Set model to deepseek-flash — that is the current deepseek flash api model name following the DeepSeek V4.1 Flash release on 10 September 2026, per the official DeepSeek API changelog. The two legacy names, deepseek-v4-flash and deepseek-v4-flash-vision-exp, have not been switched off: DeepSeek's documentation confirms they are still accepted and temporarily route to the new V4.1 Flash model at Flash pricing. So existing code keeps working, but deepseek-flash is the name to standardise on going forward, and this page covers exactly what that string now points at, what it costs, and what happens to the older models.
📺 Watch: How to use DeepSeek V4.1 Flash API for FREE!
🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →
Every fact here is drawn from DeepSeek's own API changelog and its models and pricing documentation as fetched on 13 September 2026 — no third-hand summaries, and where a benchmark number appears it is DeepSeek's vendor-reported figure, flagged as such.
The DeepSeek Flash API Model Name, Explained
DeepSeek's naming has been a moving target this year, which is precisely why the deepseek flash api model name question keeps coming up. The timeline per the official changelog runs like this: the V4 generation shipped with a Flash tier addressed as deepseek-v4-flash, an experimental vision variant appeared as deepseek-v4-flash-vision-exp, and on 10 September 2026 the company released DeepSeek V4.1 Flash — the smallest model in a new architecture family — under the simplified string deepseek-flash.
Three practical consequences follow. First, new integrations should call deepseek-flash, full stop. Second, code still calling deepseek-v4-flash or deepseek-v4-flash-vision-exp now gets V4.1 Flash answers whether it expects them or not, because those names route to the new model. The word "temporarily" in the changelog matters: DeepSeek has not committed to keeping the aliases forever, so treat them as a migration bridge rather than a permanent address. Third, because the vision-exp variant folded into the main model, vision is no longer a separate endpoint — the background on that experiment is in the DeepSeek V4 Flash vision-exp write-up, and its capability now ships as standard.
If you want to turn cheap frontier-class API access like this into actual income — with over 1,000 done-for-you AI workflows, daily tutorials and weekly live coaching calls — check out the AI Profit Boardroom → get inside here. Want a personal plan first? Book a free SEO strategy session and Julian will map it with you 1-on-1.
What deepseek-flash Points At Now
Per DeepSeek's models and pricing page, the deepseek-flash string currently resolves to DeepSeek V4.1 Flash with the following specification: a 1 million token context window, up to 384K output tokens, and support for vision input, JSON output, tool calls and thinking mode. The native multimodal visual understanding is the headline change — this is the smallest model in DeepSeek's new architecture family, which the company says is designed for a higher capability ceiling, faster inference and higher throughput.
On benchmarks, DeepSeek reports GPQA Diamond at 90.9 and a Codeforces rating of 3471 for V4.1 Flash. Treat both as vendor-reported until independent replication lands; the V4.1 Flash versus V4 Pro benchmarks page tracks how those claims stack up against the rest of the family, and the Goldie Bench write-up covers how these model brains compare in hands-on agent tests more broadly.
For release-day context — what shipped, when, and what DeepSeek said about the architecture — the DeepSeek V4.1 Flash release date page has the fuller story. This page stays focused on the API surface, because that is what breaks builds.
📺 Watch: NEW DeepSeek V4.1-Flash Is a GAME CHANGER!
Current Pricing for deepseek-flash
DeepSeek cut API prices alongside the release, and the current published rates per 1 million tokens are worth spelling out, because they are startlingly low. All figures below are from DeepSeek's pricing page as of 13 September 2026.
| Rate (per 1M tokens) | deepseek-flash | deepseek-v4-pro |
|---|---|---|
| Input, cache miss — peak | $0.30 | $1.32 |
| Input, cache miss — off-peak | $0.15 | $0.66 |
| Input, cache hit — peak | $0.006 | $0.044 |
| Input, cache hit — off-peak | $0.003 | $0.022 |
| Output — peak | $1.20 | $3.96 |
| Output — off-peak | $0.60 | $1.98 |
Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays, with off-peak rates at half the peak price. Cached input at three thousandths of a dollar per million tokens off-peak is the number that changes behaviour: agent workloads that hammer the same system prompt repeatedly become close to free on the input side. The DeepSeek V4 pricing update page covers how the earlier price structure evolved into this one if you want the trail.
