Verdict: if you are an existing deepseek-v4-pro API user, let the September 14 reroute happen. V4.1-Flash wins this comparison for agentic and coding work: it is cheaper on every published rate line and ahead of V4-Pro on DeepSeek's own agentic benchmarks. V4-Pro's genuine edges — world-knowledge recall, long-context QA and hard maths — are lost either way, because the endpoint itself is being retired until a V4.1-Pro arrives with no announced date (DeepSeek API news).
From 04:00 UTC on 14 September 2026, DeepSeek states that "all deepseek-v4-pro requests will route to V4.1-Flash at V4.1-Flash rates. This will continue until V4.1-Pro launches" (DeepSeek API news). There is no migration to write. There is a bill to re-forecast.
TL;DR
- V4-Pro is being phased out by its own lab, in favour of a smaller and cheaper model; the cutover is automatic at 04:00 UTC on 14 September 2026 (API news).
- V4.1-Flash is a 552B-parameter Mixture-of-Experts model with native image and text input, an MIT licence and a 1M-token context window (model card).
- On the published off-peak rates, cache-hit input is 7.3x cheaper, cache-miss input 4.4x cheaper and output 3.3x cheaper than V4-Pro (pricing).
- V4.1-Flash leads V4-Pro on Terminal-Bench 2.1, DeepSWE v1.1 and CyberGym in DeepSeek's own harness, and trails it on SimpleQA-Verified, MATH and LongBench-V2 (model card).
- Independent evaluation is still thin: as of 12 September the only third-party score published was Artificial Analysis's Intelligence Index reading of 40.
Last verified: 2026-09-13.
What happens to deepseek-v4-pro on September 14
Nothing you have to do, and nothing you can prevent. DeepSeek's changelog says that "Tests by multiple parties put V4.1-Flash ahead of V4-Pro on performance, cost, speed & total runtime. We're phasing out V4-Pro," and that from 04:00 UTC on 14 September 2026 every deepseek-v4-pro request is served by V4.1-Flash and billed at Flash rates (API news).
This is the same soft-retirement pattern DeepSeek used before: the older deepseek-v4-flash and deepseek-v4-flash-vision-exp names already resolve to V4.1-Flash, and the deepseek-chat and deepseek-reasoner aliases were fully retired on 24 July 2026 (API news). The model string keeps working; the model behind it changes. If your evaluation suite pins behaviour rather than the name, run it again on 15 September.
Third-party reporting matches the changelog on both the date and the rates, including the RMB-denominated off-peak figures of 0.02 cache-hit, 1 per million cache-miss input and 4 per million output (TechNode), and a separate explainer confirms the reroute holds "until V4.1 Pro is released" (changelog explainer).
How V4.1-Flash and V4-Pro compare on specs and price
V4.1-Flash is the smaller, sparser and cheaper model, and it is the one that survives. V4-Pro was a 1.6T-parameter model with 49B active per token at its April 2026 preview (API news). V4.1-Flash is 552B total parameters using a new Causal Encoder-Decoder design that activates only 8B parameters per token during prefill and 16B during decode, with the decoder KV cache projected from the final encoder hidden states (model card).
| DeepSeek V4-Pro | DeepSeek V4.1-Flash | |
|---|---|---|
| Parameters | 1.6T total / 49B active (news) | 552B total / 8B prefill, 16B decode (card) |
| Context | — | 1M tokens (card) |
| Weights | API and app only, no updated Pro weights published | MIT licence (card) |
| Off-peak cache-hit input / 1M | $0.022 (pricing) | $0.003 (pricing) |
| Off-peak cache-miss input / 1M | $0.66 (pricing) | $0.15 (pricing) |
| Off-peak output / 1M | $1.98 (pricing) | $0.60 (pricing) |
| Peak cache-miss input / output / 1M | $1.32 / $3.96 (pricing) | $0.30 / $1.20 (pricing) |
| Concurrency | 500 (pricing) | 2,500 (pricing) |
Peak rates are double off-peak on both models, and peak hours run 01:00–04:00 and 06:00–10:00 UTC, Monday to Friday (pricing). Two engineering details explain the cost gap: the global KV cache is 890 bytes per token, about a quarter of DeepSeek-V4-Flash, and the news page puts the serving requirement at a quarter of the HBM and an eighth of the SSD storage of the previous generation (card, API news). The model was trained from scratch on 45T multimodal tokens, with sparse attention trained at 64K sequence length and extended to 1M at the 34T-token mark (card).
Taking the published off-peak rates above, a workload consuming 1M cache-hit input, 1M cache-miss input and 1M output tokens every day costs about $79.9 per month on V4-Pro rates against about $22.6 per month on V4.1-Flash rates — roughly 3.5x less for identical work (pricing). If you want the fuller price-per-quality picture for the outgoing model, see our earlier analysis of V4-Pro's cost-quality position.
