What developers on one Indonesian AI gateway used last week (5-12 Nov 2026)

Published · 5 min read

Disclosure: this article is by the team behind Ezkielyna, the gateway this data comes from. The dataset is published under CC BY 4.0, so you can check and reuse every number here. Drafted with AI assistance and edited by the team.

Methodology, read this first

  • Source: the weekly leaderboard of one gateway (Ezkielyna), status/leaderboard.json, generated 12 Nov 2026 01:23 WIB.
  • Window: 7 days, 5 Nov 01:23 to 12 Nov 01:23 WIB, compared with the 7 days before.
  • Inclusion: a model is listed only if at least 3 distinct users called it in the window. The 20 listed models account for 97.59% of the requests in the leaderboard's share column; the rest are models below the threshold.
  • Unit: requests (successful answers counted by the gateway) and tokens as logged. Prompt tokens may include cached tokens.
  • Not in the data: anything per user, per country, or per prompt content. User counts are per model and cannot be added up across models.
  • Bias: one gateway, mostly Indonesian customers who pay with QRIS, and a price list that makes some models far cheaper than others. This describes our users, not Indonesia, and not "the market".

The headline numbers

  • 100,722 requests to the 20 listed models in 7 days.
  • 11.79 billion tokens processed for them: 11.72 billion prompt tokens and 74.2 million completion tokens.
  • Prompts outweighed outputs about 158 to 1.
  • Two model ids took 85.12% of all requests: deepseek-v4.1-flash-1017 (46.56%) and deepseek-v4.1-flash (38.56%).

Share by vendor (requests, listed models only)

VendorModels listedRequestsShare of requests
DeepSeek591,05688.22%
OpenAI66,6466.43%
Anthropic41,3831.35%
Zhipu AI21,1061.07%
MiniMax11980.19%
Google11710.17%
xAI11620.16%

The obvious reading: on a gateway where paying is easy and per-token prices are public, an open-weight "Flash" model carries most of the work. DeepSeek V4.1 Flash was listed at $0.04 per 1M input tokens and $0.16 per 1M output tokens on 12 Nov 2026.

The less obvious reading: closed models still matter for some jobs. OpenAI models had 6 entries in the top 20 and Anthropic 4, each with between 3 and 14 users. A small group of users pays more per request for specific tasks.

The top 10

#ModelVendorRequestsUsersSharevs prior week
1deepseek-v4.1-flash-1017DeepSeek48,0534546.56%-48.9%
2deepseek-v4.1-flashDeepSeek39,7995838.56%-52.3%
3gpt-6-lunaOpenAI2,112142.05%-29.8%
4deepseek-v4-proDeepSeek2,096102.03%+40.3%
5gpt-5.6-terraOpenAI1,810101.75%+4.0%
6gpt-5.6-lunaOpenAI1,156121.12%+34.3%
7deepseek-v4-flashDeepSeek887100.86%-54.7%
8gpt-6.1-terraOpenAI68340.66%+634.4%
9claude-haiku-5-5Anthropic65030.63%+41.3%
10gpt-5.6-solOpenAI63450.61%+67.7%

The full list of 20 is in the dataset.

Finding 1: this is agent traffic, not chat traffic

A chat app sends short prompts and gets medium answers. Our traffic looks nothing like that. Requests to deepseek-v4.1-flash averaged about 112,750 prompt tokens and about 719 completion tokens, and deepseek-v4-pro averaged about 167,900 prompt tokens per request. That is the signature of coding agents re-sending a large working context on every step and getting a short edit or tool call back.

The client data agrees, with a caveat. We only started counting client apps on 10 Nov 2026 at 14:59 WIB. In the 34 hours since, the one client above the reporting threshold (3 distinct clients and 50 requests) was OpenCode, with 4,183 requests.

Implications if you build agents:

  • Price input, not output. At a 158:1 ratio, the input price and the cache discount decide your bill. The output price barely registers.
  • Latency is prefill-bound. On the same day, the median time to first token for DeepSeek V4.1 Flash was 3,844 ms (status feed, 12 Nov 01:30 WIB). Much of that is reading 100k+ tokens, not writing.

Finding 2: concentration is extreme, and that is a risk

Two ids of the same model family carried 85% of requests. For users, that is rational: the model is good enough for most agent steps and costs little. For the gateway, it means one upstream problem with one model family touches most users.

That is why we publish per-model success and latency publicly (api.ezkielyna.store/status). If the model most people depend on has a bad day, they should be able to see it.

Finding 3: total volume fell, and we can't say why from this data

Across the 20 listed models, requests fell from 189,823 in the previous week to 100,722 (-46.9%). Almost all of the drop is the two DeepSeek Flash ids (-48.9% and -52.3%).

A few smaller models grew fast from small bases:

  • gpt-6.1-terra: from 93 to 683 requests, 4 users.
  • glm-5.3: from 120 to 482, 4 users.
  • claude-sonnet-5.5: +313.0%, 6 users.

Percentages like these, on a handful of users, are noise until they persist.

A falling week can come from a few heavy users changing their workload, from pricing or plan changes, or from routing. This file does not separate those causes, so we won't guess.

What this data cannot tell you

  • Which model is better. Volume follows price and habit as much as quality.
  • Anything about Indonesia as a whole. One gateway's customers and a skewed price list.
  • Quality of answers. We count requests, not outcomes.
  • Cache effects. Prompt tokens as logged may include cache hits that were billed at a discount.

Get the data

If you reuse it, please credit "Ezkielyna (CC BY 4.0)" and link the dataset. If you find an error, open an issue on the repo. We'd rather fix a number than defend it.

Ezkielyna is an independent API gateway: one OpenAI- and Anthropic-compatible key for many AI models, prepaid with QRIS, USDT or Binance Pay, no credit card needed. It is not affiliated with the companies that make the models.

About Ezkielyna API →