TokenScale checks each provider source every night and records what the check could reproduce. A successful observation is verified; otherwise the last known value is carried forward and marked. This page audits every missed or partial check. Provider pages remain authoritative for current published prices; TokenScale preserves the independent dated record of what it observed.
This recent night is already counted as missed in the totals above. Its row in the table below, with the reason, is added after the nightly record is audited and included in a reviewed release.
A fully verified night means a human-readable price was read from the provider's own pricing page that night, for every tracked tier. A carried-forward night means the check did not complete and the previous price was held. A partly verified night means some tiers were read and the rest were held: it is counted as a miss for the tiers it missed, and the count is stated below. Those days are drawn in grey on every chart, and they stay grey: a later check confirms only its own date, it cannot reach back and confirm one it missed. The gap is part of the record.
Disclosure is the commitment here, not perfection. The verification rate will never be 100%, because the checks run on real machines in the real world. The commitment is that every night the record missed appears on this page, with the reason.
The site launched on 18 May 2026; the first automated nightly point landed on 19 May. Coverage grew with the site: 16 providers from 19 May, 21 from 2 June, and all 22 from 26 June, when Sakana AI joined the board.
| Nights | Run | What happened | Disclosed |
|---|---|---|---|
| 18 Sep | 1 night, partial | Both official-source desks covered all 22 routes. The opening desk initially verified 60 of 66 Standard tiers. At the 9:10 PM watchdog, two fresh OpenRouter exact-endpoint reads reproduced one DeepInfra DeepSeek V3 0324 route at the unchanged $0.24/$0.90 pair and $0.135 cached input, while a distinct current model-list API confirmed the exact model. The independent close reproduced that route again and restored Perplexity's September 27 Sonar support sentence from its controlling first-party page. The final record is 61 verified and five carried: Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5. Both desks reconciled all 66 exact Batch and Cache records; public Cache activation coverage stayed unchanged. | This page, 18 Sep 2026 |
| 17 Sep | 1 night, opening partial | The 8 PM official-source desk covered all 22 routes and verified 61 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5 remained exact-model holds, so their prior rates stayed visibly carried. All 66 exact Batch and Cache records were reconciled. Perplexity's three rates remained unchanged, but its current page no longer published the earlier September 27 support sentence, so current lifecycle copy was withdrawn pending the independent 10 PM close. | This page, 17 Sep 2026 |
| 16 Sep | 1 night, partial | Both official-source desks covered all 22 routes and closed at 61 of 66 Standard tiers verified. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5 remained exact-model holds, so their prior rates stayed visibly carried. All 66 exact Batch and Cache records were reconciled in both passes. Perplexity's three rates remained unchanged while its independent close again stated that Sonar support continues through September 27, confirming the restored lifecycle fact. | This page, 16 Sep 2026 |
| 15 Sep | 1 night, partial | Both official-source desks covered all 22 routes. The opening desk initially verified 60 of 66 Standard tiers. At the 9:10 PM watchdog, two fresh cache-busted OpenRouter endpoint reads reproduced one DeepInfra DeepSeek V3 0324 route at the unchanged $0.24/$0.90 pair, with a distinct current model-list API confirming the exact model. The independent close reproduced that route again, leaving the final record at 61 verified and five carried: Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5. All 66 exact Batch and Cache records were reconciled in both scheduled branches. | This page, 15 Sep 2026 |
| 14 Sep | 1 night, partial | Both official-source desks covered all 22 routes and closed at 60 of 66 Standard tiers verified. Together GPT-OSS 20B and Kimi K2.6 remained absent from current serverless inference pricing, and similarly named fine-tuning rows did not establish inference rates. Fireworks GPT-OSS 20B was absent from pricing and its exact page said Serverless: Not supported; Cerebras ZAI-GLM-4.7 was absent from the complete current catalog; Kimi K2.5 was retired, and its directed K3 successor failed the same-tier guard; and OpenRouter DeepSeek V3 had split input and output across separate endpoints. Those six prior rates stayed visibly carried. All 66 exact Batch records and all 66 exact Cache records across 22 provider routes were reconciled. | This page, 17 Sep 2026 |
