wording	15	88.4% of 95	82.9% / 88% / 94.3%

Weakest fields: ticket.amount_paid (58% on clean documents) and lead.budget (20% on the messier set). Money amounts and free-form budgets are where a fallback to GPT-6 Astra matters. Levanto Sage 2 left fields empty correctly 84.1% of the time when they were absent. This is a small synthetic bench, not a benchmark of your data.

Effort level changes the bill

Estimated GPT-6 Astra cost per 1,000 documents by effort level, with reasoning billed as output:

Effort	Est. reasoning tokens	Per 1,000 documents
minimal	25	$14.25
low	100	$18.00
medium (API default)	300	$28.00
high	800	$53.00
xhigh	2000	$113

If the fields are copied straight from the text, a lower effort may give the same JSON; test it on your documents. When lowering effort beats switching models.

Monthly estimate by volume

“Share kept” is a scenario: the share of documents where Levanto Sage 2 is confident on every field. The rest also go to GPT-6 Astra.

Documents / month	All on GPT-6 Astra	Hybrid, 50% kept	70% kept	90% kept
10,000	$280	$146	$90.24	$34.24
100,000	$2,800	$1,462	$902	$342
1,000,000	$28,000	$14,624	$9,024	$3,424

Estimates at the assumptions above. The 30% bar needs Levanto Sage 2 to keep at least 32% of documents. For yes/no calls on GPT-6 Astra, see GPT-6 Astra vs Levanto Sage 2 for classification.

Check it on your own documents

Paste 5 to 25 documents with the JSON GPT-6 Astra returned and your field definitions into the free cutover audit. It reports agreement per field, picks the confidence threshold and projects cost with your token counts. How it works · Other models · Is my task too simple for a big model?

Run the free extraction audit