V4.1 Flash rewrites the price list, and Harness changes its session format

This was the week the model layer got cheaper and the runtime layer got less stable. DeepSeek released V4.1 Flash on September 10, cut Flash pricing by up to 60% the same day, and began retiring the flagship that the new model is said to outperform. DeepSeek Harness, meanwhile, shipped four releases in four days, one of which upgrades the session format in a way you cannot undo.

DeepSeek: the smallest model retires the biggest one

V4.1 Flash is a 552B-parameter mixture-of-experts model built on a new Causal Encoder-Decoder architecture, with only 8B parameters active on input and 16B on output. It handles images natively, takes a 1M-token context, exposes a reasoning_effort setting from 1 to 100, and ships weights on Hugging Face under an MIT license alongside a technical report.

The commercial decisions are the interesting part. DeepSeek says V4.1 Flash beats V4 Pro on performance, cost, speed and total runtime, and is acting accordingly: from 12:00 Beijing time on September 14, requests to deepseek-v4-pro will route to V4.1 Flash and be billed at Flash rates, until a V4.1 Pro exists. V4 Flash and V4 Flash Vision Exp are already retired, with their old model names temporarily pointing at the new one. For a period, DeepSeek will have no pro-tier model on sale at all — it is routing flagship traffic into the cheap model and calling that an upgrade.

The price list moved at the same moment. Off-peak rates are now ¥0.02 per million cache-hit tokens, ¥1 per million cache-miss tokens and ¥4 per million output tokens, with peak rates double and weekends billed off-peak. That is a 60% cut on cache hits, 33% on cache misses, and 11% on output — a shape that pays high-reuse agent workloads far more than generation-heavy ones. We broke down the arithmetic in a separate post the same day.

The architecture explains the pricing. Decoder global KV comes from the encoder's final hidden states instead of per-layer caches, and an FP4 cache brings storage to roughly 890 bytes per token — DeepSeek reports a quarter of the HBM and an eighth of the SSD footprint of the previous generation, and a KV cache 437 times smaller than its first-generation model. Cache-hit charges dominate long agent sessions, so compressing the cache is the direct route to a lower bill.

The gains are not uniform, and DeepSeek published the exceptions in the same table: V4.1 Flash trails V4 Pro on SimpleQA-Verified (42.3% vs 55.2%), LongBench-V2 (45.2% vs 51.5%) and HLE without tools (36.8% vs 42.7%). Agentic suites moved the other way — Terminal-Bench 2.1 from 82.7% to 90.6%, DeepSWE v1.1 from 54.4% to 74.2%. Smaller, cheaper and better at tool use is a real trade, but it is a trade: if your workload is knowledge-heavy rather than agentic, the September 14 reroute is a downgrade. Our fuller read, including third-party results and the harness-sensitivity caveat, is here.

One more thread worth tracking: reports this week said DeepSeek has begun preparing a STAR Market listing in Shanghai and has engaged CITIC Securities, alongside a large hiring round. Price cuts, MIT-licensed weights and open deployment talks read differently against an IPO calendar — they are volume and ecosystem plays, not margin plays.

DSH this week: four releases, one of them a one-way door

Zoom out and the version line itself is a story. v0.1.2-rc.1 shipped on September 3 and was superseded 29 hours later by v0.1.3-alpha.1; the 0.1.2 line never reached a stable release, and no 0.1.4 ever appeared publicly. Four days later the project was at 0.1.5-rc.2. Session format went from v2 to v3 inside a week. A community write-up on Cnblogs made the useful point: an rc promises "this version is coming", not "this version will ship" — and the only stable contract in the repository is the README's developer-preview warning, which has not moved in weeks. Treat the version number as a record of when the project re-assessed itself, not as a readiness signal.

