By Stamos Kanellakis, Founder of ATOM
For the past 17 weeks I've been tracking per-token pricing across 51 AI inference vendors and 5,000+ SKUs. This week the index posted something we haven't seen all year: a confirmed directional move on the frontier.
The numbers are small. The shape is unusually clean.
What the data shows
The frontier index (AIPI FTR GLB, covering peak-capability flagship models like Claude Opus 4.7, GPT-5.5, Gemini 3.1 Pro) declined for the third consecutive week:
Input: -0.23%
Cached input: -2.06%
Output: -0.35%
The output figure is nearly identical to last week, which is part of what makes the trend look real.
The global text benchmark (AIPI TXT GLB, which covers the full text-generation market across all tiers) moved with it. For the first time in 2026:
Input: -0.35%
Cached input: -1.01%
Output: -0.23%
The pattern is no longer confined to flagship models. It's now visible across the wider text market.
Why three weeks matters
Single-week moves on the frontier index are common in size but usually random in direction. Vendors reprice individual SKUs without coordinating with each other, which produces noise inside a tight range.
Two weeks down starts to feel like something. A third week makes it hard to explain as noise, and easier to explain as several frontier vendors pulling in the same direction.
Week 18 adds two more reasons to take the signal seriously:
The same pattern now appears on the global text benchmark, which covers many more models than the frontier subset. Getting that index to shift requires either a large share of vendors moving together, or a handful of heavy vendors pulling the rest along.
Input, cached input, and output all softened at once on the same index in the same week. Vendors rarely move all three columns together when they are only running targeted promotions.
The forward call resolved
Two weeks ago I flagged two scheduled events that should show up in this week's run:
- DeepSeek's 75% promotional discount on V4-Pro
- Alibaba Cloud Bailian's cut to cache pricing for DeepSeek V4-Pro (RMB 1 per million tokens)
Both changes touched cached pricing rather than headline rates. So the impact should appear in the cached input column of the platform channel.
That is exactly where it landed:
AIPI PLT GLB (Platform channel)
Cached input: -17.47% ← largest single move on any AIPI series this week
Input: -0.71%
Output: -1.08%
A forward call resolving in the exact column we flagged is what separates a price tracker from a benchmark.
Coverage
Models: 3,079
SKUs: 5,022
Vendors: 51
Countries: 9
Modalities: 6 (with 35 subtypes)
The 190 SKU increase from last week came mostly from continued catalog growth in audio, voice, and image generation.
Reasoning models corroborate
AIPI RSN GLB input declined 0.50%, and the reasoning premium KPI compressed from 2.2x to 1.7x.
Part of that compression comes from new reasoning entrants joining at lower price points rather than incumbents cutting rates, so the change is not a clean price signal on its own. What matters more is that reasoning is softening at the same time as the flagship segment. Reasoning is a frontier capability, and the two indexes have historically tracked together.
What's becoming calmer vs what's moving
Volatility across the wider market continues to drop:
Input volatility YTD: 0.61% (-0.34pp from Week 17)
Cached input volatility YTD: 0.30% (-0.44pp from Week 17)
Output volatility YTD: 0.45% (-0.45pp from Week 17)
The broader market is becoming calmer at the same time the frontier subset is starting to move. That is an unusual combination and worth watching.
What to watch for Week 19
Three items are likely to shape Week 19:
Whether the pattern on AIPI FTR GLB extends to a fourth week. If it does, it becomes the longest sustained directional run on the index since we started publishing.
xAI's scheduled retirement of grok-imagine-image-pro on May 15. This falls inside the Week 19 indexing window. Two more retirements are already on the calendar: Moonshot's original Kimi K2 series on May 25, and the Writer Palmyra-x-003 family on July 13.
Audio segment. AIPI AUD GLB recorded zero movement on all three pricing directions this week after the 5.77% input jump in Week 17. The segment looks like it has settled at a new baseline of 223 SKUs. The question for Week 19 is whether audio stays calm or whether new entrants pull it back into volatility.
Methodology note
ATOM indexes use a chained matched-model methodology. Only SKUs present in both the current week and the prior week contribute to the weekly percent change, which removes the composition bias that affects simple average pricing. A maximum weekly cap of ±50% is applied at the SKU level to prevent outlier movements from distorting the aggregate.
The full Week 18 edition of The Inference Price Benchmark, including additional breakdowns by channel and modality, is published every Monday at 9am ET. Read the full edition here.
Track per-token pricing across 51 AI inference vendors at a7om.com.
Top comments (0)