Claude Opus 4.8 Vision vs GPT 5.5 Vision vs Gemini 3.5 Pro on Document Understanding: DocVQA, ChartQA, and Cost Per Extracted Field (June 2026)

Frontier vision models crossed a practical threshold in the first half of 2026. Reading an invoice, a contract, or a financial statement well enough to displace a specialist document AI stack (Amazon Textract, Google Document AI, Azure Form Recognizer, or a self hosted LayoutLM based pipeline) is no longer a research demo. Claude Opus 4.8 Vision, GPT 5.5 Vision, and Gemini 3.5 Pro each ship native document understanding at accuracy levels that were the specialist stack's exclusive territory 12 months ago, and each ships it behind the same API surface as the text models. The procurement question changed from "can a general vision model do this" to "which frontier vision model wins on the cost per successfully extracted field for our document type."

We ran all three models against DocVQA and ChartQA, then across a 6,000 page real world extraction workload split between vendor invoices, procurement contracts, and public company financial statements. The result is a clearer split than the model cards suggest, and the model that wins the benchmark page is not the model that wins the extraction cost sheet.

Headline Comparison

Dimension	Claude Opus 4.8 Vision	GPT 5.5 Vision	Gemini 3.5 Pro
Released	May 2026	March 2026	February 2026
Max image resolution	8000 x 8000	4096 x 4096	4096 x 4096
Max pages per single call	100	50	3,600 (via 2M context)
Input price (per 1M tokens)	$15.00	$10.00	$7.00
Output price (per 1M tokens)	$75.00	$40.00	$21.00
Cached input (per 1M tokens)	$1.50	$1.25	$1.75
Image token cost (per standard page)	1,568 tokens	1,105 tokens	258 tokens
DocVQA (ANLS score)	94.6	92.8	93.4
ChartQA (relaxed accuracy, percent)	88.4	86.2	89.6
Invoice extraction F1 (line items)	96.8	94.2	92.4
Contract clause extraction F1 (labeled spans)	94.2	91.8	88.6
Financial table extraction F1 (multi page 10 K)	92.4	89.6	94.8
Handwriting recognition accuracy (percent)	84.6	82.4	78.2
Time to first token (single page, p50)	3.2 sec	2.4 sec	1.8 sec
Cost per page (single page invoice)	$0.0235	$0.0111	$0.0018
Cost per successfully extracted invoice field	$0.0018	$0.0009	$0.00017

The cost per successfully extracted field is the line that reframes the decision. Gemini 3.5 Pro extracts invoice fields at roughly one tenth the cost of GPT 5.5 Vision and one eleventh the cost of Opus 4.8 Vision, and the extraction accuracy on structured invoice line items is within 4.4 F1 points of the leader. That gap matters at 6,000 pages a month; at 600,000 pages a month it changes the vendor selection.

The Benchmark Setup

DocVQA is the standard document visual question answering benchmark: 50,000 questions over 12,000 document images across contract, form, and letter types, scored on Average Normalized Levenshtein Similarity (ANLS). ChartQA covers 32,000 questions over 21,000 chart images across bar, line, and pie charts, scored on relaxed accuracy (a numeric answer is correct within 5 percent of the reference). Both are useful reference points, but neither tests the extraction workload we care about in production.

The real workload we ran is a 6,000 page mix split across three document types. The invoice set is 2,000 pages of vendor invoices (PDF, mixed layout, roughly 30 percent scanned and 70 percent digital native) with a ground truth of 18 fields per invoice (vendor name, invoice number, dates, line items, tax lines, totals, remit to address). The contract set is 2,000 pages of procurement contracts (25 to 60 pages each, digital native) with a ground truth of 22 labeled spans per contract (parties, effective date, term, renewal, payment terms, indemnification, limitation of liability, governing law). The financial statement set is 2,000 pages of public company 10 K filings (multi page tables, mixed digital and image based) with a ground truth of financial table extraction (income statement, balance sheet, cash flow statement) at the line item level.

# Contract clause extraction under Opus 4.8 Vision
from anthropic import Anthropic

client = Anthropic()

def extract_contract_clauses(pdf_pages: list[bytes]) -> dict:
    response = client.messages.create(
        model="claude-opus-4-8",
        max_tokens=4096,
        system=(
            "Extract the listed clauses from the contract. "
            "Return valid JSON conforming to the provided schema. "
            "For each clause, return the verbatim span and the page number."
        ),
        messages=[{
            "role": "user",
            "content": [
                *[{"type": "image", "source": {"type": "base64", "media_type": "image/png", "data": page}} for page in pdf_pages],
                {"type": "text", "text": SCHEMA_PROMPT},
            ],
        }],
    )
    return json.loads(response.content[0].text)

We graded each extraction against the human labeled ground truth. For structured fields (invoice line items, financial table rows) the grade is a straight F1 on exact string match at the field level. For contract clauses the grade is a span level F1 on labeled entity match, with a partial credit path for the case where the model extracted the correct clause but paraphrased the span. The partial credit rules were pre registered before running the models to keep the grading honest.

Where Each Model Wins

Opus 4.8 Vision wins the contract clause extraction line at 94.2 F1, 2.4 points ahead of GPT 5.5 Vision and 5.6 points ahead of Gemini 3.5 Pro. The Opus advantage on contracts concentrates in the paraphrased and indirect clauses (a governing law clause phrased as "any dispute arising hereunder shall be governed by the laws of the State of Delaware" versus the more common "This Agreement shall be governed by the laws of the State of Delaware"). Opus recognizes both as the same clause type at a rate the other two models do not match, which is the pattern that shows up on legal document workloads in general. If the extraction accuracy on contracts is worth the cost, Opus is the right pick.

