# LLM Cutover: full text

> Paste 25 real examples of what your current LLM decides or extracts. Get agreement, a confidence-routing threshold and a monthly cost projection for a cheaper model, in under a minute. Free.

Source pages: https://llmcutover.com/, https://llmcutover.com/method/, https://llmcutover.com/docs/, https://llmcutover.com/llm-routing-calculator/, https://llmcutover.com/legal/.

## What it does
You paste 5-25 real inputs your current LLM already answered (yes/no decisions or extracted fields). The audit runs a cheaper model (Levanto Sage 2, $0.05 per 1M input tokens, $1 per 1M output) on the same task, blind, and reports:
- Agreement with your current model's answers.
- Confidence routing: for each threshold from 0.50 to 0.99, the share of calls the cheaper model keeps and the agreement on those. It recommends the threshold that keeps the most calls at your target agreement (98% by default).
- Cost projection: your monthly bill now vs a hybrid (cheaper model on every call + your current model on the share it doesn't keep). "Worth it" means at least 30% cheaper.
- A router code snippet.

## Method details
- Extraction values are compared by type: money within 0.5% and same currency, dates by day, text ignoring case and punctuation.
- Your current model's token counts are estimated (~4 characters per token). Reasoning tokens are billed at the output price; estimated per effort level: none 0, off 0, minimal 10, low 40, medium 120, high 320, xhigh 800, max 1600 (x2.5 for extraction).
- 25 samples is a screen, not a sign-off. Batch and cached-input discounts are not included.
- Samples are sent to the candidate model for the run only and not stored.

## Pricing
- Free audit: $0, up to 25 samples, 5 runs a day on the website.
- Full audit: $49 one-time, no subscription. 5 runs within 30 days, up to 500 yes/no or 200 extraction samples per run, the full disagreement list (JSON/CSV), per-field breakdown and router snippet. Seller: OPTIMUMTECH FUSION SRL, Romania. Terms and refunds: https://llmcutover.com/legal/.

## API (for agents and scripts)
- OpenAPI 3.1: https://llmcutover.com/openapi.json (same spec at https://api.llmcutover.com/openapi.json)
- POST https://api.llmcutover.com/api/v1/audit: JSON body `{mode, question|fields, samples, current, usage?, target?}`. Returns the full report as JSON.
- Anonymous, no key and no signup: 3 audits per day per client, up to 10 samples each, and a shared daily model budget of $0.02 for all anonymous callers (HTTP 503 with Retry-After when it runs out).
- Full-audit pass ($49, bought on https://llmcutover.com/#pricing): send it as `Authorization: Bearer <pass>`. 5 runs within 30 days, up to 500 yes/no or 200 extraction samples per run, disagreement list included.
- MCP (Streamable HTTP, stateless JSON): POST https://api.llmcutover.com/mcp. Tools: `run_cutover_audit` (same input as the API) and `list_models`. Same limits; the pass works as the Bearer token.
- GET https://api.llmcutover.com/api/models: model price table (public, cacheable). GET https://api.llmcutover.com/api/health: liveness.
- Errors: 400/422 bad input (message says what to fix), 401 bad pass, 413 body too large, 429 daily limit (Retry-After), 503 budget used up or temporarily paused.
- Samples are sent to the candidate model for that run only and not stored. Don't send personal data.
- Docs for humans: https://llmcutover.com/docs/

## Model prices used (list prices checked 2026-11-03)
- Claude Fable 5.1 (`claude-fable-5.1`): $10 / $50 per 1M input / output tokens; reasoning effort default: high
- Claude Opus 5.5 (`claude-opus-5.5`): $4 / $20 per 1M input / output tokens; reasoning effort default: medium
- Claude Sonnet 5.5 (`claude-sonnet-5.5`): $2 / $10 per 1M input / output tokens; reasoning effort default: high
- Claude Haiku 5.5 (`claude-haiku-5.5`): $0.1 / $0.5 per 1M input / output tokens; reasoning effort default: medium
- GPT-6 Astra (`gpt-6-astra`): $10 / $50 per 1M input / output tokens; reasoning effort default: medium
- GPT-5.6 Sol (`gpt-5.6-sol`): $4 / $20 per 1M input / output tokens; reasoning effort default: medium
- GPT-6.1 Sol (`gpt-6.1-sol`): $2 / $10 per 1M input / output tokens; reasoning effort default: medium
- GPT-6.1 Terra (`gpt-6.1-terra`): $1 / $5 per 1M input / output tokens; reasoning effort default: medium
- GPT-5 mini (`gpt-5-mini`): $0.25 / $2 per 1M input / output tokens; reasoning effort default: medium
- GPT-6 Luna (`gpt-6-luna`): $0.1 / $0.5 per 1M input / output tokens; reasoning effort default: medium
- Gemini 4 Argon (`gemini-4-argon`): $2 / $10 per 1M input / output tokens; reasoning effort default: medium
- Gemini 3.8 Flash (`gemini-3.8-flash`): $0.75 / $3.75 per 1M input / output tokens; reasoning effort default: medium
- Gemini 3.1 Flash-Lite (`gemini-3.1-flash-lite`): $0.25 / $1.5 per 1M input / output tokens; reasoning effort default: minimal
- Levanto Sage 2 (candidate): $0.05 / $1

## FAQ
### Which cheaper model does it test?
Today: Levanto Sage 2, a classify/extract model priced at $0.05 per million input tokens that returns calibrated probabilities. The audit code is model-swappable; more candidates are being added.

