AI Token Tracker — Sep 1 to Oct 19

Both TPI moves so far are composition effects, each verified by arithmetic — neither is a list‑price cut. Sep 11→18 (−3.7%): a 24th member, DeepSeek V4‑Pro, joined below the index average ($1.056 blended) — $2.406 becomes exactly $2.324. Sep 18→25 (−8.2%): the count stayed at 24 but four models were replaced by cheaper successors (GPT‑6 Sol $4.40 for GPT‑5.6 Sol $12.50, Opus 5.5 $8.80 for Opus 5 $11.00, GPT‑6 Luna $0.22 for $0.50, Grok 4.7 flat at $3.20); the new list's geometric mean is $2.1307 = the published $2.13. No member repriced in either case. There is a real signal underneath: each new generation is launching at a lower price than the one it replaces — the "falling launch prices across generations" pattern the thesis lists as structural deflation — but it shows up as successors, not as cuts to existing SKUs.

BenchLM stopped being a single‑constituent story and became a wholesale restatement. The small Sep 22/25 revisions (one model added, one repriced) were normal. Oct 1 is different in kind: the universe jumped 41→71 models and 12 of 43 historical months changed value, including months over a year old (Aug 2025 frontier 11.7→29.2). No changelog, no version bump visible in the API. Every BenchLM number on this page from before Oct 1 should be treated as measured on a different index than every number from Oct 1 on — they are not the same ruler.

SIT's Sep 28 drop is not reverting — four days in, it looks like the new normal. After the Sep 22 membership step (18→17) and a three‑day carry‑forward, Sep 28 dropped 11.5% to $1.345 with the model count still at 17, and Sep 29–30 held close by ($1.352, $1.358) rather than bouncing back the way the Sep 14 and Sep 20 spikes did. Same count, different result, and the API still doesn't expose which constituents changed. Down 14.3% since Sep 9 overall. Practical rule stands: never read a same‑day SIT print as final.

AI Pricing Guru began publishing six corrupted records on Sep 24 (Mythos 5, Mythos 5.1, Opus 4, Opus 4.1, Sonnet 4, Haiku 3.5 — price fields shifted, output prices collapsed to the cache‑read price). Left as published in the raw data, but any average built on the raw file falls ~8% on Sep 24 for no economic reason: the like‑for‑like average over unaffected models is unchanged at $3.9779. Check for cached‑input > input in future pulls; 06_build_dashboard_series.py now counts these automatically.

Marketplace prices keep moving while direct list prices sit still. OpenRouter: 122 prompt‑price cuts vs 69 increases since Sep 17 — cuts lead overall, but Sep 29 was a one‑day exception (13 increases vs 9 cuts) before cuts resumed leading Sep 30–Oct 1. Z.ai's GLM‑5.3 whipsawed −73% (Sep 22) then +294% (Sep 29), most of the way back. TPI's event log still records no list‑price cut or increase since Aug 17 (now 45 days), and AI Pricing Guru's unaffected models haven't moved. Caveat unchanged: OpenRouter's price is what the marketplace displays, which can reflect routing changes, not confirmed changes by the model's maker.

Gemini's exit from Cloudflare's consumer top 5 has held eight days (Sep 24–Oct 1). DeepSeek has held #4 every day since; #5 has settled on Doubao for three straight days after rotating with Grok earlier, and the two ranking endpoints that disagreed on Sep 29 now agree. The top two (ChatGPT, Claude) have never moved. This looks like a genuine reshuffle, not noise.

SDLLMTK has now round‑tripped twice. $0.955 (Sep 17) → peak $1.009 (Sep 18) → trough $0.962 (Sep 24) → $1.005 (Sep 28), back near the September peak. The Inference Margin Ratio has spent three weeks oscillating in a 0.357–0.384 band and sits at 0.369 today, essentially level with Sep 1's 0.368. Both the original "improving economics" read and last update's "it's flat‑to‑down" correction were each true of a moment, not a trend — the honest description is a band, not a direction.

B200 is the one chip with a clean, sustained move; everything else is oscillating. Ornn B200 is up 18.3% since Sep 10 and has risen in 9 of the last 10 sessions, now 38% above Silicon Data's rate card. H100/H200/A100 all dipped around Sep 27 and recovered by Sep 29 — net moves near zero. Compute Desk's B300, flagged two updates ago as possibly transitioning away, has fully round‑tripped ($5.94→$5.36→$6.01) and is back within 1% of its Sep 1 level — that read is retracted. Kalshi's one open B200 expiry (Oct 2) still implies a median near $8.11 against Ornn's latest $8.14.

HuggingFace's download count broke its losing streak. After 13 straight days of decline (256.5M → 242.7M, −5.4%), the top model's trailing‑30‑day downloads rose on both Sep 30 and Oct 1 — tiny moves (+0.03%, +0.04%) but the first up days since this series started. Likes never stopped climbing throughout (6,014 → 6,151). Too early to call a trend reversal off two small up-ticks; noted for the record so the next down day isn't mistaken for a continuation of an unbroken streak.

Kalshi's settled‑markets count has never behaved like a rolling window. Ten daily pulls (Sep 22–Oct 1): 152, 152, 139, 139, 152, 152, 152, 152, 152, 141 — the max settlement date hasn't moved past Sep 25 in two weeks, yet the count jumps around by double digits. A genuine rolling window would show the date advancing with the count roughly stable, or vice versa; this shows neither. Treat every Kalshi settled‑count figure in this tracker as that day's API response, not a ground truth.