REPORTS · EDITION #1 · FROZEN 2026-10-13
AI COST TRANSPARENCY REPORT
What 10 providers disclose about what they charge, what 32 models actually cost across four declared workloads, and five notes from measured usage. Every number on this page is a snapshot of 2026-10-13 and does not move: later editions are new pages, not edits to this one.
Scored providers cannot sponsor this report, this site, or any edition of it. No provider was given advance sight of the scores beyond their own right-of-reply packet. Rubric 1.0, frozen 2026-08-27 — the RFC was published before any provider was scored. Full dataset (JSON, CC BY 4.0).
1 — Provider Cost Transparency Index
Ten providers, 25 binary criteria each, 250 cells. A criterion is MET only on primary evidence from the provider's own published material; where our reviewer and verifier read the same page and disagreed, the cell resolves to the stricter reading and says so. Every non-MET cell was re-checked against the live page on 2026-09-13 before publication.
| Provider | Met | Score | Right of reply |
|---|---|---|---|
| Azure OpenAI | 23/25 | 92% | no response |
| AWS Bedrock | 22/25 | 88% | no response |
| 20/25 | 80% | no response | |
| xAI | 20/25 | 80% | no response |
| OpenAI | 18/25 | 72% | replied |
| Anthropic | 17/25 | 68% | no response |
| DeepSeek | 15/25 | 60% | no response |
| OpenRouter | 15/25 | 60% | no response |
| Mistral AI | 12/25 | 48% | no response |
| Cohere | 9/25 | 36% | no response |
What the scores say
The weakest block is billing / invoice quality: 17 of 50. Providers document what they are going to charge considerably better than what they charged. Billing predictability scores /; the invoice itself scores 17/50. The gap is the finding: the document you reconcile against is the least transparent artefact these companies produce.
Two of those invoice failures are not inherited conditions but documented regressions, declared by the providers themselves. OpenAI's help centre states that for Enterprise customers, from invoices issued on 1 April 2026, detailed API costs are no longer provided on invoices, replaced by an export flow. Google Cloud's documentation states that the 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. We record these as changes rather than absences, because that is what they are.
| Pricing clarity | 37/50 | 74% |
| Billing predictability | 42/50 | 84% |
| Spend observability | 35/50 | 70% |
| Billing / invoice quality | 17/50 | 34% |
| Disclosure | 40/50 | 80% |
2 — Effective Workload Cost Index
A price per million tokens is not a bill. This index takes four declared workloads and prices them against all 32 models on the board, using each model's real cache, batch and long-context rates where the provider publishes them. A workload profile is a cost model, not a benchmark: it says what a declared pattern of usage costs, not which model is better. Every assumption is declared. Where a provider publishes no cache-read rate the workload is computed WITHOUT cache, so its cost goes up, not down.
Interactive app: one user, many turns, nothing cached.
Same chat app with a stable system prompt + history served from cache.
Agent run: 6 steps, re-sent context, heavy output. Reasoning not priced.
Offline job: 1M requests/mo, list input, short output, batch pricing.
The spread is the point. Identical declared work, identical month, and the difference between the cheapest and the dearest model on the board runs from 134x to 275x. 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 live workload tool recomputes all of this against today's board, which will have drifted from this page.
3 — Field notes
Five measurements taken from real usage over the quarter, each published in full at the time. They are here because the quarter's pattern is visible only across them: the gap between a price table and a bill is made of cache, tokenizers and thresholds, none of which appear in the price table.
- MEASURED
- 18,700 requests of real opencode usage over four months, from the user's own export.
- FINDING
- 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.
- MEASURED
- Per-session spend across harnesses, from OpenRouter's published session-cost dataset (CC BY 4.0).
- FINDING
- In the same harness and the same 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.
- MEASURED
- A scheduled price increase announced at launch, tracked against the provider's own page until the date passed.
- FINDING
- The increase was cancelled and the introductory rate became standard. Our own note claimed the hike for two days after it was called off: a published claim with an expiry date has to be re-read at the source, never inherited from our own repo.
- MEASURED
- Published minimum prompt lengths below which caching does not apply, per provider.
- FINDING
- 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.
- MEASURED
- Token counts for identical text across model generations that changed tokenizer.
- FINDING
- 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.
4 — Right of reply
Every provider was contacted through a published channel and offered 14 days to respond before publication. 1 of 10 replied on the substance: OpenAI, criterion by criterion, with no corrections requested.
Three further providers — AWS, Anthropic and Google — answered with redirects to support channels. We record those as received but not as a position: a pointer to a help desk is not a provider's view of its own score, and presenting it as one would overstate what we were told. The remaining six did not respond within the window. No score on this page was changed by a provider.
| Provider | Contacted | Channel | Response |
|---|---|---|---|
| Azure OpenAI | 2026-08-28 | email [email protected] | No response within the window. |
| AWS Bedrock | 2026-08-28 | AWS support case 178791375200321 (Billing / Other Billing Questions, opened 2026-08-28) | Billing support agent, not a company position, replied 2026-09-15 (after the window had closed): said the enquiry is outside the scope of billing and account support, that it cannot be forwarded or escalated through the support case channel, and pointed us to press.aboutamazon.com/contact-us. Disputed no criterion. |
| 2026-08-31 | [email protected] from a personal Gmail address on 2026-08-31 (three deliveries from our own domain on 2026-08-28 bounced; Google Cloud billing support case 74848114 opened the same day was closed as out of scope) | [email protected] returned only its automatic acknowledgement. Cloud Billing Support had closed case 74848114 as out of scope: no channel to forward external assessments to press or documentation teams, pointing us back to the press address that had bounced. | |
| xAI | 2026-08-28 | email [email protected] | No response within the window. |
| OpenAI | 2026-08-28 | help.openai.com support chat, then email [email protected] (routed there by their own support) | Case 13935228, reply 2026-08-29 from [email protected]. Reviewed all seven disputed criteria (A3, C5, D1, D2, D4, D5, E2) and submitted no correction: no additional public evidence to change any classification. Explicitly endorsed the NOT VERIFIABLE reading on D4 and E2. Confirmed that from invoices issued 2026-04-01, detailed API costs are no longer on Enterprise invoices. Scope limited to published documentation. |
| Anthropic | 2026-08-28 | email [email protected], then email [email protected] (their support pointed us to the contact channels on their website) | Support agent, not a human position: said the enquiry is out of scope for customer support and pointed us to the contact channels on their website. Disputed no criterion. |
| DeepSeek | 2026-08-28 | email [email protected] | No response within the window. |
| OpenRouter | 2026-08-28 | email [email protected] | Support agent, not a company position, replied 2026-09-21 (after the window had closed): said this inbox is the customer support queue and not the right place for media inquiries, and offered no alternative channel. Disputed no criterion. |
| Mistral AI | 2026-08-28 | email [email protected] | No response within the window. |
| Cohere | 2026-08-28 | email [email protected] | No response within the window. |
The dataset
Everything above is in /reports/2026-q4.json: scores, per-block aggregates, the full workload index, right-of-reply state and the dates. No key, no sign-up, no rate limit. CC BY 4.0 — free to use with attribution: link back to https://tokenburnindex.com/reports/2026-q4
Methodology is at /methodology, the frozen rubric at /transparency-rfc, and the live index — which does move — at /transparency-index. Board prices in this edition are as of 2026-09-30; deep pricing generated 2026-10-08.