### Do you store my samples?
No. Samples are sent to the candidate model to run the audit, then dropped. Nothing is written to a database, and the report lives only in your browser. Don’t paste personal or sensitive data: the candidate model is a US service.

### Why compare against my current model and not ground truth?
Because that is the cutover question: will switching change the answers your product already relies on? If you have human-checked labels, paste those instead and the same numbers become accuracy.

### Is 25 samples enough?
It’s enough to rule a switch in or out, not to sign it off. One disagreement in 25 moves agreement by 4 points. The full audit (up to 500 yes/no or 200 extraction samples per run) narrows that.

### What if my current model is already cheap?
Then the report says so. If you already run a small model at about $0.10 per million input tokens, a switch rarely saves enough to matter, and the audit will mark it “not worth it”.

## Guides
- Reduce LLM costs without changing your answers: https://llmcutover.com/guides/reduce-llm-costs/
- A cheaper alternative to GPT for classification: https://llmcutover.com/guides/cheaper-alternative-to-gpt-for-classification/
- How to switch from GPT or Claude to a smaller model: https://llmcutover.com/guides/switch-to-a-smaller-model/
- How to choose an LLM routing confidence threshold: https://llmcutover.com/guides/llm-routing-confidence-threshold/
- Is my LLM task too simple for GPT or Claude?: https://llmcutover.com/guides/is-my-llm-task-too-simple/
- What reasoning effort costs per call: https://llmcutover.com/guides/reasoning-effort-cost/
- LLM routing calculator: https://llmcutover.com/llm-routing-calculator/

## Savings estimates by model (https://llmcutover.com/savings/)
Estimates, not bills: 500 input + 5 output tokens per yes/no call, plus our estimated reasoning tokens at each model's API-default effort; Levanto Sage 2 charged 500 in / 2 out. List prices checked 2026-11-03. "Keep" = share of calls the cheap model must keep for the hybrid to be 30% cheaper.
- Claude Sonnet 5.5: ~$4.25 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 31%. https://llmcutover.com/savings/claude-sonnet-5-5-to-sage-2/
- Claude Opus 5.5: ~$4.50 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 31%. https://llmcutover.com/savings/claude-opus-5-5-to-sage-2/
- GPT-6.1 Sol: ~$2.25 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 31%. https://llmcutover.com/savings/gpt-6-1-sol-to-sage-2/
- GPT-6.1 Terra: ~$1.13 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 32%. https://llmcutover.com/savings/gpt-6-1-terra-to-sage-2/
- GPT-5 mini: ~$0.375 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 37%. https://llmcutover.com/savings/gpt-5-mini-to-sage-2/
- Gemini 4 Argon: ~$2.25 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 31%. https://llmcutover.com/savings/gemini-4-argon-to-sage-2/
- Claude Fable 5.1: ~$21.25 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 30%. https://llmcutover.com/savings/claude-fable-5-1-to-sage-2/
- GPT-6 Astra: ~$11.25 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 30%. https://llmcutover.com/savings/gpt-6-astra-to-sage-2/
- Gemini 3.8 Flash: ~$0.844 per 1,000 calls vs $0.027 on Levanto Sage 2; keep >= 33%. https://llmcutover.com/savings/gemini-3-8-flash-to-sage-2/

## Field extraction estimates by model (https://llmcutover.com/savings/extraction/)
Estimates for the current model (700 input + 120 output tokens per document plus our reasoning estimate x2.5); Levanto Sage 2 measured on our synthetic bench (2026-11-03): 329 in / 607 out tokens per document, 96.9% of 228 field values on clean documents, 88.4% on messier ones.
- Claude Sonnet 5.5: ~$10.60 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 36%. https://llmcutover.com/savings/extraction/claude-sonnet-5-5-to-sage-2/
- Claude Opus 5.5: ~$11.20 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 36%. https://llmcutover.com/savings/extraction/claude-opus-5-5-to-sage-2/
- GPT-6.1 Sol: ~$5.60 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 41%. https://llmcutover.com/savings/extraction/gpt-6-1-sol-to-sage-2/
- GPT-6.1 Terra: ~$2.80 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 52%. https://llmcutover.com/savings/extraction/gpt-6-1-terra-to-sage-2/
- Gemini 4 Argon: ~$5.60 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 41%. https://llmcutover.com/savings/extraction/gemini-4-argon-to-sage-2/
- Claude Fable 5.1: ~$53.00 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 31%. https://llmcutover.com/savings/extraction/claude-fable-5-1-to-sage-2/
- GPT-6 Astra: ~$28.00 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 32%. https://llmcutover.com/savings/extraction/gpt-6-astra-to-sage-2/
- Gemini 3.8 Flash: ~$2.10 per 1,000 documents vs $0.624 on Levanto Sage 2; keep >= 60%. https://llmcutover.com/savings/extraction/gemini-3-8-flash-to-sage-2/

Contact: contact@llmcutover.com