Codeep 3.3.0 — DeepSeek V4.1 Flash, Qwen 3.8, and four prices that were wrong

[VladoIvankovic]

A models release, with a deadline behind it and a correction owed.

CLI 3.3.0 — npm install -g codeep
macOS 3.3.0 — https://codeep.dev/download/mac, or let the app update itself

---

DeepSeek V4.1 Flash is now the DeepSeek model

Its API id is deepseek-flash — not deepseek-v4.1-flash, which is what OpenRouter calls the same model. If you set the id by hand anywhere, that difference matters.

DeepSeek's own pricing page names it the model to use and says it outperforms V4 Pro. V4 Flash is already retired and served by it, and from 14 September, 12:00 Beijing time (04:00 UTC), every deepseek-v4-pro request is routed to V4.1 Flash and billed at its price, until a V4.1 Pro exists.

So Codeep now offers one DeepSeek model, it is the default, and a config set to either V4 id moves to it on its own. Keeping a "V4 Pro" entry after the 14th would have meant a Pro label on a Flash model. /thinking on it gains Low — V4.1 tells low, high and max apart; V4 didn't.

If you're on an older Codeep with V4 Pro selected, nothing breaks on Monday. You just silently get Flash under the old name, with the old estimate. Upgrading fixes both.

Qwen 3.8, and GLM-5.3 in China

• Qwen3.8-Max and Qwen3.8-Flash on Qwen API (pay-per-use), 3.8-Max the default. Both have a 1M window and both cost less than what they sit above — $2/$6 against 3.7-Max's $2.50/$7.50, and $0.15/$0.47 against 3.6-Flash's $0.25/$1.50. The Coding Plan's published allowlist doesn't include 3.8 yet, so it is unchanged.
• GLM-5.3 on Z.AI China, plan and pay-per-use; GLM-5.3 Flash on China pay-per-use.

---

Four prices were wrong

Codeep shows what a run cost, so a wrong price is worse than no price. A full review against every provider's own page turned up four:

	was	is
DeepSeek V4 / V4.1 Flash	$0.14 / $0.28	$0.30 / $1.20
DeepSeek V4 Pro	$0.435 / $0.87	$1.32 / $3.96
GPT-5.6 Sol	$5 / $30	$4 / $20 (promo, through at least 21 Nov)
MiniMax-M3	$0.60 / $2.40	$0.30 / $1.20 (up to 512K prompt tokens)

DeepSeek is the one that matters. DeepSeek bills peak and off-peak, and Codeep deliberately carries the peak rate so an estimate can run high but never low. These had drifted two to four and a half times below peak — the one direction the table must never err in.

And a correction to our own 3.2.0 announcement. It said GPT-6 Astra is "twice the price" of GPT-5.6 Sol. Against what Sol is actually billed today, Astra is two and a half times Sol. The gap was bigger than we told you.

Caught before shipping

Three places decide things by model name, and the new ids slipped past all three in testing:

• The Mac's context meter would have shown V4.1 Flash at 128K instead of 1M — eight times too small, with compaction free to start far too early.
• The thinking control, on both clients, would have been hidden on it.
• Qwen3.8-Max on pay-per-use would have been priced at zero, because of a rule written for the Token Plan preview.

Each is now held by a test. The same shape of bug nearly hid /thinking on GPT-6 last week, so the model-maintenance checklist now says it plainly: when an id changes shape, search every capability check for the old name.

One more: the OpenRouter shortlist offered qwen/qwen3.8-max, which OpenRouter doesn't carry — only its -0902 snapshot. Picking it before the live catalogue loaded failed. Fixed.

---

Worth watching

• DeepSeek V4.1 Pro — when it ships, V4 Pro routing ends and a Pro model returns.
• Claude Haiku 4.5 may retire from 15 October, with no replacement named yet.
• DeepSeek cache hits arrive in a field Codeep doesn't read yet, so estimates for long cached sessions run high. Corrected below — this was wrong.

Full notes: CLI · macOS

Something behaving oddly with V4.1 Flash or Qwen 3.8? Ask in Q&A — both are new, and the details differ by provider.

---

Correction, 14 September. This post said Codeep couldn't read DeepSeek's cache hits. That was wrong. DeepSeek reports the count in two fields, and Codeep has always read one of them.

The real problem was the rate. DeepSeek had no cache price of its own in Codeep, so cached tokens were estimated at the 0.1× default instead of DeepSeek's actual 0.02× — five times too high. Separately, the "saved via caching" figure used 0.1× for every provider, so it was wrong for Kimi, Qwen, DeepSeek and Claude Fable 5.1.

Both are fixed in CLI 3.3.2 and macOS 3.3.1. The mistake came from trusting a summary of DeepSeek's API reference that had the fields nested wrongly; reading the reference itself settled it.