v3.3.0

TL;DR — DeepSeek V4.1 Flash replaces both V4 models before DeepSeek routes V4 Pro to it on 14 September. Qwen 3.8 Max and Flash, GLM-5.3 in China, and four cost figures corrected — DeepSeek's by more than four times.

Added

• DeepSeek V4.1 Flash is the DeepSeek model, and the default. Its API id is
  deepseek-flash — not deepseek-v4.1-flash, which is the id OpenRouter uses
  for the same model. DeepSeek's own pricing page names it the model to use and
  says it outperforms V4 Pro. deepseek-v4-flash is already retired and served by
  it, and from 14 September, 12:00 Beijing time, every deepseek-v4-pro request
  is routed to V4.1 Flash and billed at its price until a V4.1 Pro exists.

  So the picker now offers one DeepSeek model, and a config holding either V4 id
  moves to deepseek-flash automatically. Keeping a Pro entry would have meant a
  Pro label on a Flash model. /thinking on it now offers Low as well as High
  and Max — V4.1 tells the three apart; V4 did not.

• Qwen3.8-Max and Qwen3.8-Flash on Qwen API (pay-per-use), with 3.8-Max as the
  default. Both are generally available on Model Studio international, both have a
  1M window, and both cost less than what they sit above: 3.8-Max is $2/$6
  against 3.7-Max's $2.50/$7.50, and 3.8-Flash $0.15/$0.47 against 3.6-Flash's
  $0.25/$1.50. The 3.7 models stay beside them. Qwen China and the Coding Plan are
  unchanged — the Coding Plan's published allowlist does not include 3.8.

• GLM-5.3 on Z.AI China, on both the GLM Coding Plan and pay-per-use, now as
  the default there. GLM-5.3 Flash on China pay-per-use only — the China plan
  does not list it.

Fixed

• DeepSeek estimates were two to four and a half times too low. The pricing
  rows no longer matched DeepSeek's peak rates — V4 Flash read $0.14/$0.28 against
  $0.30/$1.20, V4 Pro $0.435/$0.87 against $1.32/$3.96. Codeep carries the peak
  rate on purpose, so an estimate can run high but never low; this ran low, which
  is the one direction it must not. Cache hits are still billed at the miss rate:
  DeepSeek reports them in a field this does not read yet, so the error stays on
  the high side.

• GPT-5.6 Sol is $4/$20, a promotion OpenAI says runs at least through 21
  November, not the $5/$30 it was priced at. Correction: the 3.2.0 notes called
  GPT-6 Astra "twice the price" of Sol. Against what Sol is billed today, Astra is
  two and a half times.

• MiniMax-M3 is $0.30/$1.20 for prompts up to 512K tokens. The $0.60/$2.40 it
  was priced at is the long-context tier, so nearly every estimate was doubled.

• The OpenRouter fallback list offered a model that does not exist.
  qwen/qwen3.8-max is not on OpenRouter — only the qwen3.8-max-0902 snapshot
  is — so choosing it before the live catalogue loaded failed. Also: DeepSeek V4
  Pro there becomes V4.1 Flash, and GPT-6 Astra and Gemini 3.8 Flash, missing from
  this list since 3.2.0, are added.

• GLM-5.3 Flash was described as "a twentieth of the price" of GLM-5.3. At
  $0.15/$0.50 against $1.40/$4.40 it is about a ninth.

• Kimi K3's context window is 1,048,576 tokens, as Moonshot lists it, not a
  round million.

Caught before shipping

deepseek-flash does not start with deepseek-v4, which is what the thinking
control checked for — so the new default would have arrived with /thinking
hidden and nothing to explain it. It is the same trap GPT-6 nearly fell into in
3.2.0 behind a gpt-5 prefix, and the model-maintenance checklist now says to
search every capability check whenever an id changes shape.