AI Cost Transparency Report Examines 10 Providers and 32 Models

Summary

The frozen Q4 2026 report evaluates 10 AI providers against 25 cost-transparency criteria each, using primary evidence from provider-published material. Azure OpenAI ranked first at 23 of 25 criteria, followed by AWS Bedrock at 22, while Cohere ranked last at 9; the report says the weakest area was billing and invoice quality, which scored 17 of 50 points. It records two documented changes that moved detail away from invoices: OpenAI removed detailed API costs from Enterprise invoices issued from April 1, 2026, and Google Cloud says detailed usage line items had already been removed from invoices in 2021. A separate workload index prices four declared usage patterns across 32 models using published cache, batch and long-context rates. The cheapest-to-most-expensive monthly spread ranged from 134x for a coding-agent workload to 275x for a batch pipeline, with the report emphasizing that a per-million-token price is not the same as an actual bill. Five field measurements found that cache accounted for 97% of input volume in one four-month coding-agent export, model choice produced a roughly sevenfold median session-cost difference in a Cursor dataset, and tokenizer changes could raise token counts by about 30% at unchanged token prices. The report also notes that caching may not apply below provider-specific prompt-length minimums. OpenAI was the only provider to give a substantive response within the publication window; no provider changed a score.