Requests to deepseek-v4-pro start returning V4.1-Flash on 14 September at 04:00 UTC

The cutover asks nothing of your code. A client with `deepseek-v4-pro` in a config file keeps getting valid responses after 04:00 UTC on 14 September; the model behind the string is V4.1-Flash, and it stays that way until V4.1-Pro ships ^[1]. DeepSeek published the change in advance, so the way to see it in your own traffic is to diff outputs across the boundary. The dev.to write-up documenting the switch went up on 12 September, which leaves about two days ^[21][3].

The stated reason is short. Testing by multiple parties found V4.1-Flash beat V4-Pro on capability, cost, speed and total runtime, according to the post ^[2]. It quotes the announcement's most direct line: "We're phasing out V4-Pro." ^[3]

The two benchmarks in play measure different things. Terminal-Bench 2.1 scores work the model has to execute, and V4.1-Flash posts 90.6 there, above Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8 ^[4][5]. SimpleQA measures how accurately a model answers factual questions, and it falls from V4-Pro's 55.2 to 42.3 ^[6][7]. That is 12.9 points, a 23 percent relative decline ^[1]. For the 90.6 to describe your system, your workload has to look like the harness that produced it: shell and tool loops with a pass condition. If your product answers from model memory without retrieval, the SimpleQA number is the one that transfers, and that one fell.

Cross-lab comparison in the post rests on Artificial Analysis Intelligence Index v4.3, where overall scores differ by little and cost per task differs by up to eight times ^[16]. The index version moves the score more than the model does in at least one case. GLM-5.3-Flash scored 57 on index v4.1.1, the figure Z.ai cites in its own launch post, and 42 on v4.3, released 7 September ^[17]. Same weights, 15 points apart, a 26 percent drop, because the test set got harder ^[2].

V4.1 shipped as Flash only. DeepSeek has not given a ship date for V4.1-Pro, and the API points at Flash in the interim ^[9]. The new architecture is larger at 552 billion parameters while activating less per token, and DeepSeek states KV cache memory drops to one quarter and storage to one eighth ^[10][11]. Z.ai announced GLM-5.3 at 753 billion parameters and GLM-5.3-Flash at 320 billion on the same page, and published that the smaller model uses 3.0 times less attention compute and a 4.4 times smaller KV cache, with both supporting 1 million tokens of context ^[12][13]. Moonshot's Kimi K3 exists in one variant at 2.8 trillion parameters, priced at $3.00 and $15.00 per million tokens ^[14][15].

The DeepSeek saving depends on prefix reuse. Its cache read price is $0.006 per million tokens, fifty times below the normal rate ^[20], and V4.1-Flash sits under a dollar per million tokens to begin with ^[8]. With its two nearest rivals, the quoted price understates what you pay: reasoning mode is always on in GLM-5.3 and Kimi K3 and cannot be turned off, and K3 only dials between low, medium and max, with max as the default ^[18].

One provenance note, since the numbers all arrive through one post. The write-up was produced by deepseek-v4.1-flash through Hermes Agent, edited by Nokka, and published on dev.to on 12 September ^[21]. Nokka states in it that every cited test figure comes from a lab or an independent measurer that chose its own test set, so a cross-lab comparison means checking whether the set is the same ^[22].