SYFTLY/ answers
light estimateLast updated 2026-08-20

Mistral AI vs Z.ai

Mistral OCR 4 vs GLM-OCR for ocr & document extraction. Score: 85.66 vs 95.22 — edge Z.ai Max pages: 1000 vs 100 — edge Mistral AI Computed from public benchmarks with dated sources; updated 2026-08-20.

Head to head
Mistral OCR 4 compared with GLM-OCR per decision axis
AxisMistral OCR 4GLM-OCREdge
Price$4 per 1000 pages—*—
Score85.6695.22Z.ai
Max pages1000100Mistral AI
tables✓✓—
handwriting✓✓—
JSON✓✓—

* token-/credit-priced — the headline understates real per-unit cost, so no price edge is awarded.

Strengths & caveats
Mistral AIDedicated single-model OCR API (OCR 4, released 23 June 2026) with markdown plus HTML-table output, paragraph-level bounding boxes, block classification and inline confidence scores, 170 languages across 10 language groups, and a 50% batch discount ($2/1000). OCR 4.1 (public preview since 16 July 2026) adds structural block labels and block-level confidence; annotated Document AI extraction is $5/1000 pages. The OCR 4 release doubled the price from $2 to $4 per 1000 pages, making it the most expensive comparable per-page option here — the hyperscalers' basic OCR tier is $1.50/1000. The 85.66 OmniDocBench composite is the independent v1.6 leaderboard's untagged 'Mistral OCR' entry (April 2026), so it predates OCR 4; Mistral self-reports 93.07 on OmniDocBench, 85.20 on OlmOCRBench and a 72% average human-preference win rate for OCR 4, all Tier 3. The 1000-page async cap is carried over from the June 2026 check and could not be re-verified.Z.aiHighest published independent OmniDocBench composite of any hosted API in this table (95.22 on v1.6; Z.ai cites 94.62 on v1.5), callable at api.z.ai/api/paas/v4/layout_parsing with text, markdown and image-link output, HTML table reconstruction and handwriting support; the underlying 0.9B model is also self-hostable, and throughput is quoted at 1.86 PDF pages/second. Priced per token ($0.03 per million tokens for both input and output), not per page, so no comparable $/1000-pages rate exists without estimating — excluded from the price axis. Tight caps (100 pages per document, PDF 50MB, single image 10MB) rule it out for large documents, the documented language list is narrower than the hyperscalers', and China-hosted inference may raise data-residency questions for EU/US buyers.
Sources