DeepSeek Ships V4.1-Flash, Then Retires the Model It Replaced
On September 10 at 04:00 UTC, DeepSeek pushed DeepSeek-V4.1-Flash to general availability on its API and dropped a 50-page technical report alongside MIT-licensed open weights — 475 GiB spread across 48 shards. Four days later, on September 14, the company starts routing every request that names deepseek-v4-pro to the new model and billing it at Flash rates. There is no deprecation window, no maintenance page, no migration guide asking politely. If your code says "Pro," it will quietly start running on Flash.
That is an unusual move for any lab, and an even stranger one for DeepSeek, which spent the first half of 2026 raising prices on V4 Pro and telling the market that frontier capability costs money. By September the story had inverted: DeepSeek says V4.1-Flash "has comprehensively surpassed V4 Pro in performance, cost, speed, and total time." When a vendor tells you the cheaper tier beats the expensive tier on all four axes at once, the honest reading is that the model family changed underneath the naming scheme. V4.1 is a new lineage, not a point release.
What Actually Changed in the Architecture
The headline number developers keep repeating is 890 bytes per token of KV cache — roughly a quarter of V4-Flash's footprint. DeepSeek got there with a causal encoder-decoder split and FP4 quantization of the cache, a combination that matters far more to your invoice than to any leaderboard. Long-running agents re-read a huge prefix on every turn. If the cache is cheap to keep resident, the cache-hit price can be absurdly low without burning the provider's margin, and that is exactly how the rate card reads: $0.006 per million cache-hit input tokens at peak, $0.003 off-peak, against $0.30 and $0.15 for cache misses.
The model itself is a 552B-parameter mixture of experts activating roughly 8B parameters on input and 16B on output. The technical report also describes Engram, a 196B-parameter conditional memory component sitting outside the backbone. It is not a retrieval index and it is not a scratchpad you can inspect; it is a trained component whose job is to hold conditional state, and it is the most interesting architectural idea in the release.
The Benchmark Table, Read Carefully
On agentic evaluations the new model posts numbers that would have been a frontier result a year ago. Terminal-Bench 2.1 lands at 90.6 against V4 Pro's 87.9 and Opus 5's 89.1. DeepSWE v1.1 reaches 74.2 against V4 Pro's 62.7. AutomationBench hits 54.8 versus 43.2. DeepSeek's Codeforces rating climbs to 3,471, a modest bump over Pro's 3,348.
- Where it wins: terminal work, software engineering tasks, security-heavy CyberGym-style evaluations (88.1 versus 84.5 for the best competitor in the group).
- Where it loses: Terminal-Bench 3.0 at 30.0 against Opus 5's 43.3, and GPQA Diamond at 90.9 behind GPT-5.6 Sol's 94.1.
- Where it is honest about limits: Humanity's Last Exam, 36.8 overall and 39.1 on the starred configuration. Hard expert reasoning is still a closed-model game.
The most useful line in the technical report is not a benchmark at all. The same V4.1-Flash checkpoint scored 65.5% to 74.2% on DeepSWE v1.1 depending only on which harness wrapped it — Claude Code, Codex, OpenCode, or DeepSeek's own scaffold. An 8.7-point swing produced by the loop around the model, not the model. Anyone who has shipped an agent already knows this in their bones, but seeing the same weights swing that far across harnesses should end the era of treating a single benchmark row as a purchasing decision.
Price-Per-Token Versus Cost-Per-Task
Peak pricing is $0.30 per million input tokens and $1.20 per million output tokens, halved off-peak to $0.15 and $0.60. Against GPT-6 Astra and Claude Fable 5.1 — both sitting at $10 input and $50 output per million — that is a thirtyfold gap on input and a fortyfold gap on output. Concurrency is 2,500 requests for Flash against 500 for Pro, which matters more than the rate card when you are fanning out subagents.
But the number that spread fastest was cost per completed task. On OpenDesign's public design arena, V4.1-Flash reached 98% of GPT-6 Astra's average score at about 1.4% of the cost, finishing a task in 5.3 minutes for roughly $0.023 where Astra took 11.1 minutes and $1.61. That is one third-party arena on one task family, not a universal ranking, and the same release notes that celebrate it also document the failures: DeepSeek's own prior generation landed three to four points below vendor-reported Terminal-Bench figures when independent analysts ran it.
The sane response is not to move your whole stack on a Tuesday. It is to close the gap between what you believe your agent costs and what it actually costs. Measure cost per completed task, not price per million tokens. Route the boring, high-volume steps — file reads, log triage, boilerplate refactors, classification — to a model in this price band, and keep the last mile of hard reasoning on a frontier model until your own evals say otherwise.
Why September 2026 Feels Like a Reset
This release did not land in quiet. Anthropic shipped Claude Fable 5.1, cut cache-read pricing by 75%, and introduced the trusted-access Mythos 5.1 twin on September 1. Google followed with Gemini 3.8 Flash and a defenders-only Cyber variant the next day, both on introductory pricing that doubles on January 1, 2027. Meta's Muse Spark 1.3 arrived the same week with a contributor tier priced at $0.10 and $0.20 per million. OpenAI's GPT-6 Astra — the first model to trip the company's critical-cyber safeguard threshold — went out on September 3.
Five frontier-ish releases in ten days, and not one of them was a price cut on the expensive end. The cheaper direction came from the open-weight side, and it came with a catch that the marketing tends to skip: the economics only hold if your workload reuses context. A stateless application that pays cache-miss prices on every call gets a good deal; an agent that rushes its prompts and never reuses a stable prefix gets a much less impressive one. In that sense DeepSeek's release is less a gift than an audit — it tells you, in dollars, how well you have engineered your context.
OpenAI DevDay lands on September 29 in San Francisco, where a wider Astra rollout and a full evaluation suite are expected. If past cycles hold, the announcement that matters most will not be a new model name. It will be the tooling around it — harnesses, memory layers, evaluation harnesses, and pricing mechanics — because that is where the 8.7-point harness gap actually lives. Model weights are becoming a commodity. The scaffolding you build around them is not.
