By Nokka | September 12, 2026
This article was written by AI (deepseek-v4.1-flash) through Hermes Agent, reviewed and edited by Nokka.
September 14, 2026, 04:00 UTC is about to be a strange date in AI history.
From that second onward, every request sent to deepseek-v4-pro gets routed to V4.1-Flash instead. That will continue until V4.1-Pro ships [1].
Put plainly: DeepSeek is retiring its own flagship name and handing the work to the model labeled Flash.
The company's stated reason is short. Testing by multiple parties found V4.1-Flash beats V4-Pro on capability, cost, speed, and total runtime [1].
The most direct line in the announcement: "We're phasing out V4-Pro."
Looking only at work the model actually has to execute, the gap is not small [2].
| Benchmark | V4.1-Flash | V4-Pro |
|---|---|---|
| Terminal-Bench 2.1 | 90.6 | 87.9 | | Terminal-Bench 4.0 | 31.2 | 12.4 | | DeepSWE v1.1 | 74.2 | 62.7 | | AutomationBench | 54.8 | 43.2 | | Agents' Last Exam | 31.8 | 25.7 | | CyberGym | 88.1 | 83.3 |
That 90.6 on Terminal-Bench 2.1 does not just beat its older sibling in-house. It also beats Claude Opus 5 at 89.1 and GPT-5.6 Sol at 88.8 [3].
The point matters more than it appears, because Opus 5 launched after Opus 4.8. Beating the newer model is a different statement than beating the older one.
All from a model priced under a dollar per million tokens.
Capability did not improve across the board. World knowledge actually regressed [2][3].
| Benchmark | V4.1-Flash | V4-Pro |
|---|---|---|
| GPQA Diamond | 90.9 | 92.4 |
| HLE (no tools, text-only) | 39.1 | 42.7 |
| SimpleQA-Verified (base) | 42.3 | 55.2 |
SimpleQA falling from 55.2 to 42.3 is the clearest regression. That benchmark measures how accurately a model answers factual questions, and it is the dimension people miss, because the headline scores are all about the agentic jump.
The picture that emerges is a model that got better at doing things and knows less about the world.
Many assume all three major labs offer a full lineup of both flagship and Flash tiers. Checking official sources, that is not how it looks [1][4][5].
Z.ai does have the pair. GLM-5.3 at 753 billion parameters and GLM-5.3-Flash at 320 billion, announced together [4].
DeepSeek is all-in on one. V4.1 shipped as Flash first, and V4.1-Pro has no date. Until then, the old model name routes to Flash [1].
Moonshot has neither. Kimi K3 exists in a single variant at 2.8 trillion parameters. The cheaper option is K2.7 Code at $0.95/$4.00, which is a different model, not a scaled-down K3 [5][6].
That difference is not a trivial detail. It says which lab believes the Flash path is the future.
Put the three labs side by side and each is betting differently. Price is not the only axis.
Z.ai bets on shrinking while keeping both. It builds the big model and the small one, and publishes that the small one uses 3.0 times less attention compute and a 4.4 times smaller KV cache than its flagship [4].
That number matters more than it looks, because both support 1 million tokens of context. The cost of serving that context does not match.
DeepSeek bets everything on one lane. It did not make a small model as a secondary option. It moved the entire API to Flash, with a new architecture that is genuinely bigger at 552 billion parameters while activating less per token [1].
It also states KV cache memory drops to one quarter and storage to one eighth [1].
Moonshot bets on size. It chose to build the largest open-weight model in the world and refuses to make a scaled-down price competitor. The result is that K3 is no longer in the cheap market. At $3.00/$15.00 it is more expensive than what Chinese labs used to charge [5][6].
| Model | Total/active params | Price per M (in/out) | AA Index | Cost per task |
|---|---|---|---|---|
| GLM-5.3-Flash | 320B / 18B | $0.15 / $0.50 | 42 | $0.25 |
| DeepSeek V4.1 Flash | 552B / 16B | $0.30 / $1.20 | 40 | $0.27 |
| GLM-5.3 | 753B / 40B | $1.40 / $4.40 | 45 | $2.01 |
| Kimi K3 | 2.8T / 104B | $3.00 / $15.00 | 44 | $2.00 |
The last column comes from Artificial Analysis measurements on Intelligence Index v4.3 [7][8].
Read it and the picture is clear: overall scores differ by little, while cost per task differs by up to eight times.
