Methodology

How we collect, normalise, and maintain AI model pricing data.

Data sources

All prices are sourced directly from official provider documentation and pricing pages. We do not use third-party aggregators as a primary source. Primary sources for each provider:

Provider	Source
OpenAI	openai.com/api/pricing
Anthropic	anthropic.com/pricing
Google	cloud.google.com/vertex-ai pricing
AWS Bedrock	aws.amazon.com/bedrock/pricing

Normalisation rules

To allow fair comparison across providers, all prices are standardised as follows:

• All rates expressed as USD per 1 million tokens.
• Input and output prices tracked separately — they differ significantly between models.
• Cached input rates tracked where officially documented by the provider.
• Batch pricing discounts documented but not used as the primary comparison figure — real-time rates are shown by default.
• Context window sizes quoted in tokens, not words or characters.

Model status labels

Each model in the calculator carries an implicit status:

• Current — actively recommended by the provider, no deprecation notice.
• Legacy — superseded by a newer version but still accessible via API.
• Deprecated — provider has announced a retirement date. We remove deprecated models from the default calculator dropdown on or after their retirement date.

Update process

Prices are reviewed manually against official documentation. We aim to reflect provider price changes within 7 days of announcement. Major changes (new model releases, significant price drops) are typically reflected within 48 hours.

The calculator shows a "Prices last updated" timestamp in the pricing overrides section. The models comparison page subtitle indicates the month of the most recent full review.

Calculator methodology

The cost calculator models the full operational cost of an AI system, not just token costs. Each component is calculated as follows:

• Core LLM inference: (input tokens × input rate + output tokens × output rate) × monthly calls, adjusted for cache hit rate and batch share.
• Web search: search calls per workflow × monthly workflows × price per 1,000 calls.
• Embeddings: monthly embedding tokens × embedding price per million.
• OCR: pages per workflow × monthly workflows × price per page.
• Fixed infrastructure: user-supplied monthly fixed cost for hosting, monitoring, vector DB.
• Contingency buffer: percentage applied to all variable costs to account for usage spikes and estimation error.

Scenario ranges (low / expected / high) apply multipliers to usage volume, token counts, and tool call frequency. They do not alter pricing rates.

Limitations

• Only publicly listed API prices are tracked. Enterprise, committed-use, or negotiated discounts are not reflected.
• Prices can change without notice. Always verify at the official provider pricing page before committing to a budget.
• Token counts are estimates. Actual token usage depends on the specific text, model tokeniser, and implementation. Test with real data before scaling.
• The calculator uses a simplified cost model. Real production systems may have additional cost vectors not modelled here (e.g. image generation, audio transcription, fine-tuning).
• We are not affiliated with OpenAI, Anthropic, or any AI provider. We have no advance notice of pricing changes.

Editorial standards

ModelPricing is an independent resource. Our editorial decisions — which models to include, how to describe them, which comparisons to highlight — are not influenced by providers or affiliate relationships. Affiliate partnerships (disclosed separately) cover tool recommendations only, not model pricing data.

If you spot a pricing error or have a correction, the best way to flag it is to verify against the provider's official page and note the discrepancy. We take accuracy seriously.

Data use

Pricing data displayed on this site is sourced from public documentation and is freely available. We ask that you credit ModelPricing if you republish data from this site. Commercial redistribution of bulk pricing data in database or API form is not permitted without prior agreement.