DeepSeek V4 Pro: What Changes? — routed to V4.1 Flash from 9/14
On 2026-09-10 DeepSeek released DeepSeek-V4.1-Flash and stated that it plans to retire V4 Pro in an orderly manner. This page is for teams still running on V4 Pro or about to adopt it: it lays out what changes and what does not, using only what DeepSeek's official documentation actually says. Pricing, specs, benchmarks and setup for the successor model live in the DeepSeek V4.1 Flash guide.
⚠️ Bottom line (3 lines)
- From 13:00 JST on 2026-09-14 (12:00 Beijing time / 04:00 UTC), and until V4.1 Pro is released, all requests to
deepseek-v4-proare routed to V4.1 Flash and billed at the V4.1 Flash price. - You do not have to change your API key or code. Requests keep being accepted, but the model actually answering becomes V4.1 Flash.
- DeepSeek has not announced a V4.1 Pro release date, so there is no guarantee you can return to V4 Pro's pricing or behaviour.
→ For the successor model, see the DeepSeek V4.1 Flash guide (pricing, benchmarks, setup, FAQ).
How we got here (based on official announcements)
- 2026-04-24 — DeepSeek V4 (V4-Pro / V4-Flash) becomes available on the API; the legacy
deepseek-chat/deepseek-reasoneraliases enter a migration window. - 2026-07-24 — Legacy aliases
deepseek-chat/deepseek-reasonerare discontinued. - 2026-07-31 — V4-Flash reaches general availability as a public beta.
- 2026-08-13 — V4-Pro (-0813 / GA) ships with substantially improved agentic performance, and the move to peak/off-peak pricing is announced.
- 2026-08-16 16:00 UTC — All V4 models move to two-tier peak/off-peak pricing (off-peak is half the peak rate).
- 2026-08-21 — The experimental
deepseek-v4-flash-vision-expbecomes available. - 2026-09-10 — DeepSeek-V4.1-Flash is released. V4 Flash and V4 Flash Vision Exp are retired and consolidated into V4.1 Flash, and DeepSeek states V4 Pro will be retired in an orderly manner.
- 2026-09-14 13:00 JST — Routing of
deepseek-v4-prorequests to V4.1 Flash begins (scheduled).
Source: DeepSeek Change Log. Release details are also collected in the release timeline.
What actually changes from 9/14
| Item | Changes? | Detail |
|---|---|---|
| API key / endpoints | No change | Base URLs stay https://api.deepseek.com (OpenAI format) and /anthropic (Anthropic format) |
Model name deepseek-v4-pro | Still accepted | Requests keep working, but they are served by V4.1 Flash |
| Billing rate | Changes (lower) | V4 Pro rates → V4.1 Flash rates; the gap widens further with cache hits |
| Context / max output | Effectively same | Both 1M-token context and 384K-token maximum output |
| Image input (Vision) | Newly available | V4 Pro had no vision support; V4.1 Flash supports it natively |
| Response quality / style | May differ | The served model changes, so re-check long or specialised prompts before relying on them in production |
| Concurrency limit | Increases | V4 Pro 500 → V4.1 Flash 2500 |
| Thinking mode controls | Check docs | V4.1 Flash supports both non-thinking and thinking (thinking is the default). Check DeepSeek's Thinking Mode page for any reasoning-effort parameter |
V4 Pro reference rates (valid through 2026-09-13)
| Item | Off-peak | Peak |
|---|---|---|
| Input (cache miss) | $0.66 | $1.32 |
| Input (cache hit) | $0.022 | $0.044 |
| Output | $1.98 | $3.96 |
Effective 2026-08-16 16:00 UTC. Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays (JST 10:00-13:00 / 15:00-19:00); weekends are off-peak all day. The old promo rates ($0.435/$0.87) are retired. For post-migration rates see the V4.1 Flash pricing table. Source: api-docs.deepseek.com.
Migration checklist
- Grep your code for the model name. Find every place that sets
deepseek-v4-pro. It keeps working, but new code is clearer withdeepseek-flash. - Look for retired names.
deepseek-v4-flash,deepseek-v4-flash-vision-exp,deepseek-chatanddeepseek-reasonerare retired or discontinued. - Re-check production responses once. Because the served model changes, compare outputs on a few representative prompts.
- Update your cost estimate. The token calculator lists both off-peak and peak rows for V4.1 Flash.
- Shift batch work off-peak. Peak is 01:00-04:00 and 06:00-10:00 UTC on weekdays (double the off-peak rate); weekends are off-peak all day.