| 13 Sep | 1 night, partial | Both official-source desks covered all 22 routes and closed at 60 of 66 Standard tiers verified. Together GPT-OSS 20B and Kimi K2.6 had no exact current serverless inference rows; Fireworks GPT-OSS 20B had no rate and its exact page said Serverless: Not supported; Cerebras ZAI-GLM-4.7 was absent from the complete catalog and its exact endpoint returned 404; retired Kimi K2.5 had no unique provider-directed successor; and OpenRouter DeepSeek V3 split the lowest input and output across different endpoints. Those six prior rates stayed visibly carried rather than mixing routes or changing the tracked lineup. All 66 exact Batch records and all 66 exact Cache records across 22 provider routes were reconciled. | This page, 17 Sep 2026 |
| 12 Sep | 1 night, partial | Both official-source desks covered all 22 routes and verified 60 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and OpenRouter's split DeepSeek V3 route remained exact-model holds, so their prior rates stayed visibly carried. All 66 exact Batch and Cache records were reconciled. Current Cohere model pages and the catalog reproduced all three rates and Live status after the landing-page table disappeared. | This page, 16 Sep 2026 |
| 11 Sep | 1 night, partial | Both official-source desks opened all 22 routes and verified 60 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and OpenRouter's split DeepSeek V3 route remained exact-model holds, so their prior rates stayed visibly carried. The opening desk also recorded that OpenRouter's former single V3 route and cached-input field had disappeared; the independent close reproduced the split routes. All 66 exact Batch and Cache records were reconciled. | This page, 15 Sep 2026 |
| 10 Sep | 1 night, partial | Both official-source desks opened all 22 routes and the closing watchdog recovered MiniMax M2 from its current API overview and explicit-cache guide, leaving 61 of 66 Standard tiers verified. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7 and retired Kimi K2.5 remained exact-model holds, so their prior rates stayed visibly carried. DeepSeek Lite moved prospectively to the provider-directed V4.1 Flash successor after the complete guarded-successor review passed. All 66 exact Batch and Cache entries were reconciled. | This page, 14 Sep 2026 |
| 9 Sep | 1 night, partial | The 8 PM official-source desk opened all 22 routes and verified 60 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and legacy MiniMax M2 remained exact-model holds, so their prior rates stayed visibly carried. All 66 exact Batch and Cache entries were reconciled. Perplexity's three rates verified while its earlier September 27 support sentence was absent and withdrawn from current copy. | This page, 13 Sep 2026 |
| 8 Sep | 1 night, partial | The two official-source desks opened all 22 routes, and the closing watchdog resolved all three rendered AWS Bedrock holds without changing their rates. The final record verifies 60 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and legacy MiniMax M2 remained exact-model holds, so their prior rates stayed visibly carried. All 66 exact Batch and Cache entries were reconciled. | This page, 12 Sep 2026 |
| 7 Sep | 1 night, partial | The recovered closing desk freshly opened all 22 official routes and verified 57 of 66 Standard tiers. Together Small and Flagship, AWS Bedrock's three rendered rows, Fireworks Small, Cerebras Flagship, retired Kimi K2.5 and legacy MiniMax M2 remained exact-model holds. All nine prior rates stayed visibly carried, and all 66 exact Batch and Cache entries were reconciled without inventing replacements. | This page, 11 Sep 2026 |
| 6 Sep | 1 night, partial | Both official-source desks freshly opened all 22 provider routes and verified 60 of 66 Standard tiers. Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and legacy MiniMax M2 remained exact-model holds, so their prior rates stayed visibly carried rather than being silently replaced. The closing desk also reconciled all 66 exact Batch and Cache entries. | This page, 10 Sep 2026 |
| 5 Sep | 1 night, partial | The 8 PM desk freshly opened all 22 official routes and initially verified 59 of 66 standard tiers. At 9:10, two fresh cache-busted Cohere pricing reads reproduced Command R+ (08-2024) at the unchanged $2.50/$10 rate, while Cohere's separate current model table still marked the exact SKU Live. The same-day point was upgraded, leaving 60 verified and 6 carried exact-model holds: Together GPT-OSS 20B and Kimi K2.6, Fireworks GPT-OSS 20B, Cerebras ZAI-GLM-4.7, retired Kimi K2.5 and legacy MiniMax M2. All 66 Batch entries and all 22 Cache provider routes were compared. | This page, 5 Sep 2026 |
