The AI Cost Transparency Report Q4 2026, frozen on 2026-10-13, maps what ten providers disclose about pricing against what 32 models actually cost across four declared workloads. The report scores each provider on 25 binary criteria drawn from published material, producing an index where higher scores indicate greater transparency. Azure OpenAI leads with 92% (23 of 25 criteria met), followed by AWS Bedrock at 88% (22/25). At the bottom, Cohere scores 36% (9/25). No provider was given advance sight of scores beyond their own right-of-reply packet.

Scoring breakdown and what the numbers reveal

Scores across five disclosure blocks show consistent patterns. Pricing clarity reaches 74% (37 of 50 cells met), billing predictability 84% (42/50), spend observability 70% (35/50). The weakest block is billing and invoice quality at 34% (17/50). Two invoice failures are documented regressions declared by providers themselves: OpenAI's help centre states that from invoices issued on 1 April 2026, detailed API costs are no longer provided on Enterprise invoices, replaced by an export flow. Google Cloud's documentation states that detailed usage cost line items were removed from invoices in 2021 and that invoice usage details do not include a column for the SKU price. Both moved detail off the invoice and into a separate surface.

Disclosure scores stand at 80% (40/50), indicating that most providers publish some pricing information, but the gap between quoted prices and actual bills remains significant. The report emphasizes that billing predictability scores higher than invoice quality: providers document what they will charge considerably better than what they actually charge.

Effective Workload Cost Index

The report prices four declared workloads against all 32 models on the board using each model's real cache, batch and long-context rates where published. Workload profiles are cost models, not benchmarks: they say what a declared pattern of usage costs, not which model is better. Four workloads are defined: an interactive app with one user and many turns with nothing cached; the same app with a stable system prompt and history served from cache; an agent run of 6 steps with re-sent context and heavy output (reasoning not priced); and an offline job of 1M requests per month with list input and short output using batch pricing.

The spread between the cheapest and dearest model on the board runs from 134x to 275x depending on workload. The batch pipeline spreads widest because the discount compounds: the cheap model applies a batch tier on an already low rate, and the dear one has no batch tier to apply. The report notes that where a provider publishes no cache-read rate, the workload is computed without cache, so its cost goes up, not down.

Five field notes from measured usage

Five measurements taken from real usage over the quarter reveal gaps between price tables and actual bills, gaps made of cache, tokenizers and thresholds that do not appear in headline pricing.

  • Measured from 18,700 requests of real opencode usage over four months: cached tokens were 97% of all input volume — 2.28 billion against 65.9 million uncached. Any cost model that ignores the cache is not off by a margin; it is measuring a different thing.
  • Per-session spend across harnesses from OpenRouter's published session-cost dataset: in the same harness and turn range, the model choice moves the median session from under $4 to over $27 — a 7x spread that no per-token table makes visible.
  • A scheduled price increase announced at launch, tracked against the provider's own page until the date passed: the increase was cancelled and the introductory rate became standard. A published claim with an expiry date has to be re-read at the source, never inherited from the repo.
  • Published minimum prompt lengths below which caching does not apply: a prompt under the minimum pays full list price however stable it is. The discount people design around does not exist below the threshold, and the threshold is not in the headline price table.
  • Token counts for identical text across model generations that changed tokenizer: a newer tokenizer produced about 30% more tokens for the same text. At an unchanged per-token price, the bill rises without a single line of the price table moving.

Right of reply

Every provider was contacted through a published channel and offered 14 days to respond before publication. OpenAI replied on the substance, criterion by criterion, with no corrections requested. Three further providers — AWS, Anthropic and Google — answered with redirects to support channels, recorded as received but not as position: a pointer to a help desk is not a provider's view of its own score. The remaining six did not respond within the window. No score on the page was changed by a provider.

What this means for teams

Teams selecting models based on per-token pricing should account for cache effects, tokenizer differences, and minimum prompt length thresholds that can change bills substantially. The 134x to 275x spread between cheapest and dearest models on identical declared work shows that published price per million tokens is a starting point, not a bill. Billing predictability outpaces invoice quality, meaning what providers promise to charge and what they actually deliver diverge. Teams building cost models should test against real usage patterns including cached input, and verify whether their provider's invoice includes the detail needed for reconciliation. The dataset is available under CC BY 4.0 for those who wish to recompute workloads against today's board.