DeepSeek to reroute all V4-Pro API traffic to cheaper V4.1-Flash on September 14

DeepSeek is retiring its flagship name

Starting at 04:00 UTC on September 14, 2026, DeepSeek will redirect every API call addressed to deepseek-v4-pro to V4.1-Flash instead, and it will keep doing so until a V4.1-Pro model ships — a release that currently has no announced date. According to a dev.to post by Nokka citing DeepSeek's announcement, the company summarized the move in one line: "We're phasing out V4-Pro." Its stated rationale is that testing by multiple parties found V4.1-Flash ahead of the old flagship on capability, cost, speed and total runtime.

Two caveats on provenance: the two dev.to sources for this story are English and Thai versions of the same piece, which was drafted by an AI agent and edited by its author. The underlying figures come from DeepSeek's API documentation and announcements, competitor launch posts, Coursiv and Artificial Analysis, so cross-vendor comparisons deserve some skepticism.

Agentic benchmarks jumped, factual ones slipped

DeepSeek's changelog, as quoted in the dev.to post, shows V4.1-Flash clearly ahead on hands-on work: 74.2 versus 62.7 on DeepSWE v1.1, 54.8 versus 43.2 on AutomationBench, 31.8 versus 25.7 on Agents' Last Exam and 88.1 versus 83.3 on CyberGym. The widest gap is Terminal-Bench 4.0, where Flash scores 31.2 against V4-Pro's 12.4.

On Terminal-Bench 2.1, Flash's 90.6 also edges out Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8, according to Coursiv — a comparison against competitors' newer releases, not merely DeepSeek's own previous generation.

The same tables show regressions elsewhere. SimpleQA-Verified, which measures factual answer accuracy, drops from 55.2 on V4-Pro to 42.3 on Flash. GPQA Diamond slips from 92.4 to 90.9, and text-only HLE from 42.7 to 39.1. The emerging picture is a model that got better at executing tasks while knowing somewhat less about the world, a trade-off that will hit factual Q&A workloads even as agentic headlines dominate.

A larger model that costs less to run

V4.1-Flash is listed at 552 billion total parameters with 16 billion active per token, priced at $0.30 per million input tokens and $1.20 per million output tokens. DeepSeek says the new architecture cuts KV cache memory to a quarter and cache storage to an eighth of the previous design. Its cache-hit rate of $0.006 per million tokens is fifty times below the standard price, which matters for workloads that reuse context heavily.

Rivals are betting differently

The dev.to post frames DeepSeek's move against two competitors taking opposite paths. Z.ai ships both a flagship GLM-5.3 (753 billion parameters, $1.40/$4.40) and GLM-5.3-Flash (320 billion total, 18 billion active, $0.15/$0.50), claiming the smaller model needs 3.0 times less attention compute and a 4.4 times smaller KV cache while matching the 1M-token context window.

Moonshot sells a single Kimi K3 variant at 2.8 trillion parameters for $3.00/$15.00, with reasoning mode always on and no way to disable it — adjustable only between low, medium and max, defaulting to max — so every request bills reasoning tokens. Artificial Analysis figures cited in the post put cost per task at roughly $0.25 to $0.27 for the two Flash models versus about $2.00 to $2.01 for the flagships, despite modest differences in composite scores.

Read the comparisons carefully

That comparison carries a warning the post itself flags: Artificial Analysis released Intelligence Index v4.3 on September 7, and GLM-5.3-Flash scores 57 on index v4.1.1 but 42 on v4.3 — the same model against a harder test set. Conflicting numbers usually mean different index versions, not model changes. Benchmark figures also come from vendors or evaluators choosing their own test sets, and a single composite score says little about whether a model fits a specific job.

What to do before September 14

Anyone with systems pinned to deepseek-v4-pro will see behavior change at the cutover with no opt-in. The practical advice from the report: pin model names explicitly, retest before relying on them in production, and measure new releases against your own real workloads rather than headline scores, since the knowledge regressions may outweigh agentic gains for some tasks.

Why it matters

A model labeled Flash beating the flagship it replaces inverts the convention that a suffix signals a cut-down tier. If the reported numbers hold, model names no longer tell you capability, and per-token price tells you even less — V4.1-Flash outscored V4-Pro on agentic tasks at a lower cost. The deprecation also shows how differently the major Chinese labs now structure their lineups: DeepSeek consolidates everything into one efficient lane, Z.ai keeps a two-tier pair, and Moonshot pushes a single very large model with billing behavior that can quietly inflate bills. Whether V4.1-Pro eventually beats its cheaper sibling, and whether Moonshot answers with a smaller K3, will decide if this was a one-off or the new shape of the market.