SpanveroModelsGPUsCalculatorCompareTrendsPricing

State of Inference Costs — July 2026

Published July 9, 2026 · data June 20, 2026 → July 9, 2026 (20 daily snapshots) · live edition — extends with each daily snapshot until the next report ships

What running AI actually cost this month — computed straight from Spanvero's append-only daily price record, the same data you can download. No estimates, no smoothing: every figure carries its date.

Daily snapshots
20
Jun 20 → Jul 9
Prices logged
8,766
rows appended to the log
Models tracked
339
live market, latest day
Real price moves
44
25 cuts · 19 rises

What the data shows

Between June 20, 2026 and July 9, 2026, Spanvero logged 8,766 prices across 20 daily snapshots: the full live market listing (339 priced models on the latest day), 83 GPU rental rates, and the dated per-model prices Spanvero publishes. The log is append-only — a day, once written, is never revised.

The floor held. The cheapest paid model on the market cost $0.02 per 1M blended tokens on June 20, 2026 (inclusionAI: Ling-2.6-flash) and $0.02/1M on July 9, 2026 (still inclusionAI: Ling-2.6-flash) — no meaningful move at the bottom of the market.

Above the floor, prices did move: 44 real day-over-day changes cleared our 0.5% noise threshold — 25 cuts against 19 rises. The largest single move was Qwen: Qwen3 235B A22B Thinking 2507, rising from $0.1 to $0.8223/1M on July 1, 2026 (+722%); the largest cut was Nex AGI: Nex-N2-Pro, from $1.5 to $0.625/1M on June 23, 2026 (−58%).

On the rental side, the cheapest tracked GPU went from $0.26/hr on June 20, 2026 (NVIDIA RTX 3090 24GB) to $0.06/hr on July 9, 2026 (NVIDIA RTX 3060 12GB). Read that with care: our GPU coverage grew from 10 to 83 SKUs over the window, and the new floor comes from a card that entered the record mid-window — so much of that drop is wider tracking, not vendor repricing. Like-for-like — only the SKUs tracked for the whole window — the floor held at $0.26/hr (NVIDIA RTX 3090 24GB).

Finally, the market ended the window 3 models larger than it began (entries minus exits), and truly-free endpoints ($0 in, $0 out) held at 27.

The cheapest way to run AI, start vs end

Biggest market moves

Single-day changes on the live market listing, largest first — 44 in this window. The full feed lives on the price record.

NameBeforeAfterChangeDate
Qwen: Qwen3 235B A22B Thinking 2507$0.1/1M$0.8223/1M+722%Jul 1
Qwen: Qwen3 30B A3B Thinking 2507$0.24/1M$0.845/1M+252%Jul 2
Nex AGI: Nex-N2-Pro$1.5/1M$0.625/1M−58%Jun 23
MiniMax: MiniMax M2.5$0.525/1M$0.3/1M−43%Jun 26
Anthropic Claude Sonnet Latest$9/1M$6/1M−33%Jul 1
Z.ai: GLM 5.1$2.03/1M$2.6375/1M+30%Jun 30
Tencent: Hy3$0.5/1M$0.36/1M−28%Jul 9
Qwen: Qwen3 8B$0.225/1M$0.286/1M+27%Jul 2
MiniMax: MiniMax M2.7$0.6/1M$0.45/1M−25%Jun 26
TheDrummer: Rocinante 12B$0.3/1M$0.375/1M+25%Jun 24

Published-model movers

Window-start vs window-end change in the dated price Spanvero publishes per model — each links to that model's live cost page. 12 of 62 tracked models moved.

NameBeforeAfterChangeLast changed
Qwen3 30B A3B Thinking 2507 (entered Jun 23)$0.24/1M$0.845/1M+252%Jul 7
Qwen3 Coder 480B A35B Instruct$0.61/1M$1.01/1M+66%Jun 23
MiniMax M2.5 (entered Jun 23)$0.525/1M$0.3/1M−43%Jun 30
MiniMax M2.7 (entered Jun 23)$0.625/1M$0.45/1M−28%Jun 30
Qwen3-8B (entered Jun 23)$0.225/1M$0.286/1M+27%Jul 7
GLM 5.2 (entered Jun 23)$2.5/1M$1.8823/1M−25%Jul 7
gpt-oss-120b$0.11/1M$0.09/1M−18%Jun 30
DeepSeek V3.2 Exp$0.32/1M$0.34/1M+6.2%Jun 23
Phi-4 (entered Jun 23)$0.1025/1M$0.105/1M+2.4%Jun 30
GLM 5.1$2.03/1M$2.001/1M−1.4%Jul 7

GPU rental movers

Hourly rates that changed over the window, per tracked SKU — each links to that card's rate page.

NameBeforeAfterChangeLast changed
NVIDIA A100 40GB$1.10/hr$1.99/hr+81%Jul 3
NVIDIA H200 141GB$3.49/hr$4.39/hr+26%Jul 3
NVIDIA RTX A5000 24GB$0.36/hr$0.27/hr−25%Jul 3
NVIDIA H100 80GB SXM$2.69/hr$3.29/hr+22%Jul 3
NVIDIA L40S 48GB$0.86/hr$0.99/hr+15%Jul 3
NVIDIA L4 24GB$0.43/hr$0.39/hr−9.3%Jul 3

Sources, labeled: the market floor and single-day moves come from the live OpenRouter listing (the full market); published-model rows are the dated prices Spanvero publishes; GPU rates are the vendors' own posted prices. We sell no compute and mark nothing up. How we stay honest →

Cite this report

Free to quote with a link. The underlying record: history.json · history.csv · prices.json · gpus.json · the full feed →

Spanvero, “State of Inference Costs — July 2026” (data 2026-06-20 → 2026-07-09) — spanvero.com/reports/inference-costs-2026-07/

Want this level of analysis run on your own workload? We sell flat-fee written cost audits — $49 / $249 / $499 — priced from the same dated record. How the audits work →

All reports → · The daily price record · Cost calculator · Open the free Spanvero advisor →

Data June 20, 2026 → July 9, 2026; published July 9, 2026. $0 markup, your own accounts, we never resell compute. © 2026 Cynosure LLC.

The weekly price index

A short email of real AI price moves, straight from the daily log — no hype. We're collecting the list now; the first issue goes out when it opens. Unsubscribe with one click.