DeepSeek V4.1 Flash matches V4 Pro accuracy at 3–6x lower cost per answer

What shipped on September 10

DeepSeek released V4.1 Flash on September 10, and a hands-on evaluation published on dev.to works through what changed. The model is described by DeepSeek as the smallest in a new architecture family: 552 billion total parameters arranged in a causal encoder-decoder, with one 20-layer stack reading the prompt and another writing the answer, 8 billion parameters active per input token and 16 billion per output token. It adds native image input — a first for a Flash model — plus a 1M-token context window, a 384K output cap and an MIT license.

Pricing, per DeepSeek's pricing page as relayed by the dev.to post, is $0.30 per million input tokens and $1.20 per million output tokens at peak (01:00–04:00 and 06:00–10:00 UTC, weekdays), halved off-peak, with cache hits at $0.006 and $0.003. V4 Pro sits at $1.32/$3.96 peak. The old V4 Flash is retired, along with the Vision Exp build, and their API names now route to V4.1 Flash at Flash pricing. DeepSeek had also planned to reroute V4 Pro traffic to V4.1 Flash from September 14, but dropped the plan from its change log the same day, citing user demand, so Pro remains available at Pro prices.

The dev.to author notes the plumbing changed less than the price: the tokenizer is identical across the three builds, so token budgets port. Cache pages halved from 1,024 tokens to 512, a 902K-token prompt with a planted value returned it correctly, and prompts beyond the 1M window now return a 400 error rather than silently truncating.

Official numbers versus an independent check

DeepSeek's own model card puts V4.1 Flash above the retired Flash everywhere and above Pro on agent and coding work — Codeforces 3471 versus 3348, Terminal-Bench 4.0 at 31.2 versus 12.4, DeepSWE 74.2 versus 62.7, CyberGym 88.1 versus 83.3, AutomationBench 54.8 versus 43.2 — but behind Pro on knowledge tasks: 90.9 versus 92.4 on GPQA Diamond and 36.8 versus 42.7 on HLE.

The independent test on dev.to covers the checkable half of that table: 11 tasks with a single verifiable numeric answer (prime sums, grid paths, a rule applied 40 times, 7 to the power 222 with the remainder kept, a knapsack, and others), run three times each. With thinking enabled, all three builds scored 33 of 33, so the suite separated them on cost and speed rather than accuracy. At peak pricing, a correct answer cost $0.00097 to $0.00149 on V4.1 Flash depending on reasoning effort, against $0.00286 to $0.00825 on V4 Pro — a factor of 2.9x at low effort, 5.2x at the default high and 6.4x at max — and 26% less than V4 Flash at low effort.

Speed favoured the new model too. Streamed with thinking off, V4.1 Flash decoded at 121–134 output tokens per second with 1.4 seconds to first token, versus 86–94 tokens per second for V4 Flash and 46–49 for Pro. In a two-turn function-calling loop, every build made exactly one call per turn, with V4.1 Flash costing $0.00019 per leg against $0.00103 on Pro.

Behaviours to design around

The dev.to evaluation flags two failure modes. Turning thinking off drops every build to 21 of 33, and V4.1 Flash in that mode tends to answer multi-step arithmetic with a single incorrect token rather than failing visibly. On five questions about entities that do not exist, V4.1 Flash with thinking on exhausted the full 16,384-token output window in three of five runs and invented an answer in the other two; with thinking off, it declined all five. In other words, the hallucination guard behaviour inverts depending on the thinking switch, which is worth probing before a migration.

The reasoning controls are also narrower than documented. The API accepts reasoning_effort as low, high or max; thinking_budget values from 0 to 1,024 are accepted but ignored, and the model card's claimed continuous integer 1–100 control is not exposed on the API, which rejects integers.

Why it matters

For API users doing verifiable work — agent tasks, coding, math with a checker — the price of a correct answer has collapsed, and anyone still calling the retired V4 Flash names is already being rerouted to the new model whether they planned it or not, with a different pricing structure attached. The remaining argument for V4 Pro is the knowledge gap on benchmarks like HLE and GPQA, and evidently enough users agreed that DeepSeek reversed its plan to fold Pro traffic into Flash. The practical takeaway from the dev.to data: the savings are real, but the single-wrong-token failure mode with thinking disabled means the cheap setting should ship with checks, not as a default.