What is working, and what is not
What is working, what is not, and what we are late on. Published by us, before anyone else finds it.
Sources last read 2026-09-18 at 13:51 UTC. Page built 2026-09-18 at 16:03 UTC. 755 published values are overdue for a re-check, and 50 contradictions are open, and 758 values are frozen because their source stopped publishing them. Collection runs every six hours; the site is rebuilt every hour, so the two timestamps above are never the same and neither one stands in for the other.
Every source we read
| Source | Last run | Result | Values written | Last failure | Why it failed |
|---|---|---|---|---|---|
| anthropic_models | 2026-09-18 01:15 | ran clean | 20 | 2026-09-02 | values we knew went missing from the source |
| anthropic_pricing | 2026-09-18 01:22 | ran clean | 85 | 2026-09-02 | the page changed shape and our reader lost its footing |
| aws_bedrock | 2026-09-18 01:30 | ran clean | 3 880 | 2026-08-26 | the source gave two different values for the same cell |
| baseten_pricing | 2026-09-18 01:37 | ran clean | 44 | never | |
| cerebras_pricing | 2026-09-18 10:14 | ran clean | 0 | 2026-09-11 | values we knew went missing from the source |
| cloudflare_workers_ai | 2026-09-18 01:11 | ran clean | 1 | 2026-09-15 | the page changed shape and our reader lost its footing |
| deepinfra | 2026-09-18 10:11 | ran clean | 1 | never | |
| deepseek_pricing | 2026-09-18 01:39 | ran clean | 0 | 2026-09-11 | values we knew went missing from the source |
| ecb_fx | 2026-09-18 13:20 | ran clean | 0 | 2026-09-01 | the page changed shape and our reader lost its footing |
| fireworks_pricing | 2026-09-18 13:22 | ran clean | 0 | never | |
| google_gemini | 2026-09-18 01:49 | ran clean | 274 | 2026-08-27 | the page changed shape and our reader lost its footing |
| groq_pricing | 2026-09-18 01:52 | ran clean | 65 | never | |
| hf_router_perf | 2026-09-18 02:52 | ran clean | 351 | never | |
| litellm_registry | 2026-09-18 01:57 | ran clean | 118 | 2026-08-27 | values we knew went missing from the source |
| mistral_pricing | 2026-09-18 02:10 | ran clean | 46 | 2026-08-27 | the page changed shape and our reader lost its footing |
| moonshot_pricing | 2026-09-18 02:12 | ran clean | 35 | 2026-09-02 | values we knew went missing from the source |
| nebius_token_factory | 2026-09-18 01:17 | ran clean | 4 | never | |
| novita_api | 2026-09-18 13:32 | ran clean | 28 | 2026-09-12 | the run did not complete |
| novita_pricing | 2026-09-18 13:51 | ran clean | 0 | 2026-09-15 | the run did not complete |
| openai_pricing | 2026-09-18 02:26 | ran clean | 390 | never | |
| openrouter_models | 2026-09-18 10:12 | ran clean | 69 | 2026-09-13 | the run did not complete |
| openrouter_uptime | 2026-09-18 02:27 | ran with reservations | 1 010 | 2026-09-18 | the run did not complete |
| ovh_ai_endpoints | 2026-09-18 02:29 | ran clean | 2 | never | |
| perplexity_pricing | 2026-09-18 02:39 | ran clean | 0 | never | |
| replicate_pricing | 2026-09-18 01:36 | ran clean | 0 | 2026-09-05 | the page changed shape and our reader lost its footing |
| snowflake_cortex | 2026-09-18 01:28 | failed | 0 | 2026-09-18 | the page changed shape and our reader lost its footing |
| vendor_entities | 2026-09-14 03:09 | ran clean | 0 | never | |
| xai_pricing | 2026-09-18 02:46 | ran clean | 21 | 2026-08-27 | values we knew went missing from the source |
| zai_pricing | 2026-09-18 02:48 | ran clean | 3 | 2026-09-11 | values we knew went missing from the source |
« Values written » counts what a run actually recorded, not what it read. A run that finds every figure unchanged writes nothing, and that is the normal case: we record a value when it moves, never once per pass.
What we are holding back, and why
| Count | What it is |
|---|---|
| 17 465 | Parameters seen only by a third-party aggregator. Level 5 is never published on its own, so these stay out until a vendor page confirms them. |
| 159 | Parameters our own code stopped producing after we changed how we read a page, and which a current parameter already covers with the same value. Kept in the archive, removed from the pages. |
| 755 | Published values not re-checked within three times the cadence we announced for them. They stay visible and carry their real observation date, which is the point. |
| 50 | Open contradictions: two sources disagree, or a movement looked like a redesign rather than a repricing. Each one is recorded rather than silently resolved in our favour. |
| 758 | Values frozen because the source stopped publishing them. We keep them with the date they were last observed, which is what a registry is for. Nothing can bring this number down, and nothing should. |
What this page deliberately does not show
Three figures would look good here and would not be honest yet. The median delay between a vendor changing a price and us detecting it, because we do not yet record the vendor's own effective date and would be measuring our own collection interval. A breakdown of held-back parameters by reason, because there is only one reason so far. And an uptime percentage, because this site has been collecting since 25 August 2026 and a percentage over three days says nothing.