| 4 Sep | 1 night, missed in shipping history | The separate September 4 desks and durable ledger preserve an original closing result of 60 verified and 6 carried tiers. A later local feature chain was mistakenly rebased on the live September 3 build and dropped all 66 September 4 points from the shipping data. The September 5 recovery restored the missing public date conservatively as 66 carried points instead of overstating what the replacement candidate itself could prove. | This page, 5 Sep 2026 |
| 3 Sep | 1 night, partial | The 8 PM desk freshly opened all 22 official routes and verified 61 of 66 standard tiers. Together GPT-OSS 20B and Kimi K2.6, Cerebras ZAI-GLM-4.7 and Fireworks GPT-OSS 20B remained absent or unsupported, while Moonshot's Kimi K2.5 remained explicitly retired with several possible successors. Those five rates were carried rather than silently changing the lineup. Anthropic's explicit generally available Fable 5.1 successor passed the guarded model-change gate at the same standard and Batch rates, with its exact Cache terms recorded. All 66 Batch entries and the 22-route Cache watch were also compared. | This page, 3 Sep 2026 |
| 2 Sep | 1 night, partial | Both desks freshly opened all 22 official routes. The 10 PM close verified 61 of 66 standard tiers. Together GPT-OSS 20B was reproducible at 8 PM but disappeared from both current first-party Together surfaces at close; Together Kimi K2.6, Cerebras ZAI-GLM-4.7 and Fireworks GPT-OSS 20B remained absent, while Moonshot's Kimi K2.5 remained explicitly retired with several possible successors. The previous same-model rates were carried rather than silently changing the lineup. All 66 Batch entries and the 22-route Cache watch were also compared in both branches. | This page, 2 Sep 2026 |
| 1 Sep | 1 night, missed | The scheduled 8 PM and 10 PM desks and both watchdogs were blocked while candidate v1053 awaited review and live remained v1052. No official-source route was opened for the September 1 record. Approval arrived after midnight, when a new read could no longer prove the prior date, so all 66 tiers were explicitly carried from 31 August and remain grey. No closing-desk or secondary-ledger evidence is claimed for the missed night. | This page, 2 Sep 2026 |
| 25–31 Aug | 7 nights, ended 1 Sep | Together’s tracked Kimi K2.6 and Cerebras’s tracked ZAI-GLM-4.7 remained absent from their current public catalogs, so the 25 and 26 August records verified 64 of 66 tiers and carried those two flagship rates grey. On 27 and 28 August, OpenRouter’s DeepSeek V3 route also split its lowest input and output across different endpoints, so those records verified 63 of 66 rather than inventing a mixed route. One exact DeepInfra route again owned the full $0.24/$0.90 pair on 29 and 30 August, returning OpenRouter Mid to verified. The 30 August first desk verified 63 of 66: Together and Cerebras remained grey, and Fireworks GPT-OSS 20B joined them after its rendered model page said Serverless: Not supported while the static serverless pricing table exposed no current row. The prior $0.07/$0.30 rate stayed grey and no successor was chosen silently. On 31 August the record verified 62 of 66; Together, Cerebras and Fireworks remained carried, and Moonshot’s retired Kimi K2.5 route was also carried rather than silently replaced. | This page, 29 Aug 2026 · extended 2 Sep 2026 |
| 18–24 Aug | 7 nights, ended 25 Aug | Cerebras’s tracked ZAI-GLM-4.7 remained absent from its complete public model list, so the flagship rate held grey at $2.25/$2.75 on all seven nights. The 18 to 20 August records each verified 65 of 66 tiers. On 21 August, Together’s rendered inference table also stopped listing tracked Kimi K2.6, so its $1.20/$4.50 flagship rate was carried and the record closed at 64 of 66. Together's row returned at the 22 August 8 PM read but disappeared again at the independent 10 PM close. The recovered 23 and 24 August full passes again found the row absent, so both nights verify 64 of 66 and carry the same two flagship rates grey. No replacement was chosen silently. | This page, 19 Aug 2026 · extended 24 Aug 2026 |
| 16–17 Aug | 2 nights, ended 18 Aug | DeepSeek replaced flat V4 pricing with separate peak and off-peak rates on 16 August. All three DeepSeek tiers were held rather than have the record pick a time window on the reader’s behalf, so that night closed at 63 of 66 verified. On 17 August three more first-party routes stopped reading at once: Groq’s small and mid rows during its production-model change, MiniMax’s rendered page hiding the tracked legacy row, and Cerebras’s public model list no longer containing the tracked ZAI-GLM-4.7. That night recorded 59 of 66, the lowest since 10 August. The run ended on 18 August, when the peak rate was adopted as the conservative canonical basis and the Groq lineup was settled, and both providers verified again. | This page, 19 Aug 2026 |
