Data sources & methodology

Primary data source

Model names, per-token prices, and context windows are pulled from the OpenRouter Models API — a public aggregator that tracks pricing across major LLM providers. We run an automated fetch script; we do not manually edit individual model prices.

Dataset version and last fetch time are shown in the site header on the calculator page.

Model selection (automatic)

Not every OpenRouter model is listed. An auto-selection pipeline applies:

• Auto providers — Any OpenRouter vendor with qualifying text models (e.g. Z.ai, Moonshot) is discovered on each fetch; synthetic ~vendor namespaces are blocked.
• Minimum context — Models below 32K context are excluded (agent workloads need usable context).
• Exclusions — Free tiers, distill models, image-only endpoints, legacy generations (e.g. GPT-3, Claude 3).
• Family dedup — One representative per model family (latest version; prefers standard over “fast” variants).
• Scoring cap — Up to 5 models per provider, 50 globally; flagship names score higher, niche variants lower.
• Daily refresh — GitHub Actions re-fetches OpenRouter and re-runs selection; new vendors and flagships appear without manual catalog edits.

Full rule patterns live in selection_rules.json in our repository — we keep the site summary short so it stays in sync with the pipeline.

Cost formula

For each model, estimated cost is:

cost = (input_tokens / 1M × input_price) + (output_tokens / 1M × output_price)

Where:

• input_tokens = requests × input tokens per request (monthly or daily ÷ 30)
• output_tokens = requests × output tokens per request
• Prices are USD per million tokens from the API

What we do not include

• Batch API discounts
• Cached / prompt-caching input rates
• Regional or enterprise pricing
• Tool-call overhead, vision token surcharges, or reasoning-token billing quirks
• Taxes, platform fees, or minimum spend commitments

Always verify on the provider's official pricing page before production commitments.

Capability & benchmark references

LabAgenticFlow does not publish its own benchmark scores or quality rankings. Capability chips on the calculator (Tools, Vision, context window) are derived from OpenRouter API metadata — not from our testing.

For independent quality, latency, and throughput comparisons, see:

• Artificial Analysis — LLM leaderboard
• LMSYS Chatbot Arena

Flagship models on the calculator may show a See provider benchmarks → link. URLs are configured in data/selection_rules.json (benchmark_links) and point to third-party aggregators or official provider docs — we do not scrape or display scores on-site.

Official provider pricing pages

• OpenAI
• Anthropic
• Google AI
• DeepSeek
• Mistral
• AWS Bedrock (Nova)

Update frequency

Pricing and catalog membership refresh on a daily cron (06:00 UTC via GitHub Actions). We update when providers change rates or ship new models — not on a meaningless hourly ticker. The live counts in the box above reflect the latest committed data/models.json.