Setup in VS Code GitHub Copilot Chat (3 steps)
The record of how the operator (okamo channel) actually set this up. Extension labels and the model picker may change over time.
⚠️ This figure records the setup as of 2026-07 and shows deepseek-v4-pro as the configured model. For new setups, choose deepseek-flash where the picker offers it (extension labels may change over time).
Get an API key
Sign up at platform.deepseek.com. Top up with PayPal or a credit card (VISA/Mastercard), then create an API key beginning with sk-. It is fully prepaid, with a $2 minimum top-up and no auto-charge.
Install the extension
Install DeepSeek V4 for Copilot Chat (publisher: Vizards) from the VS Code Marketplace.
Configure and start
Run DeepSeek: Set API Key from the command palette (Cmd/Ctrl+Shift+P) to register your key, then pick a DeepSeek V4-family model in the Copilot Chat model picker.
Source: okamo channel operator's setup record (2026-07-04). For current pricing and specs see the V4.1 Flash guide.
Reference: V4 Pro vs Claude Sonnet in practice (historical note)
In the operator's hands-on experience at the time, V4 Pro delivered near-Claude-Sonnet practical quality on text-centric coding tasks at a much lower cost. Note that the comparison was not like-for-like: the Sonnet side billed for two roles (summariser + critical engineer) while the DeepSeek side billed for one role (critical engineer).
An earlier version of this page cited a measured cost example of "Claude Sonnet $5.33 vs DeepSeek V4 Pro $0.26 (about 95% cheaper)". That figure was measured under the old promo pricing ($0.435/$0.87) and does not carry over to current V4.1 Flash rates (output $0.60 off-peak), so we removed it. Use the token calculator and the pricing comparison for today's numbers.
Original reports: savings-mode comparison report / Claude vs DeepSeek code-review comparison report
FAQ
Q. What happens to V4 Pro after September 14?
A. From 13:00 JST on 2026-09-14 (12:00 Beijing time / 04:00 UTC), and until V4.1 Pro is released, all requests to deepseek-v4-pro are routed to V4.1 Flash and billed at the V4.1 Flash price. You do not have to change your API key or code, but the model answering becomes V4.1 Flash.
Q. Do I need to change my extension or app settings?
A. DeepSeek states that requests specifying deepseek-v4-pro are routed to V4.1 Flash — i.e. they keep being accepted. So no rewrite is strictly required for requests to keep working. For new code, specifying deepseek-flash makes the intent explicit.
Q. Is V4 Pro pricing still valid?
A. V4 Pro's rates (cache-miss input $0.66 off-peak / $1.32 peak, output $1.98 off-peak / $3.96 peak) are reference values valid only through 2026-09-13. From 2026-09-14 13:00 JST you are billed at V4.1 Flash rates: cache-miss input $0.15 off-peak / $0.30 peak, output $0.60 off-peak / $1.20 peak.
Q. When will V4.1 Pro be released?
A. No release date has been announced as of 2026-09-11. The official docs only say the routing lasts "until the future release of V4.1 Pro", with no committed date.
Q. V4 Pro was text-only — can I use images after migrating?
A. Yes. V4.1 Flash supports native multimodal visual understanding and the official feature table marks Vision as supported (V4 Pro did not). For image, audio, video and PDF support across models, see the multimodal guide.
Sources
- DeepSeek API Docs — Models & Pricing (retrieved 2026-09-11; notes (1) and (2) describe the routing and retirement)
- DeepSeek API Docs — Change Log (2026-09-10, DeepSeek-V4.1-Flash Release)
- Related: DeepSeek V4.1 Flash guide · pricing (DeepSeek) · token calculator · release timeline
⚠️ Disclaimer
- Information on this page is based on DeepSeek's official pages (api-docs.deepseek.com) as retrieved on 2026-09-11. Pricing and model availability may change without notice. Always verify on the official page before contracting or production use.
- The routing start time is shown converted from the officially stated "12:00 Beijing time on 2026-09-14" (Beijing = UTC+8, JST = UTC+9).
- Hands-on notes reflect the operator's own environment and usage; they do not guarantee the same results for everyone. In particular, the Claude Sonnet comparison was not a strict like-for-like test.
- This site is not a recommendation or an agent for any provider. DeepSeek is a trademark of DeepSeek. GitHub and GitHub Copilot are trademarks or registered trademarks of GitHub, Inc.