The thread connecting both: the harness is now part of the model

V4.1 Flash launched on September 10; the same day, DSH made it the default model for new sessions. That is fast, and it is also the point. When one vendor ships the model, the runtime and the evaluation scaffold, "official benchmark" quietly comes to mean "our stack" — DeepSeek's code-agent numbers were produced on its own minimal Harness profile with a 1M-token context, which is a legitimate setup but not a neutral one.

The size of that effect is not hypothetical. An engineering evaluation reported this week found a 16.77-point drop when the same model was moved into DeepSeek Harness, with functional correctness of 78 against Qwen 3.8 Flash at 82. Harness choice is not a detail you delegate to a vendor chart. If you run agents in production, the unit you should be versioning is the pair — model plus harness — and the evaluation that matters is the one you own.

Elsewhere in AI

• NVIDIA is buying Hugging Face for more than $12.9B. The open-weights hub landing inside the GPU vendor that profits from serving those weights raises an obvious question about neutrality, and nobody has answered it yet.
• OpenAI shipped GPT-Image 2.5 with reported latency reductions around 50%. Smaller item from the same release: generated images still carried 2.0 metadata at launch, which tells you something about how fast these rollouts move.
• OpenAI is co-developing next-generation chips with Samsung, extending a relationship that started around memory toward foundry work — the compute supply chain keeps moving in-house.
• Anthropic's Fable 5.1 and Mythos 5.1 cut costs up to 45%, arriving against reporting that its enterprise sales momentum has stalled. Same week, same direction as DeepSeek: cheaper per task, not merely smarter.
• Efficiency claims kept coming — a startup saying it matched DeepSeek V4 Pro's pretraining results with roughly 50x less compute, and a fused decode kernel claiming 1.58x over vLLM. Treat both as claims pending replication; the direction is consistent.
• Last week's AGI argument is still unresolved. GPT-6 Astra's launch claims met benchmark reporting that put it behind Anthropic and Meta on several suites, and the "AGI is here" and "AGI is a marketing term" quotes are both still standing.

On the blog this week

• GPT-6 Astra: OpenAI says "welcome to the AGI era" — what shipped, what the benchmarks say, and what the claim is worth.
• DeepSeek cuts Flash prices by up to 60% — why the cache-hit line is the one that matters.
• DeepSeek V4.1 Flash is live — the agentic gains, and the three evaluations where it still trails V4 Pro.

Our take

Three observations, in order of how much they should change your behaviour.

First, this week's competition was about cost per completed task, not headline capability. DeepSeek cut cache pricing 60%, Anthropic cut up to 45%, OpenAI cut image latency, and two startups claimed order-of-magnitude efficiency wins. Capability still sells the launch; economics decides what stays in production.

Second, the retirement of V4 Pro before V4.1 Pro exists is the most interesting decision of the week, because it is a claim with a deadline attached. Either Flash really is sufficient for pro-tier work — in which case the pro tier was mostly marketing — or DeepSeek is accepting a capability gap for a few weeks to move its installed base onto a cheaper, more cache-efficient model. Watch what happens to the knowledge-heavy evaluations when V4.1 Pro arrives; that is where the answer will be visible.

Third, and most practical: pin your versions and back up before upgrading DSH. Session format V3 is explicitly not downgrade-readable, plugin APIs were removed in the same release, and the profile defaults changed. A project that ships corrected behaviour four times in four days is doing the work — but it is still pre-release by its own description, and this week it told you so in the changelog.

Sources

• DeepSeek — V4.1 Flash release announcement
• github.com/deepseek-ai/deepseek-harness/releases
• IT之家 — Flash 系列新价格表
• AI Primer — benchmark table and third-party results
• 21世纪经济报道 — DeepSeek 启动科创板上市筹备
• 博客园 — 0.1.2 没有正式版：一个 rc 是怎么被放弃的
• 证券时报 — 英伟达超 129 亿美元收购 Hugging Face
• 界面新闻 — OpenAI 推出 GPT-Image 2.5
• 澎湃新闻 — Anthropic 上新 Fable 5.1