What Did DeepSeek Announce About V4 Pro in September 2026?
DeepSeek's V4.1 Flash announcement on 10 September 2026 stated three things about the existing lineup: V4 Flash and V4 Flash Vision Exp are retired with their legacy names temporarily routed to V4.1 Flash; tests by multiple parties put V4.1 Flash ahead of V4 Pro on performance, cost, speed and total runtime, so DeepSeek is phasing out V4 Pro; and starting 04:00 UTC on 14 September 2026 all deepseek-v4-pro requests would route to V4.1 Flash at V4.1 Flash rates until V4.1 Pro launches.
On 11 September 2026 the API pricing page, which still lists deepseek-v4-pro as a model, added a footnote: in response to user demand, DeepSeek has decided to continue providing API services for V4 Pro after 14 September with the billing method unchanged, and will provide further notice should there be any changes. Nothing on the announcement page was edited to match. At the time of writing, a reader of the announcement believes V4 Pro is being switched off this week and a reader of the pricing page believes it is not.
| Source | Date | What it says about V4 Pro | How to weigh it in 2026 |
|---|---|---|---|
| V4.1 Flash announcement | 10 September 2026 | Phasing out; all V4 Pro requests route to V4.1 Flash at Flash rates from 04:00 UTC 14 September until V4.1 Pro launches | States the intent and the eventual destination. |
| API pricing page, footnote 2 | 11 September 2026 | V4 Pro API service continues after 14 September, billing unchanged, further notice if anything changes | Later, more operational, and on the page that governs billing. Treat as current. |
The reasonable synthesis is that V4 Pro is a model with an announced end of life and a paused switch-off. The August V4 Pro rate change was also announced with days of notice, so a team on V4 Pro in 2026 should assume the next notice will be short too. The V4.1 Flash overview covers what the replacement model is.
Why Should Most V4 Pro Callers Migrate to V4.1 Flash Anyway in 2026?
Because V4.1 Flash is cheaper on every line of the rate card, higher on most of DeepSeek's own agentic benchmarks, and has five times the concurrency, and because DeepSeek has told you where the traffic is eventually going. The pricing guide has the full rate comparison; the short version is that V4.1 Flash is about 7 times cheaper on cache hits, 4.4 times on cache-miss input and 3.3 times on output.
On DeepSeek's instruct benchmarks at maximum effort, V4.1 Flash beats V4 Pro on Terminal-Bench 2.1, 3.0 and 4.0, DeepSWE, AutomationBench, Agent's Last Exam, CyberGym, SEC-Bench Pro, HLE with tools and Codeforces. V4 Pro's base model retains an edge on knowledge-heavy rows such as SimpleQA-Verified (55.2 versus 42.3), MultiLoKo (50.9 versus 45.5) and LongBench-V2 (51.5 versus 45.2), and on GPQA Diamond (92.4 versus 90.9). The benchmark guide decodes the full table.
What Changes for Callers Moving to V4.1 Flash in 2026?
Most integrations will work by changing the model string, because both models sit behind the same OpenAI-format and Anthropic-format endpoints with the same feature list. The differences are in behaviour and limits, and each one needs a line in a migration plan.
| Dimension | DeepSeek V4 Pro 0813 | DeepSeek V4.1 Flash | Migration action |
|---|---|---|---|
| Model name | deepseek-v4-pro | deepseek-flash | Change the string. Legacy deepseek-v4-flash names also resolve to V4.1 Flash. |
| Endpoints | api.deepseek.com (OpenAI format), api.deepseek.com/anthropic | Same | None. |
| Vision | Not supported | Native image input | Retire any separate vision model or OCR step; test image prompts. |
| Thinking mode | Non-thinking and thinking, thinking default | Same | Confirm the mode your calls use; explicit is safer than default. |
| Reasoning effort | Not described as a 1 to 100 dial on the pricing page | Integer 1 to 100; benchmarks at 100 | Set per workflow. Do not inherit 100 into production. |
| Context and output | 1M context; 384K max output | 1M context; 384K max output | None. |
| Concurrency limit | 500 | 2,500 | Raise client-side rate limiters if they were tuned to 500. |
| Features | JSON output, tool calls, Responses API, Anthropic API, chat prefix (beta), FIM non-thinking only (beta) | Same | None. |
| Pricing | 0.044 / 1.32 / 3.96 USD per 1M at peak | 0.006 / 0.30 / 1.20 USD per 1M at peak | Re-run the budget; expect a reduction, then verify token counts. |
| Prompt format (self-host only) | V4 prompt format | V4.1 prompt format via encoding.py or deepseek-recipe | API callers unaffected. Self-hosters see the self-host guide. |
How Should a Team Migrate From V4 Pro to V4.1 Flash in 2026?
