DeepSeek V4.1 Flash 1017: Your Model Identifier Moved and Nothing Told You

The identifier code>deepseek-flash/code> does not point at the model it pointed at in September. DeepSeek's own documentation says the name is now served by DeepSeek-V4.1-Flash-1017, the October checkpoint — and that two legacy identifiers, code>deepseek-v4-flash/code> and code>deepseek-v4-flash-vision-exp/code>, are still accepted, but that the models behind them have been retired, that requests are served by the 1017 checkpoint, and that they are billed at the Flash price.

Nothing about the price changed. That is precisely why this is easy to miss: a bill that looks identical is not evidence that the model is identical, and a baseline collected in September is now describing a checkpoint nobody can call any more.

What "1017" means, and what it does not

DeepSeek dates checkpoints by month and day, and 1017 is 17 October. The V4.1 line's architecture is unchanged — a 552-billion-parameter Causal Encoder-Decoder mixture-of-experts activating 8 billion parameters during prefill and 16 billion during decode — so a checkpoint update here is a post-training move rather than a new model. That is the factor that most affects how much your evaluations should shift: same shape, same context, different weights.

What is unchanged on the card: 1M-token context, 384K maximum output, 2,500 concurrent requests, vision supported, and $0.15 per million input tokens against $0.60 output off-peak, with a cache hit at $0.003. What is not published anywhere we can cite: which evaluations the checkpoint improved and by how much, in a form a reader can verify. Anything you read about "the 1017 jump" is either a vendor statement or a third party's reading of it.

The three identifiers, and what each buys

• deepseek-flash — the current name, served by V4.1-Flash-1017.

• deepseek-v4-flash — accepted, retired model, served by 1017, billed at the Flash price.

• deepseek-v4-flash-vision-exp —/strong> the same three facts. This was the experimental vision model that landed in August and was retired inside three weeks, absorbed into the V4.1 release.

All three are one endpoint call apart and now resolve to the same checkpoint. The practical consequence is that an integration written against either legacy name keeps working, keeps being billed at the same rate, and starts returning a different model's outputs without any error or notice. If you have attributed a change in quality to prompt drift or traffic mix since mid-October, this is a competing explanation that costs nothing to rule out.

How to tell which checkpoint answered

Three checks, none of which requires a vendor statement.

Read the models table. DeepSeek's Models & Pricing page lists a MODEL VERSION row per column: today it reads DeepSeek-V4.1-Flash-1017 for deepseek-flash and DeepSeek-V4-Pro-0813 for deepseek-v4-pro. That row is the authoritative statement about what the identifier resolves to, and it changes when the checkpoint changes.

Keep a frozen evaluation set. A twenty-prompt suite run monthly against the same identifier is the only way to notice a silent swap in time to matter. Log the date next to the score, the way you would log a model revision.

Log the usage response. Per-request metadata rarely names the checkpoint, but the bill does tell you the rate — and a rate that did not change alongside a score that did is the signature of exactly this situation.

What this means for a cost model

Nothing, and that is the trap. The card is identical: $0.15 and $0.60, $0.003 cached, doubling in the two weekday peak windows. So a spreadsheet built in September still reconciles, while the thing it is pricing has changed underneath it. Cost models survive identifier swaps; quality models do not.

Two habits are worth adopting. Pin the evaluation, not the name: the name is DeepSeek's to move, and it has moved once already this month. And keep the cache-hit share in the model, because on this card $0.003 against $0.15 is a 50× line that dominates everything except a change of model class.

What to call today

DeepSeek V4.1 Flash — under any of the three identifiers — is on OrcaRouter at DeepSeek's published rates, passed through with no markup, on the same OpenAI-compatible endpoint as a 224-model catalogue, with automatic failover across provider paths. DeepSeek V4 Pro sits beside it at $0.66 and $1.98 off-peak for text-only flagship work. Both are billed exactly as DeepSeek defines them, including the peak doubling — which cuts both ways: the discount is live the day it is published, and so is the surcharge.

What to watch: the MODEL VERSION row on DeepSeek's page. It is one line, it is the only place the checkpoint is named, and it has already changed once.