⚡ Key takeaways
- Surprise: open-weight models fell from 53% to 22% of language-model releases in our data (2023 → 2026).
- Yet open models now arrive close behind the best closed ones (Mistral Large 4, Qwen, Kimi, Inkling 2 Small).
- Prices keep falling: cheaper tiers each quarter and lower list prices.
- Most commoditised: open-weight models, then frontier models, harnesses and frameworks.
- Scarcest: chips and independent verification and environments.
🗺️ Share versus quality
Thesis 1: Fewer open releases, but closer to the frontier.medium confidence
🔧 Reproduce this: commoditisation_index.py
commoditisation_index.py · reads data/model-releases.json; layer inputs are our assumptions
"""commoditisation_index.py — how fast each layer is turning into a commodity (our index).
Part 1 uses the site's data: the share of LLM releases in data/model-releases.json that
were open-weight, per year. Part 2 scores eight layers 0-10 from three inputs (our
assumptions, explained in the article): open alternative quality, price decline per
year, and number of serious competitors. Higher = more commoditised. Stdlib only.
"""
import json, os
m = json.load(open(os.path.join(os.path.dirname(os.path.abspath(__file__)), "..", "..", "model-releases.json")))
for y in ("2023", "2024", "2025", "2026"):
xs = [x for x in m["items"] if str(x["date"]).startswith(y) and "llm" in x["mod"]]
op = sum(1 for x in xs if x.get("open") in (1, 2))
print(f"{y}: {len(xs):3d} LLM releases, {op / len(xs):.0%} open-weight")
L = [ # layer, open alternative (0-1), price decline per year (0-1), serious competitors
("Chips", 0.1, 0.2, 4), ("Frontier models", 0.6, 0.6, 6), ("Open-weight models", 1.0, 0.7, 10), ("Harnesses", 0.6, 0.5, 8),
("Frameworks", 1.0, 0.0, 10), ("Runtimes", 0.5, 0.3, 6), ("Builders", 0.4, 0.2, 8), ("Verification & environments", 0.2, 0.1, 3)]
print(f"{'layer':28} index (0 = scarce, 10 = commodity)")
for n, o, p, c in sorted(L, key=lambda r: -(r[1] * 4 + r[2] * 4 + min(r[3], 10) / 10 * 2)):
idx = o * 4 + p * 4 + min(c, 10) / 10 * 2
print(f"{n:28} {idx:4.1f} {'#' * round(idx)}")$ python3 commoditisation_index.py
2023: 15 LLM releases, 53% open-weight
2024: 29 LLM releases, 45% open-weight
2025: 55 LLM releases, 27% open-weight
2026: 76 LLM releases, 22% open-weight
layer index (0 = scarce, 10 = commodity)
Open-weight models 8.8 #########
Frontier models 6.0 ######
Harnesses 6.0 ######
Frameworks 6.0 ######
Runtimes 4.4 ####
Builders 4.0 ####
Chips 2.0 ##
Verification & environments 1.8 ##
So what: Commoditisation is about how close the alternatives are, not how many there are.
📊 The index
Thesis 2: Models and agent tooling behave like commodities; chips and verification do not.medium confidence
Commoditisation index (0 = scarce, 10 = commodity; our model)
| Open-weight models | |
| Frontier models | |
| Harnesses | |
| Frameworks | |
| Runtimes | |
| Builders | |
| Chips | |
| Verification & environments |
Real script output. Inputs: quality of the open alternative, yearly price decline, number of serious competitors (our assumptions).
So what: Value moves to what is hard to copy: hardware, data, distribution and trust.
📄 Method and sources
- Open-weight shares computed from the Model Release Map.
- Layer index inputs and weights are our assumptions, stated in the script.
- Price trend from the site's token-price article and lab announcements.
Facts are sourced; figures marked ⚠️ come from secondary reports or company statements. Scores, weights and predictions are our judgement and are labelled. Disclosure: Claude, made by Anthropic, helps write this site. Anthropic is one of the companies analysed here, under the same rules as the others.
🧭 Foundations: the AI market, Nov 2026
Live data: AI Market Map & Investments · AI Investment Tracker.
Read
Read
Commoditisation
You are here