The same way any model migration should go in 2026: as a configuration change validated by an evaluation set, not as a search-and-replace followed by hope. The steps below assume a team already has some V4 Pro traffic in production.
- Inventory the callers. List every workflow, script and vendor integration that sends deepseek-v4-pro, with its monthly token volume by type. If the model string is hardcoded in more than one place, fix that first so the next migration is one edit.
- Freeze an evaluation set. Twenty to fifty real tasks per workflow with known good outputs. Include the knowledge-heavy and long-context cases, because those are where V4 Pro's base model still leads.
- Run V4.1 Flash at three effort levels. Low, mid and high, recording accuracy, output tokens and latency at each. Pick the lowest level that passes.
- Compare cost per passing task, not cost per token. A model that needs two attempts is not cheaper at half the price.
- Test the vision path if you have one. If a workflow currently runs OCR or a separate vision model before V4 Pro, try V4.1 Flash on the raw image; DocVQA at 95.6 on the model card suggests it may replace the step.
- Re-tune rate limiters and retries. The concurrency limit rises from 500 to 2,500. Client-side throttles set for V4 Pro will leave throughput unused.
- Schedule flexible work off-peak. Peak is 06:30 to 09:30 and 11:30 to 15:30 IST on weekdays. Anything batchable belongs outside those windows at half price.
- Cut over per workflow, not all at once. Move the workflows that passed, keep the rest on V4 Pro while it lasts, and log the model name on every request so the two are separable in analytics.
- Watch the two DeepSeek pages weekly. The retirement plan changed within 24 hours in September 2026. The next change may be as fast.
DeepSeek's 10 September plan was that deepseek-v4-pro requests would be served by V4.1 Flash at V4.1 Flash rates. If that plan is reinstated, callers who have not migrated will get the new model silently, with different vision support, a different effort dial and different output behaviour, at a lower price. Silent is the problem. A team that has already run the evaluation set knows what it will get; a team that has not will find out from its customers.
When Should a Team Stay on DeepSeek V4 Pro in 2026?
When its own evaluation set shows V4 Pro ahead on the workflow and the cost difference does not matter at that workflow's volume. On DeepSeek's base-model table V4 Pro leads on fact recall (SimpleQA-Verified by 13 points), multilingual knowledge (MultiLoKo by 5 points) and long-context understanding (LongBench-V2 by 6 points). A workflow that depends on the model knowing things rather than doing things, in a language other than English or Chinese, or across very long documents, is the candidate for staying. Everything else is a candidate for moving. Either way, staying is a decision with an expiry date, because DeepSeek has stated the destination.
What Are the Common Migration Mistakes in 2026?
- Acting on the announcement without reading the pricing page. The pricing page is the later document and currently says V4 Pro continues.
- Acting on the pricing page and forgetting the announcement. The end-of-life intent is published. Plan for it.
- Search-and-replace migration. Vision, effort and thinking defaults differ. Test first.
- Inheriting effort 100. It is the benchmark setting and the most expensive one.
- Leaving the rate limiter at 500. Five times the concurrency is available.
- Not logging the model name per request. If routing changes on DeepSeek's side, you need to know which model answered.
- Migrating everything at once. Per-workflow cutover with the evaluation set as the gate is slower for a week and safer for a year.
Key Takeaways for 2026
- DeepSeek announced V4 Pro routing to V4.1 Flash from 14 September 2026, then said V4 Pro continues with billing unchanged. Both pages are live.
- Treat V4 Pro as end-of-life with no date. Have the migration tested.
- V4.1 Flash is 3 to 7 times cheaper per line, ahead on most agentic benchmarks, and offers 2,500 concurrency versus 500.
- Caller-facing changes: model string deepseek-flash, native vision, reasoning effort 1 to 100, thinking default, Flash pricing.
- Migrate per workflow through an evaluation set; compare cost per passing task; set effort per workflow.
- Stay on V4 Pro only where your tests show it ahead on knowledge-heavy, multilingual or long-context work.
Distk runs model migrations for growth teams across India and internationally as configuration changes gated by evaluation sets, so that the next vendor notice, however short, is an afternoon's work. If you have DeepSeek V4 Pro in production in 2026, that migration plan is where we start.