Story · Changelog · No 15

The short version, in eight cards

parallaxlens.com/s/2026-09-21-open-models-four-months-behind
Parallax · Story 1 / 8
Parallax medallion, tech desk
Changelog · No 15

Open models are four months behind, not years

Kimi K3 took 84 days to reach the frontier OpenAI had in April. You can download it free. Almost nobody has.

21 Sep 2026 · 5 min read · 13 sources
Swipe up Swipe
Changelog · No 15 2 / 8
WHAT YOU ASSUME

Two measurements, one answer

What most people think

Open models are hobby projects. The paid ones are a generation ahead, and closing that gap will take years.

What the data shows

4months

Two measurements, one answer

Source · Epoch AI, open-closed capability gap · May 2026

Changelog · No 15 3 / 8
FIVE DATES

Three months behind, then four

2025
Epoch AI reads the gap
2026
GPT-5.5 · ECI 158.22
Epoch AI reads it again
Kimi K3 · ECI 158
GPT-6 Astra · ECI 166

Two readings of the same gap, seven months apart. In between, the day an open model caught April's best closed one.

Source · Epoch AI data insights and model pages · September 2026

Changelog · No 15 4 / 8
INSIDE THE AVERAGE

Eight points, spread very unevenly

points
58 Mystery Game Puzzles

Eight points ahead on the index, four months ahead in time. Those are two readings of the same gap: the points say how far behind, the months say how long the catch-up took.

Source · Epoch AI model pages · September 2026

Changelog · No 15 5 / 8
WHO IS ACTUALLY THERE

Not the labs you have heard of

8 of 10 open
DeepSeek · V3.2DeepSeek · V4-Pro, MIT licenceDeepSeek · V4-Flash, MIT licenceMoonshot · Kimi K2 ThinkingMoonshot · Kimi K2.6Moonshot · Kimi K3, 2.8TZhipu · GLM-5Zhipu · GLM-5.1OpenAI · GPT-5.5OpenAI · GPT-6 Astra
  1. DeepSeek · V3.2 open
  2. DeepSeek · V4-Pro, MIT licence open
  3. DeepSeek · V4-Flash, MIT licence open
  4. Moonshot · Kimi K2 Thinking open
  5. Moonshot · Kimi K2.6 open
  6. Moonshot · Kimi K3, 2.8T open
  7. Zhipu · GLM-5 open
  8. Zhipu · GLM-5.1 open
  9. OpenAI · GPT-5.5 closed
  10. OpenAI · GPT-6 Astra closed

Those gaps run from three points to fifty-eight. So who releases these open models? Three Chinese labs: DeepSeek, Moonshot, which makes Kimi, and Zhipu, which makes GLM.

Source · Epoch AI Benchmarking Hub model pages · September 2026

Changelog · No 15 6 / 8
THE WHOLE STACK

Free is only the top layer

  1. Weights Downloadable today. Kimi K3, DeepSeek V4, Qwen. The layer everyone means.
  2. Licence Permissive for 81% of Chinese releases above 20B. Only 29% of American ones.
  3. Code Inference code usually. Training code rarely.
  4. Training data Almost never released. You cannot see what the thing was trained on.
  5. Compute and serving Not downloadable at any licence. The cluster is the real bill.

Eight of those ten releases are free to download. Nothing under them is, so the rest of the stack stays somebody else's bill, and that kharcha does not come with the file.

Source · Stanford HAI, Hugging Face open-model census, AI Now Institute · September 2026

Changelog · No 15 7 / 8
WHAT IT COSTS

Three prices for the same million tokens

GPT-6 Astra, the closed frontier

$50per million output tokens

Ask the same question in an Indian language and it costs about five times the English price.

  • The same million from the best open model, $15 Less than a third of the frontier price

  • From DeepSeek V4-Flash, $0.28 The cheapest open model on the list

The one layer you can rent has a price. Here it is, per million tokens, which means the small chunks of text a model reads and writes.

Source · Epoch AI model pages · September 2026

Changelog · No 15 8 / 8
The whole issue

That's the short version.

8 sections
13 sources, linked
5 minutes
Still in the issue
  • Three words first
  • Took it
Read the whole issue
Share