Workload Benchmark & Cost Model
High-Volume JSON Extraction & ETL LLM Cost Comparison
Structured ETL and invoice/document extraction pairs a strict JSON schema & few-shot prefix (1,200 tokens) with incoming raw payloads (1,800 dynamic tokens) and compact JSON output (250 tokens) at 10,000 requests/day—maximizing both cache reuse and Batch savings.
Extraction (JSON / ETL) Cost Comparison (20 models)
100% Client-Side · Zero Prompt TransmissionAvg input/turn: 3,000 tok · Daily turns: 10,000 · Registry verified: 2026-11-12
| Model & Host | Active Tier & Rules | Eff. In / Out ($/1M) | Cache Read ($/1M) | Daily | Monthly (30d) | $/1K Turns | Savings vs Sync |
|---|---|---|---|---|---|---|---|
LOWESTClaude Haiku 5.5 Anthropic·Anthropic First-Party·1000K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 100K) | $0.050 $0.250 | $0.0050 | $1.64 | $49.21 | $0.164 | -61.4% $78/mo saved |
GPT-6 Luna OpenAI·OpenAI First-Party·922K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 272K) | $0.050 $0.250 | $0.0050 | $1.64 | $49.23 | $0.164 | -61.4% $78/mo saved |
Mistral Small 4 (2603)Open-Weight Mistral·Mistral La Plateforme·256K max in Verified 2026-11-12Official Docs ↗ | Batch -50% Standard (<= 256K)
| $0.075 $0.300 | — | $3.00 | $90.00 | $0.300 | -50.0% $90/mo saved |
GPT-OSS 120B (Groq)Open-Weight Groq·Groq LPU·131K max in Verified 2026-11-12Official Docs ↗ | Batch -50% Standard (<= 131K)
| $0.075 $0.300 | — | $3.00 | $90.00 | $0.300 | -50.0% $90/mo saved |
GPT-OSS 120B (Amazon Bedrock)Open-Weight AWS Bedrock·Amazon Bedrock (US Regions)·128K max in Verified 2026-11-12Official Docs ↗ | Batch -50% US Standard (<= 128K)
| $0.075 $0.300 | — | $3.00 | $90.00 | $0.300 | -50.0% $90/mo saved |
Llama 4 Maverick 17B (Amazon Bedrock)Open-Weight AWS Bedrock·Amazon Bedrock (US Regions)·1000K max in Verified 2026-11-12Official Docs ↗ | Batch -50% US Standard (<= 1M)
| $0.120 $0.485 | — | $4.81 | $144.38 | $0.481 | -50.0% $144/mo saved |
GPT-OSS 120B (Together AI)Open-Weight Together AI·Together AI Serverless (MXFP4)·131K max in Verified 2026-11-12Official Docs ↗ | Batch -0% Standard (<= 131K)
| $0.150 $0.600 | — | $6.00 | $180.00 | $0.600 | 0.0% |
DeepSeek Flash (V4.1-Flash-1017)Open-Weight DeepSeek·DeepSeek First-Party·1000K max in Verified 2026-11-12Official Docs ↗ | Cache On Standard (<= 1M)
| $0.300 $1.200 | $0.0060 | $8.82 | $264.74 | $0.882 | -26.5% $95/mo saved |
DeepSeek V4.1 Flash (Together AI)Open-Weight Together AI·Together AI Serverless (FP8)·1000K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -0% Standard (<= 1M)
| $0.300 $1.200 | $0.0060 | $8.82 | $264.74 | $0.882 | -26.5% $95/mo saved |
Mistral Large 4Open-Weight Mistral·Mistral La Plateforme·1000K max in Verified 2026-11-12Official Docs ↗ | Batch -50% Standard (<= 1M)
| $0.350 $1.050 | — | $13.12 | $393.75 | $1.312 | -50.0% $394/mo saved |
Google·Google AI Studio / Gemini API·1049K max in Verified 2026-11-12Official Docs ↗ | Batch -50%Promo ends 2026-12-31 Intro Promo (<= 1M, thru 2026-12-31)
| $0.375 $1.875 | — | $15.94 | $478.13 | $1.594 | -50.0% $478/mo saved |
Anthropic·Anthropic First-Party·1000K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 1M) | $1.000 $5.000 | $0.0500 | $32.27 | $968.06 | $3.227 | -62.0% $1,582/mo saved |
OpenAI·OpenAI First-Party·922K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 272K) | $1.000 $5.000 | $0.0500 | $32.28 | $968.50 | $3.228 | -62.0% $1,582/mo saved |
Qwen3.8 27B Preview (Groq)Open-Weight Groq·Groq LPU·131K max in Verified 2026-11-12Official Docs ↗ | Standard (<= 131K)