Which model actually scores better on agentic benchmarks
V4.1-Flash leads on agentic coding and trails on knowledge and long context. In DeepSeek's own harness, with instruct models at maximum reasoning effort, V4.1-Flash reaches 90.6 on Terminal-Bench 2.1 against V4-Pro's 87.9 and Claude Opus 5's 89.1; 74.2 on DeepSWE v1.1 against V4-Pro's 62.7; and 88.1 on CyberGym against 83.3 (model card).
The ceiling is visible on newer and harder tests. V4.1-Flash scores 30.0 on Terminal-Bench 3.0 where Opus 5 records 43.3, 90.9 on GPQA Diamond against 93.4, and 36.8 on Humanity's Last Exam without tools against 56.3 (model card). At base-model level, V4.1-Flash-Base edges V4-Pro-Base on MMLU-Pro (74.1 vs 73.5), HumanEval (79.4 vs 76.8), BigCodeBench (60.6 vs 59.2) and Codeforces rating (3471 vs 3348), while V4-Pro-Base holds SimpleQA-Verified (55.2 vs 42.3), MATH (64.5 vs 61.1), MultiLoKo (50.9 vs 45.5) and LongBench-V2 (51.5 vs 45.2) (model card).
Treat all of that as vendor-reported. As of 12 September, the only published third-party measurement was Artificial Analysis's Intelligence Index score of 40 at $0.30 per million input and $1.20 per million output peak rates, placing it just below GLM-5.3-Flash and above the outgoing V4-Pro (collected in Latent Space's AINews digest, 12 September 2026). Every agentic figure in circulation is DeepSeek's own. Our head-to-head against rival frontier models sits in the V4.1-Flash vs Opus 5 vs GPT-5.6 Sol vs Kimi K3 comparison, and the small-model field is covered in our GLM-5.3-Flash vs Qwen 3.8 Flash Next comparison.
Who gains most from the switch, and what you lose
Nobody can stay on V4-Pro, so the practical question is who gains. Heavy cache users gain most: cache-hit input falls from $0.022 to $0.003 per million tokens off-peak, a 7.3x cut, which matters for long system prompts and repeated document context (pricing). Agent operators gain twice, because the concurrency ceiling moves from 500 to 2,500 on the Flash tier (pricing).
The losses are real. Workloads that depended on V4-Pro's stronger factual recall, long-context QA or maths lose those numbers with no drop-in replacement, since the Pro checkpoint stayed API-and-app-only after its 13 August 2026 general availability and no updated Pro weights were published. There is no V4.1-Pro date. If SimpleQA-style recall is load-bearing in your pipeline, plan a second model rather than waiting.
Across three trials each on an identical seven-constraint article-planning task, Gemini 3.8 Flash (High) and Claude Opus 4.6 (Thinking) both scored 17 of 17 on machine-checked constraint adherence. Median wall time was 23 seconds for Gemini against 67 seconds for Opus (n=6, measured 2026-09-13). A cheap model matching an expensive flagship on a tightly constrained production task, at a fraction of the runtime, is the same shape as DeepSeek routing everyone to the cheaper model.
For hands-on setup, our V4-Flash guide covers the endpoint basics, and Kimi K3 vs Claude Fable 5 for coding covers alternatives if you want a second provider in the loop.
FAQ
Q: Do I need to change my code before September 14?
A: No. The deepseek-v4-pro model string keeps working and is served by V4.1-Flash from 04:00 UTC on 14 September 2026, billed at Flash rates (API news). Re-run your evaluation suite afterwards, because the model behind the name changes.
Q: Will my bill go up or down?
A: Down, on every published rate line. Off-peak, cache-hit input drops from $0.022 to $0.003, cache-miss input from $0.66 to $0.15 and output from $1.98 to $0.60 per million tokens (pricing).
Q: Is V4.1-Flash actually better than V4-Pro?
A: On agentic coding, per DeepSeek's own harness, yes: 90.6 vs 87.9 on Terminal-Bench 2.1 and 74.2 vs 62.7 on DeepSWE v1.1. On world knowledge and long context, no: V4-Pro-Base leads on SimpleQA-Verified and LongBench-V2 (model card).
Q: When does V4.1-Pro arrive?
A: DeepSeek has not announced a date. The changelog says the reroute continues until V4.1-Pro launches, with no timeline attached (API news).
Q: Can I run V4.1-Flash myself?
A: Yes. V4.1-Flash weights are published under an MIT licence, unlike V4-Pro, which remained API-and-app-only (model card).
Q: How much of the benchmark data is independent?
A: Very little. As of 12 September, the only third-party figure published was Artificial Analysis's Intelligence Index score of 40 (Latent Space's AINews digest); the agentic results are all vendor-reported.
Last verified: 2026-09-13. Corrections and updates are logged on this page. This article was researched and drafted with AI assistance under named human accountability; see how we work.
Top comments (0)