What Happens to DeepSeek V4 Pro
The changelog carries one more decision that matters for anyone running production workloads: DeepSeek confirmed it will continue providing API services for DeepSeek V4 Pro after 14 September 2026, with the billing method unchanged. Earlier guidance had pointed at a cutoff around that date, so this is a reprieve — V4 Pro remains addressable as deepseek-v4-pro with its 1 million token context and 384K output ceiling, though notably without vision support, which is now a point in Flash's favour rather than Pro's.
The sensible reading of the two announcements together: DeepSeek wants everyone on the new architecture family, Flash is the first member shipped, and Pro's stay of execution exists so nobody's production pipeline breaks while the larger new-architecture models arrive. If you are choosing between them today, Flash is roughly a quarter of Pro's price, adds vision, and posts vendor benchmarks near or above it — which is why the routing of legacy Flash names to V4.1 was uncontroversial while Pro kept its own name.
📺 Watch: DeepSeek V4.1 Flash + Hermes Agent (FREE!)
Using the New Model Name in Agent Stacks
Swapping the model string is a one-line change in most setups, but the place it pays off fastest is agentic use. DeepSeek models slot into coding agents and automation harnesses where the token bill is the whole game, and a Flash-class model with cached input this cheap changes which tasks are economical to automate at all. The DeepSeek harness guide covers the agent-side setup, and the DeepSeek V4 tutorial walks through the API basics that still apply unchanged to V4.1 — authentication, endpoints and request shape did not move, only the model name did.
Within an Agent OS style setup, the model name lives in one config layer precisely so releases like this stay boring: update the string, run your checks, carry on. Teams that hardcoded deepseek-v4-flash across a dozen scripts are this week discovering why that layer exists.
One caution worth repeating from the changelog language: aliases described as temporary do eventually disappear. The cheap insurance is to grep your codebase for both legacy strings today, switch them to deepseek-flash, and note the change — five minutes now against a mystery outage later, on whatever day DeepSeek decides the bridge has served its purpose.
Common Mistakes When Switching the Model String
A few failure modes are predictable enough to name. The first is assuming deepseek-flash behaves identically to the old deepseek-v4-flash because the alias still answers: it does not — V4.1 Flash is a different architecture, so outputs, latencies and token habits can shift even though the request shape is unchanged, and any prompt tuned hard against V4 Flash deserves a re-check. The second is forgetting the vision-exp consumers: code that called deepseek-v4-flash-vision-exp for image work now hits the mainline model, which is an upgrade, but rate assumptions and output formats built around an experimental endpoint should be revalidated. The third is scheduling batch jobs blind to the clock — with peak windows at 01:00 to 04:00 and 06:00 to 10:00 UTC on weekdays and everything else billed at half price, simply moving a nightly pipeline a few hours can halve its API bill. None of these are exotic; all of them are cheaper to catch this week than in an invoice next month.
Bottom Line on the Model Name
Call deepseek-flash. It resolves to DeepSeek V4.1 Flash, released 10 September 2026, with a 1M context window, 384K max output, native vision, JSON output, tool calls and thinking mode, at $0.30 per million input tokens peak (cache miss) and $1.20 output — half that off-peak. The old names deepseek-v4-flash and deepseek-v4-flash-vision-exp still work but only as temporary routes to the same model, and deepseek-v4-pro continues past 14 September 2026 with billing unchanged. All of it per DeepSeek's official changelog and pricing documentation, and all of it liable to move again — this is DeepSeek, after all, and the next architecture-family sibling is presumably already in the queue.
If you want the exact agent workflows that squeeze value out of releases like this the week they drop, check out the AI Profit Boardroom → join 3,000+ operators inside. And if you would rather talk it through first, book a free SEO strategy session — bring your project, leave with a plan.