| 14–15 Aug | 2 nights, ended 16 Aug | The 8 PM official-source desk initially verified 64 of 66 tiers. At the independent 10 PM close, Kimi K2.5's direct official USD model page reproduced $0.60/$3, upgrading that same-day point without adding a duplicate. The final record is 65 verified and one carried: OpenRouter's previous DeepSeek V3 route disappeared, while current routes split the lowest input and output prices between two providers. That same route was still unreadable on 15 August, so the tier carried a second night and the board again recorded 65 of 66 verified. The 16 August close ended the run: fresh endpoint reads found one reproducible cheapest route for each pinned tier and that night verified at the unchanged $0.24/$0.90. | This page, 14 Aug 2026 · 15 Aug added here, 19 Aug 2026 |
| 10 Aug | 1 night | The official-console pass verified 18 of 66 tiers across Cohere, SambaNova, Cerebras, OpenRouter, xAI and Meta. The other 48 points were explicitly carried because their official routes were not completed that day. No carried point was presented as freshly checked. | This page, 11 Aug 2026 |
| 6–9 Aug | 4 nights, ended 11 Aug | A second run with no full verification pass, days after the first. Every tier was carried forward on 6 and 7 August. The 8 August run reached only 3 of the 66 tracked tiers before it stopped. Daytime passes on 9 August eventually read 48 of 66 clean, with zero price changes, while 18 tiers remained behind missing or rendered-only routes or model-choice decisions. The run ended on 11 August when all 66 tiers received a same-day official-source check. | This page, 9 Aug 2026 · closed 11 Aug 2026 |
| 1–4 Aug | 4 nights, ended 5 Aug | The nightly kept appending data but no verification pass ran and no provider was stamped as checked. Found on 4 August while auditing the site against its own record. The run went one night longer than this page first disclosed: 4 August was carried forward in full as well. It ended on 5 August, when a pass verified 39 of the 66 tracked tiers and held the other 27. | This page + journal, 4 Aug 2026 · extended and closed here, 9 Aug 2026 |
| 13–17 Jul | 5 nights | A macOS update reset a permission the scheduled job needed, and it failed quietly five nights in a row. Not detected from inside the system; it took three nights to notice. The job now raises an alarm the moment it fails. | Changelog, 18 Jul 2026 |
| 13 Jun | 1 night | The nightly carried every tier forward unchanged without reading the sources. A full re-verification the next day confirmed most values as correct and fixed six tier mappings it found along the way. | This page, 4 Aug 2026 |
| 31 May | 1 night | Both the primary and the backup nightly jobs skipped the run entirely. The miss was found the next day and the previous prices were held. | This page, 4 Aug 2026 |
| 27–30 May | 4 nights | The nightly job wrote prices it could not actually confirm at source. A data audit on 31 May caught it and reverted all 36 affected points to the last verified values. The public site never showed the bad numbers, but those nights were not verified. | This page, 4 Aug 2026 |
Partly verified nights. 39 nights read part of the board and held the rest: 5 August (39 of 66 tiers verified), 9 August (45 of 66), 10 August (18 of 66), 14 August (65 of 66), 15 August (65 of 66), 16 August (63 of 66), 17 August (59 of 66), 18 August (65 of 66), 19 August (65 of 66), 20 August (65 of 66), 21 August (64 of 66), 22 August (64 of 66), 23 August (64 of 66), 24 August (64 of 66), 25 August (64 of 66), 26 August (64 of 66), 27 August (63 of 66), 28 August (63 of 66), 29 August (64 of 66), 30 August (63 of 66), 31 August (62 of 66), 2 September (61 of 66), 3 September (61 of 66), 4 September (60 of 66), 5 September (60 of 66), 6 September (60 of 66), 7 September (57 of 66), 8 September (60 of 66), 9 September (60 of 66), 10 September (61 of 66), 11 September (60 of 66), 12 September (60 of 66), 13 September (60 of 66), 14 September (60 of 66), 15 September (61 of 66), 16 September (61 of 66), 17 September (61 of 66), 18 September (61 of 66) and 19 September (61 of 66). Each is covered by its row in the table above and counted as a miss for every tier it did not reach. Those tiers stay grey on their charts. These counts are the points the record actually holds for each night; a row above may describe a different number of tiers as read, because reading a page and writing down a verified price are not the same act.