| $0.800 $4.000 | — | $34.00 | $1,020.00 | $3.400 | 0.0% |
DeepSeek V4 Pro (0813)Open-Weight DeepSeek·DeepSeek First-Party·1000K max in Verified 2026-11-12Official Docs ↗ | Cache On Standard (<= 1M)
| $1.320 $3.960 | $0.0440 | $35.72 | $1,071.58 | $3.572 | -27.8% $413/mo saved |
Google·Google AI Studio / Gemini API·1049K max in Verified 2026-11-12Official Docs ↗ | Batch -50%Promo ends 2026-11-30 Intro Promo (<= 1M, thru 2026-11-30)
| $1.000 $5.000 | — | $42.50 | $1,275.00 | $4.250 | -50.0% $1,275/mo saved |
Gemini 3.1 Pro Preview Google·Google AI Studio / Gemini API·1049K max in Verified 2026-11-12Official Docs ↗ | Batch -50% Standard (<= 200K)
| $1.000 $6.000 | — | $45.00 | $1,350.00 | $4.500 | -50.0% $1,350/mo saved |
Claude Sonnet 5.5 (Amazon Bedrock) AWS Bedrock·Amazon Bedrock (Cross-Region Inference)·1000K max in Verified 2026-11-12Official Docs ↗ | Cache On Global CRIS (<= 1M)
| $2.000 $10.000 | $0.1000 | $64.54 | $1,936.13 | $6.454 | -24.1% $614/mo saved |
Claude Fable 5.2 Anthropic·Anthropic First-Party·1000K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 1M) | $4.000 $20.000 | $0.1000 | $128.00 | $3,839.86 | $12.800 | -62.4% $6,360/mo saved |
OpenAI·OpenAI First-Party·922K max in Verified 2026-11-12Official Docs ↗ | Cache OnBatch -50% Standard (<= 272K) | $5.000 $25.000 | $0.5000 | $164.12 | $4,923.48 | $16.412 | -61.4% $7,827/mo saved |
Budgeting & Tokenizer Disclaimer: Estimates assume USD list pricing verified on 2026-11-12. Actual token counts vary across provider tokenizers (for example, Claude 4.7+ uses a newer tokenizer that generates ~30% more tokens on identical prose than legacy models), and reasoning models bill hidden thinking tokens as output tokens.
Token Anatomy & Scaling Assumptions
- Cacheable Prefix (1,200 tokens): JSON Schema, field validation rules, edge-case enum definitions, and 2–3 canonical extraction examples.
- Dynamic Input (1,800 tokens): Raw HTML, email body, OCR invoice text, or support ticket payload.
- Completion Output (250 tokens): Deterministic structured JSON payload.
- Daily Scale: 10,000 requests/day across 24 hours/day (300,000 requests/month).
Worked Note: At 10,000 req/day (1,200 prefix / 1,800 dynamic / 250 output, 90% hit, 1 write/hr, 1h TTL, Batch ON), Claude Haiku 5.5 (50% Batch stacking with 1h caching) costs $1.64/day ($49.21/mo vs. $127.50/mo sync uncached—61.4% savings) and GPT-6 Luna stacks 50% Batch with caching for $1.64/day ($49.23/mo). Mistral Small 4 (50% Batch, prompt caching disabled in Batch) costs $3.00/day ($90.00/mo vs. $180.00/mo). By contrast, Together AI's DeepSeek-V4.1-Flash charges standard rates in Batch (0% discount, caching active), landing at $8.82/day ($264.74/mo).
Caching & Batch Strategy
Because all 10,000 daily jobs share the exact same 1,200-token schema prefix, cache hit rate defaults to 90% with 1 cold write/hour (24 writes/day) on a 1h TTL. Even with Anthropic's 2.0x 1-hour write rate ($0.20/1M on Haiku 5.5), 24 daily writes cost pennies while 90% cache reads save substantially on 360M monthly prefix tokens.
ETL pipelines are the textbook Batch API workload: 24-hour turnaround slashes input and output costs by 50% on Anthropic (stacking with 1h caching), OpenAI GPT-5/6.x (stacking with caching), Mistral (50% off without Batch caching), and Amazon Bedrock open-weight models like gpt-oss-120b (50% off without caching). Avoid Together AI for serverless chat Batch if expecting a discount (0% Batch discount), and remember Groq Batch does NOT stack with prompt caching.