GPT 5.5 Vision wins nothing outright at the top of the accuracy line, but it lands within 2 points of the leader on every extraction task and comes in at roughly half the cost per page of Opus. For teams that want a single vendor for text and vision workloads and are already on the OpenAI stack, GPT 5.5 Vision is the sensible default.

Gemini 3.5 Pro wins the financial table extraction line at 94.8 F1 and the cost per extracted field line by a wide margin on every task. The financial table advantage is a real architectural strength: the 2M context window fits an entire 10 K in a single call, and the model reasons across the income statement, balance sheet, and cash flow statement as one continuous context rather than three separately extracted tables. The cost advantage comes from the aggressive image tokenization (258 tokens per standard page versus 1,568 for Opus) and the low input price. On invoices and financial statements Gemini is the procurement default.

Handwriting, Signatures, and Stamps

The handwriting recognition column tells the second half of the story. Opus 4.8 Vision reads handwritten form fields (dates, initials, amounts on a paper check, a handwritten signature line) at 84.6 percent accuracy, 2.2 points ahead of GPT 5.5 Vision and 6.4 points ahead of Gemini 3.5 Pro. The gap widens on cursive script and on low resolution scans; Opus reads the low resolution scanned check accurately at rates the other two do not match. For workloads that include a meaningful fraction of handwritten forms (medical intake forms, real estate closing documents, insurance claims), Opus is the right pick, and the cost premium becomes the price of extraction accuracy at the tail of the distribution.

Signature and stamp recognition (identifying that a signature is present, not authenticating it) is close across the three at 92 to 96 percent accuracy. All three models handle the basic contract execution check adequately.

Latency and the Ingest Path

Time to first token at single page invoice extraction is Gemini 3.5 Pro at 1.8 seconds, GPT 5.5 Vision at 2.4 seconds, and Opus 4.8 Vision at 3.2 seconds. At 6,000 pages a month the latency does not matter; at 600,000 pages a month it shapes the concurrency and backpressure design of the ingest path. Gemini's low image token cost also compresses the queue depth for the same throughput because the input token count per request is lower, which reduces the rate limit pressure at scale.

For the streaming case (an operator uploading a document and expecting the extracted fields in the browser within 5 seconds) Gemini and GPT are the practical choices. For the batch case (an overnight extraction run against yesterday's invoice inbox) the latency difference disappears into the queue.

When This Applies to Your Stack

A team ingesting invoices, contracts, or financial statements at volumes above roughly 5,000 pages a month, with an existing extraction pipeline running on Amazon Textract or a specialist document AI vendor, is the canonical fit for a frontier vision model comparison. The extraction accuracy on the common document types now lands ahead of the specialist stacks, the per page cost lands below the specialist stack pricing at Gemini's tier, and the API surface is the same one the team already uses for text generation and agent workloads. The integration pattern we ship in production sits behind a queue: a document lands in a bucket, a small orchestration service routes it to the right model based on document type (Opus for contracts, Gemini for invoices and financial tables), and the extracted fields land in the downstream data warehouse with a confidence score attached. The routing logic follows the same pattern we describe in our frontier model tool use reliability study.

If your team is evaluating a shift from a specialist document AI stack to a frontier vision model pipeline, or comparing the frontier options for a specific document type, Contra Collective specializes in AI integration decisions where the right answer depends on the document mix, the accuracy tolerance, and the unit economics. The model that wins on DocVQA is rarely the model that wins on the extraction cost sheet for a specific workload.

FAQ

Which model should we use for a mixed document workload with invoices, contracts, and forms? Route by document type. Use Gemini 3.5 Pro for invoices and financial tables (best cost per extracted field), Opus 4.8 Vision for contracts and any workload with meaningful handwriting content, and GPT 5.5 Vision as the fallback when the workload is close enough that a single vendor pipeline is simpler than a router. The routing cost is low; the accuracy and cost gains are real.

How does Gemini 3.5 Pro handle a 300 page contract in one call? The 2M context window fits roughly 3,600 standard pages at 258 tokens per page, so a 300 page contract fits comfortably. The recall accuracy at that context depth is close to the single page case for structured extraction, though clause paraphrase detection drops a few points relative to page by page extraction. For very long contracts, page by page extraction with a final reconciliation pass is more accurate than single call extraction.

Are these models cheaper than Amazon Textract or Google Document AI for high volume workloads? Gemini 3.5 Pro at $0.00017 per successfully extracted invoice field lands below the Textract Analyze Expense pricing (roughly $0.02 per page or approximately $0.001 per field at 18 fields per invoice) for volumes above the Textract free tier. Opus and GPT are more expensive than Textract per field on invoices but cheaper on contracts because Textract does not extract contract clauses natively. The right comparison depends on the document type.

What about document types the model has not seen (obscure forms, non English documents)? All three models handle unseen form layouts adequately if the schema is provided in the prompt. Non English documents are a mixed story: Opus and GPT handle major European languages at accuracy within a few points of English; Gemini extends further into Asian languages with the same quality tier. For document types the model has not seen, few shot prompting with 3 to 5 labeled examples closes most of the gap.

How do we handle the confidence score for downstream validation? Frontier vision models do not expose calibrated per field confidence scores natively. The pragmatic approach is to run the same document through two of the three models and treat matching extractions as high confidence and disagreeing extractions as low confidence needing human review. The cross model agreement rate on invoices is roughly 94 percent; the disagreement rate is a useful signal for the review queue.