DeepSeek's September Beta: A Historical Test Window
MODELS & COSTS
DeepSeek's September Beta: A Historical Test Window
The endpoint name included its expiry date: deepseek-v4.1-flash-expires-on-0910. That short September window gave developers something to try before a full release, not a stable integration target. The reported 507 tok/s peak also belongs to uncontrolled beta observations rather than a matched speed test.
In this article
At the time, the beta used V4 Flash rates and a 20-concurrent-request account cap. Developer reports described around 420 output tokens per second on average and peaks above 500, under uncontrolled conditions. Those timings should be kept separate from an official, reproducible speed benchmark.
The September 8 temporary endpoint
Around 15:00 Beijing time, DeepSeek's team told its official community groups that an intermediate version — described in Chinese as 中间版本, an in-between checkpoint — was opening for limited testing inside the real API environment ahead of a final build. Access could not be simpler: keep your base URL and key, set the model name to deepseek-v4.1-flash-expires-on-0910, and you are in. No waitlist, no new console, no separate endpoint.
| Item | The beta envelope |
|---|---|
| Model ID | deepseek-v4.1-flash-expires-on-0910 |
| Window | September 8 → around September 10 (about two days) |
| Access | Any DeepSeek API key, same base URL, model name swap only |
| Concurrency | 20 requests/account (production V4 Flash allows 2,500) |
| Billing | Same rate card as V4 Flash, no beta surcharge |
| Paperwork | None — no model card, no benchmarks, no tech report yet |
That 20-concurrency cap is doing a lot of quiet talking. It is barely one percent of what production Flash allows, which tells you exactly how DeepSeek sees this build: functional validation and evaluation, not a lane for production traffic. And tipsters expect the official release post very soon — the two-day fuse reads like DeepSeek does not intend to keep anyone waiting long.
Developer-reported speed
The number doing the rounds on day one is roughly 420 output tokens per second in long generations, with peaks reported above 500 — the highest shared run hit 507 tok/s, and one developer clocked 328 tok/s while having the model generate an SVG animation of a pelican riding a bicycle. One widely shared test pushed more than 71,000 tokens through in under three minutes. For scale, Artificial Analysis measures the public V4 Flash 0731 endpoint at roughly 128 tok/s, so the early reports point to a raw-throughput jump on the order of three times or more.
The back-to-back task comparisons against V4-Flash-Vision-Exp are where the gap looks widest: testers reported roughly 6x faster end-to-end on SVG code generation, about 5.2x on a 49k-token long-context retrieval, around 5x on a large SQL generation-and-optimization task, 4.6x on an algorithmic problem, and 3.9x on an async refactor. Same-task end-to-end numbers fold in thinking time, first-token latency, and generation speed all at once, and they come from individual testers rather than a controlled suite — so I read the consistent direction across all five as the signal, not any single multiplier. Nobody should wire production to a model ID that is scheduled to stop answering, but as a two-day evaluation window, this is a generous one.
The architecture described for the beta
Here is the part that made me sit up. The previous refresh, Flash 0731, was explicitly a post-training update that kept the same architecture and scale as its preview — DeepSeek said so in the changelog. V4.1 Flash is described, in the company's own wording, as a brand-new model structure, and several outlets read that as a sign the model was re-pretrained rather than fine-tuned. That breaks the pattern, and it matters, because an architectural iteration can move capability ceilings in a way that post-training alone usually cannot.
Bundled with the restructure is the second claim: native multimodal support. This is a precise phrase with a precise contrast behind it. The Vision-Exp build from August 21 bolted a vision encoder onto the V4 Flash text base — DeepSeek's own materials described that pair as a text model plus an external vision path. V4.1 Flash is described as multimodal from the ground up, with text and image inputs handled by the base model itself. What testers have actually exercised so far is text-plus-images, and I would treat multimodal as confirmed only in that text-and-image sense until a model card lands. Some coverage mentions audio inputs processed in a unified manner too, but the modality spec has not been published, so audio stays in the unknown column for now rather than the confirmed one.
Rates, tokens and the bill
DeepSeek said the test bills at V4 Flash rates, and the Flash rate card since the August 16 repricing is peak/off-peak: $0.22 per million input tokens and $0.66 output off-peak, doubling at peak ($0.44/$1.32), with cache hits at $0.007/$0.014. Peak hours are 01:00–04:00 and 06:00–10:00 UTC on weekdays. Against Pro at $0.66/$1.98 off-peak, the Flash tier sits at a clean 3x discount on every line.
Now the subtlety several first-day testers noticed the expensive way: DeepSeek's “lower cost” claim is an efficiency claim, not a rate cut, because per-token prices did not move at all. A model generating tokens three times faster drains your balance three times faster per minute of wall-clock time — one five-minute session visibly cost more than the same five minutes on V4 Flash. Whether per-task cost actually falls depends on whether the model finishes the job with fewer tokens and fewer retries, which nobody can judge from two days of speed tests. Keep both ideas in your head at once: cheaper per token than Pro by 3x, but the meter spins faster.
The Pro-replacement survey question
The most strategic detail of the whole beta is hiding in the feedback form. Alongside questions about agent frameworks and real-world scenarios, DeepSeek asks testers directly whether this intermediate build can fully replace the online V4 Pro. That is a remarkable question to put in front of users, because Pro sits three price tiers above Flash — if a Flash-tier model approaches Pro-level capability at Flash-level prices, the migration math answers itself for a huge slice of users, and DeepSeek knows it.
Read together with the new-structure phrasing, the questionnaire hints that V4.1 Flash is the first visible step of something bigger than a point refresh — a Flash line repositioned to compress the gap to the flagship, the way DeepSeek has repeatedly used a cheaper, faster model to reset what a price tier buys.
That is my strategic read, not a confirmed fact — no benchmark exists yet that could confirm it. What would answer it is exactly what is missing: independent evals on the agentic and reasoning suites that separate the Flash tier from the Pro tier today. Until those land, the honest position is that a two-sentence survey question is suggestive, not evidentiary. But it is easily the most interesting thing DeepSeek said all week.
The preceding release sequence
Zoom out and the cadence is striking: Flash 0731 public beta on July 31, Pro GA plus Harness v0.1 on August 13, Vision-Exp live on August 21, open weights on August 31 — four to five major moves in under 40 days. The V4.1 Flash beta looks like the opening salvo of a new cycle, not a one-off experiment.
The intermediate label and short window set a clear scope for the beta. The unanswered question was performance per dollar against Pro under matched tasks. The feedback form asked about that replacement, but the survey itself could not establish it; a model card, release rates and independent tests were still needed.
Sources: OrcaRouter, “DeepSeek V4.1 Flash Hits a Two-Day API Beta” (Sep 8, 2026, community measurements + beta details); DeepSeek API Docs changelog (Flash 0731 / Pro GA benchmark tables, vendor-reported); BigGo Finance beta coverage (Sep 8, 2026). All benchmark and speed figures are vendor-reported or community-measured as labeled — no independent verification exists yet for the V4.1 Flash build.
Comments
Post a Comment