One caution: these scores depend on the index version too, not just the model. The clearest example is GLM-5.3-Flash scoring 57 on index v4.1.1, the figure the vendor cites in its own launch post [4], but 42 on v4.3 released September 7 [10].
That is 15 points apart without the model changing at all. The test set got harder. When you see two conflicting sets of numbers, check which index version each one cites.
Both GLM-5.3 and Kimi K3 have a mechanism that pushes real cost above what the per-token price suggests: reasoning mode is always on and cannot be switched off [6][9].
Kimi K3 only lets you dial between low, medium, and max, with max as the default. The model was trained to always think, with no off switch [5].
Every request therefore generates reasoning tokens, and those are tokens you pay for.
A second trap is specific to Kimi K3. Multi-turn use requires sending the complete assistant message back, including the reasoning and tool call fields. If your harness only preserves the content field, the model degrades silently with no warning [5].
One Every benchmark figure here comes from individual labs or an independent evaluator choosing its own test sets. Cross-lab comparison requires checking whether the test set is the same, not just the numbers [2][4].
Two A single composite score does not tell you which model fits your job. Long coding tasks and factual Q&A use different numbers, and the results diverge sharply.
Three Published prices usually include promotional windows and cache rates that differ from standard pricing by orders of magnitude. DeepSeek's cache rate sits at $0.006 per million tokens, fifty times below the standard rate [1]. Workloads that reuse context get far cheaper than the sticker price implies.
Four DeepSeek rerouting the old model name means any system still pinned to it will change behavior immediately. Pin model names explicitly and retest before relying on it in production [1].
I run all three labs daily for writing and review, and the lesson here is that picking a model by name or by per-token price is no longer enough.
I used to assume anything labeled Flash is capability reduced to save money. What DeepSeek did flips that assumption, because its Flash model beats the one without the suffix outright.
What I will be watching next is not this round of numbers but how the other two labs respond. If Moonshot actually ships a scaled-down K3, the budget market gets more competitive. And if V4.1-Pro launches and beats Flash, that confirms this story is not finished.
My working advice is not to marry one vendor. Configure your stack so switching models takes one value change, then measure with your own real work whenever a new release lands. This board changes monthly, and a name that once signaled a capability tier no longer signals that.
[1] DeepSeek, "DeepSeek-V4.1-Flash: Smarter, Faster, More Efficient" (Sep 10, 2026), [https://api-docs.deepseek.com/news/news260910/](https://api-docs.deepseek.com/news/news260910/)
[2] DeepSeek API Docs, "Change Log" (2026), [https://api-docs.deepseek.com/updates/](https://api-docs.deepseek.com/updates/)
[3] Coursiv, "DeepSeek V4.1 Flash Replaces V4 Pro: Pricing and Benchmarks" (2026), [https://coursiv.io/blog/deepseek-v4-1-flash](https://coursiv.io/blog/deepseek-v4-1-flash)
[4] Z.ai, "GLM-5.3-Flash: Frontier Intelligence, Flash Cost" (2026), [https://z.ai/blog/glm-5.3-flash](https://z.ai/blog/glm-5.3-flash)
[5] Moonshot AI, "Kimi K3 Tech Blog: Open Frontier Intelligence" (Jul 16, 2026), [https://www.kimi.ai/blog/kimi-k3](https://www.kimi.ai/blog/kimi-k3)
[6] BenchLM.ai, "Kimi API Pricing (September 2026): Kimi K3 at $3/$15" (2026), [https://benchlm.ai/moonshot/api-pricing](https://benchlm.ai/moonshot/api-pricing)
[7] Artificial Analysis, "DeepSeek V4.1 Flash (Reasoning, Max Effort) vs GLM-5.3-Flash" (2026), https://artificialanalysis.ai/models/comparisons/deepseek-v4-1-flash-vs-glm-5-3-flash
[8] Artificial Analysis, "GLM-5.3-Flash vs Kimi K3 (max)" (2026), https://artificialanalysis.ai/models/comparisons/kimi-k3-vs-glm-5-3-flash
[9] OpenRouter, "Z.ai: GLM 5.3 Flash" (2026), https://openrouter.ai/z-ai/glm-5.3-flash [10] Artificial Analysis, "Announcing the Artificial Analysis Intelligence Index v4.3" (Sep 7, 2026), https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v4-3