DeepSeek has shipped a lot this year, and the updates come fast enough that it’s easy to lose track of what’s current. This DeepSeek news roundup sorts it out: what’s live today, what changed, what’s still rumor, and how to tell the difference.
The information below is based upon what was reported at the early stage of November 26th, 2026.
This is the newest version: V4.1-Flash-1017
The newest model is DeepSeek-V4.1-Flash-1017. It was released by DeepSeek on a Saturday as an update to V4.1-Flash, which they claim can complete tasks faster, and produce more concise answers. October 17 happened to be a Saturday.
Some practical information:
Vendor benchmarks tend to flatter the vendor. Independent testing is a better yardstick, and it usually arrives a few days after launch.
- The API name and pricing didn’t change, and developers reached the new model through the same deepseek-flash name.
- The weights were published on Hugging Face under the MIT license.
- DeepSeek’s own model card calls it a “minor update” that cuts task time and output length, without saying by how much.
If you already use deepseek-flash, you’re on the new version without doing anything.
What the numbers say (and don’t)
DeepSeek published benchmark results with the release. Its change log lists 42.6 on Humanity’s Last Exam without tools, up from 36.8 for the September model, and 42.3 on Terminal-Bench 4.0, up from 31.2. On the DeepSWE software engineering test, the company reported 77% of tasks solved, compared with 74.2% before.
Keep one thing in mind: these are DeepSeek’s own figures. Vendor benchmarks have a tendency to make the vendor look good. An independent test is a much better measure and it typically comes a few days after launch. For example, when the V4 flagship went generally available, an independent leaderboard ranked it third with a score of 53, only a few points above the much smaller Flash model and short of what the model card implied. That’s a useful reminder to read a launch table as a claim, not a verdict.
2026 release timeline
Here’s the short version of how the year unfolded:
| Date | Release | Notes |
| April 24 | V4-Pro and V4-Flash (preview) | First V4 models |
| July 31 | V4-Flash-0731 | Official release, open weights |
| August 12–13 | V4-Pro-0813 | Flagship goes generally available |
| September 10 | V4.1-Flash | Smaller model, image understanding |
| October 17 | V4.1-Flash-1017 | Faster, shorter outputs |
Dates come from third-party trackers and DeepSeek’s own posts, so double-check anything critical against the official changelog.
Pricing and access changes
Two pricing stories are worth knowing.
First, DeepSeek raised API prices on August 17 and split them into peak and off-peak rates. If you run batch jobs, scheduling them off-peak could lower your bill.
Second, the September V4.1-Flash release came with lower API prices, and the model was described as the smallest in a new architecture family with built-in image understanding. That cut didn’t go unnoticed. The release was reported to have pushed down share prices of several rival companies.
Pricing changes often, so check the official pricing page before you budget.
What’s next: V4.1-Pro, and a V5 that isn’t confirmed
This is where rumor and fact mix, so I’ll separate them.
Confirmed: The only future model DeepSeek has named publicly is V4.1-Pro. Its news page and API changelog have no V5 entry.
Reported, not confirmed: One outlet reported on November 3, citing industry sources, that V5 training is still being planned and that a release is likely in 2027. Yet another report indicates that DeepSeek is training a 2 trillion-parameter model.
Use them as indicators and not timetables. Plans are evolving and anonymous sources are the least compelling proof.
You can evaluate any DeepSeek claim with the News Filter. Any DeepSeek claim can be evaluated by using the News Filter.
AI model headlines outpace reality. When you are sharing or using a story, please verify the following 4 items.
Is it found on the official page?
DeepSeek’s news page, API changelog or Hugging Face account should list a real release. If it doesn’t, it’s not out.
- Is there a model name or ID? Real releases have a version string you can call. Vague “coming soon” claims don’t.
- Who’s the source? Official posts and named reporters beat anonymous “industry sources.”
- Is the benchmark independent? Consider scores from subjective tests as marketing until verified by a third party.
A story that passes all four is solid. One that fails the first two is rumor, however confidently it’s written. This is a handy habit whenever you’re following DeepSeek news or any fast-moving AI story.
What this means for different readers
- Developers: Keep using deepseek-flash. Try the new version with some of your prompts as the new version might alter the output style because it is “faster and shorter”.
- Business teams: Keep an eye on fluctuations in prices and the peak/off-peak split when setting a budget.
- Researchers: The weights are designed under the MIT license, meaning they can be inspected and tested locally.Open-source fans: The weights are available for inspection and local testing under the MIT license.
- Regular users: Either the app or the website currently utilizes the update, so you don’t need to install it.
The takeaway
So, the situation is straightforward. V5 is still only rumours, V4.1-Pro is the next named release, and pricing is now more complex than ever with the latest live model, V4.1-Flash-1017. If you follow DeepSeek news, you should always refer to their official changelog, consider vendor benchmarks with a pinch of salt, and wait for independent benchmarks to draw your conclusions.