Correction, 9 August 2026. This page understated its own record. From 4 to 9 August it showed 14 carried-forward nights while the price record held 17, left the August run marked ongoing after it had ended, and did not disclose the nights of 4, 6 and 7 August at all. On the page whose whole job is disclosure, that is the worst place to be wrong, so it is stated here rather than quietly fixed. Two things changed with it: the counts above are now computed from the price record itself on every build, and the build now fails if any night in the record is not covered by a row in this table. Logged in the changelog.
How this table was built, honestly. Before 4 August 2026 the record did not store each night's verified-or-carried status; it can now, permanently, in the data itself. The rows above are reconstructed from build logs and the change log. If a missed night ever surfaces that this table does not show, it gets added, dated, with the reason, and the correction is noted in the changelog.
Verification is per provider page, not all-or-nothing. On most verified nights all 22 pages read cleanly; on some, a page that will not render without scripts gets held while the rest verify. A rendered-source pass on 11 August cleared the current holds. Cohere is pinned to the live dated Command R7B 12-2024, Command R 08-2024 and Command R+ 08-2024 rows. AWS Bedrock is pinned to Claude Haiku 4.5, Claude Sonnet 5 and active Llama 4 Maverick 17B Instruct, with its exact US East (Ohio) rate. Kimi K2.5 was also matched on its rendered official model page. Earlier carried dates remain grey permanently; the successful 11 August check does not rewrite those gaps.
List price is not effective price. Prices are recorded as published, in dollars per million tokens. On 30 June 2026, Claude Sonnet 5 shipped with a new tokenizer that turns the same text into roughly 30% more tokens at an unchanged per-token rate: a real cost increase that never appeared on a pricing page. TokenScale records the published rate; it does not yet measure cost per unit of actual work.
Token estimates are an average. Roughly 1.33 tokens per word, which holds for typical English prose and will not hold for your code, your JSON, or your Mandarin.
The nightly record starts on 19 May 2026. Anything earlier is a baseline drawn from the sources available at the time, and is marked as such on the charts.
Promotional rates never enter the history. They are shown live, clearly labelled, and excluded from the record, because a limited-time rate is not what the thing cost.
You don't have to take my word for any of it.
Starting 5 August 2026, TokenScale's key pages are submitted to the Internet Archive every night by a watchdog that runs outside the machine the nightly runs on. That builds a third-party, timestamped copy of what this site said on any given date, one I don't control and can't edit. Earlier snapshots exist too: the Archive holds captures of this site going back to its launch window, through 9 June 2026, which corroborate the early record.
The providers' own pricing pages are archived on a rotating schedule as well. Where those captures work, you can compare what TokenScale recorded against an independent snapshot of the source. Where a provider's page is JavaScript-rendered and archives as an empty shell, that provider will be marked here as not independently checkable, because pretending otherwise would defeat the point. A provider-by-provider table appears here once the watchdog has a full week of archive results to report.
Tell me and I'll correct it in public, in the changelog, within a day. Use the "Report a pricing error" link in the calculator's footer, or email projectsbilton@gmail.com. That is the whole arrangement.
Last updated: 19 September 2026. The counts above are computed from the price record on every build and run through the night of 19 Sep.