The migration most developers never opted into
At 04:00 UTC on September 14, 2026, every request pointed at deepseek-v4-pro began being answered by DeepSeek-V4.1-Flash and billed at Flash rates. There is no migration project, no SDK bump, and no deprecation window to plan around. DeepSeek said the new release had comprehensively surpassed V4 Pro in performance, cost, speed, and total time, so the flagship tier simply stops serving its own traffic until a V4.1 Pro arrives. The new model identifier is deepseek-flash. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoints are retired, with their legacy names temporarily routed to the new checkpoint during a transition window.
That is an aggressive move for a model that is only five days old. It is also the clearest signal yet about where price competition in frontier inference has actually moved: not to output tokens, but to the cost of re-reading context your agent has already seen.
DeepSeek-V4.1-Flash landed on September 10, 2026, roughly a week after OpenAI's GPT-6 Astra, Google's Gemini 3.8 Flash and its Fairwind-gated Cyber variant, Anthropic's Claude Fable 5.1 and Mythos 5.1, and Meta's Muse Spark 1.3. Five frontier-class releases in ten days is the new normal, and the differentiator this month was not raw capability. It was the shape of the bill.
What V4.1-Flash actually is
The release is a mixture-of-experts model with a 552-billion-parameter backbone, up roughly 94 percent from the previous Flash tier, plus 196 billion conditional memory parameters that DeepSeek calls Engram. The company says only about 8 billion parameters activate during prefill and 16 billion during decode, which is what keeps serving costs low despite the much larger backbone. Native vision is included, the context window is one million tokens, and the weights ship under an MIT license across 48 shards that total roughly 475 to 510 gigabytes depending on how the repository is counted.
The architectural headline is a combination DeepSeek describes as an encoder-decoder split with cross-layer attention reuse, FP4 expert weights, and sliding-window attention with bounded replay. The company claims this cuts KV cache memory requirements to about a quarter of what a comparable dense design needs and SSD storage to about an eighth. Fewer bytes of working memory per token means more concurrent sessions per GPU, and that efficiency is precisely what the new rate card is built on.
The number everyone keeps repeating
Off-peak, a cache-hit input token costs $0.003 per million. A cache miss costs $0.075 to $0.15 per million depending on the source page, and output runs $0.30 to $0.60 per million. Peak hours, defined as 01:00 to 04:00 UTC and 06:00 to 10:00 UTC on weekdays, double those figures. For scale, Anthropic's Claude Opus 5 lists at $5 per million input tokens and $25 per million output tokens, making the off-peak output rate more than 80 times cheaper on that axis alone.
Two details matter more than the headline price. First, the gap between a cache hit and a cache miss is roughly fiftyfold on the same model, which means the way your requests reuse context now moves your bill more than the rate card does. Second, DeepSeek raised the concurrency ceiling from 500 on V4 Pro to 2,500 on Flash, so the cheaper tier is also the one you can push harder.
Why cache reads dominate agent bills
An agent loop reads far more than it writes. System instructions, tool definitions, retrieved documents, prior turns, and file contents are re-sent on every step, while only the newest tool result is genuinely new. That asymmetry is why the useful benchmark for agent builders is completed work per cached token, per second, and per dollar, rather than raw intelligence per token. A model that is cheaper per token but needs an extra retry or a longer reasoning trace can easily cost more in production than a pricier one that finishes on the first pass.
Where the benchmarks hold and where they do not
At its maximum effort setting, DeepSeek reports 90.6 on Terminal-Bench 2.1 against 87.9 for V4 Pro, 74.2 on DeepSWE v1.1 against 62.7, 31.2 on Terminal-Bench 4.0 against 12.4, 31.8 on Agents' Last Exam, 88.1 on CyberGym, and 54.8 on AutomationBench. Its Codeforces rating reaches 3,471, above V4 Pro's 3,348. On OpenDesign's public design arena, one comparison put V4.1-Flash at 98 percent of GPT-6 Astra's average score at roughly 1.4 percent of the cost, finishing a task for about two cents where Astra cost over a dollar and a half.
The counterweight is just as real. On Terminal-Bench 3.0 it scored 30.0 against Opus 5's 43.3. On Humanity's Last Exam it reported 36.8 against 56.3 for Opus 5. On GPQA Diamond it reached 90.9, behind GPT-5.6 Sol's 94.1 and Opus 5's 93.4. In other words, this is a price-performance release, not a new capability ceiling, and the gap on hard reasoning and long-horizon reliability is still measured in double-digit points.
The most useful disclosure in the technical report is quieter. The same V4.1-Flash checkpoint scored between 65.5 and 74.2 percent on DeepSWE v1.1 depending solely on which agent harness wrapped it. An 8.7-point spread came from the surrounding evaluation scaffold rather than the weights. Read that as a caution about every vendor benchmark table you see this quarter, and as a strong argument for piloting on your own workflow instead of adopting someone else's leaderboard.
Who should switch this week
- High-volume, cache-heavy agent steps. Classification, extraction, retrieval summarization, tool-result formatting, and long-context reading are exactly where a fiftyfold cache discount compounds across thousands of daily calls.
- Batch and overnight pipelines. Off-peak pricing is half price, and any schedulable workload gets that discount for free.
- Teams experimenting with self-hosting. The MIT license and FP4 expert weights make local deployment viable, but the 196B Engram tables are CUDA-bound in the reference implementation and do not yet offload cleanly to host RAM, so budget VRAM realistically before you buy hardware.
Who should wait
- Regulated or sensitive data. The cheapest inference in the world does not help if your data is not allowed to leave your jurisdiction or your own tenancy.
- Hard reasoning and long-horizon agent reliability. The HLE and Terminal-Bench 3.0 gaps are large enough to matter where a wrong answer is expensive.
- Anyone pinning model snapshots for reproducibility. Traffic routing changed underneath you today. If your product depends on identical behavior, capture a before-and-after output set before you ship anything new.
A ten-minute checklist for the switch
- Log your cache-hit ratio per agent step, not per request. A single blended number will hide the steps where you are paying full price.
- Reorder prompts so static content comes first: system instructions, tool schemas, few-shot examples, then variable content such as the user message and retrieved chunks.
- Pin the model identifier explicitly instead of relying on a legacy alias that now routes elsewhere.
- Record cost per completed task for your five most frequent agent actions, then re-measure after a week on Flash.
- Decide in advance which steps stay on a frontier model, so routing decisions are a config change rather than an incident.
What this means beyond one release
DeepSeek's migration is a preview of the next twelve months of model operations. Tiers will be swapped underneath you, cache economics will decide your gross margin, and the harness you wrap around a model will matter as much as the model itself. The teams that come out ahead are the ones that instrument cost per completed task, design prompts for cache reuse, and treat model routing as a first-class part of the architecture rather than a constant they set once and forgot. The cheapest model is only the cheapest model when your context layout lets it be.
