Lesson 07 · DeepSeek Mastery Pro+ ~10 min read Updated August 15, 2026

Migrating to V4: the painless path off the legacy models.

If you built on an older DeepSeek model alias, the legacy names were retired on July 24, 2026 — calls to them now fail. The migration is small but has a couple of traps. Here is the checklist. Updated August 13, 2026: V4 Pro has now left preview — see section 03.

01What actually changes

Mostly just the model name in your request. Swap the retired alias for a current V4 identifier (Flash or Pro). The endpoint shape, message format, and your SDK stay the same — that is the benefit of the compatible API.

02The traps to check

Hard-coded model strings scattered across your code — search and replace them all. Behavior drift: V4 may format or reason slightly differently, so re-run your evals or spot-checks. Reasoning vs non-reasoning: make sure you are calling the variant you intend.

Legacy aliases are now retired: deepseek-chat and deepseek-reasoner stopped working on July 24, 2026 (15:59 UTC). If your code still calls them, those requests are failing now. Pin to explicit V4 names rather than floating aliases so you control when behavior changes.
Migration checklist

Find every model string → swap to a current V4 name → re-run your tests/spot-checks → confirm reasoning vs non-reasoning → set a spend limit → ship. Ten minutes for most projects.

03August 13, 2026: V4 Pro leaves preview

The flagship spent nearly four months in preview. On August 13, 2026 DeepSeek shipped the general-availability build, DeepSeek-V4-Pro-0813, and the deepseek-v4-pro endpoint now serves it. The smaller Flash model had already graduated on July 31, 2026.

What this means for your code: nothing, if you did the migration above. The endpoint name did not change, the context window (1M tokens) and maximum output (384K tokens) did not change, and the API rates carried over from preview unchanged. If you pinned to deepseek-v4-pro, you have quietly been moved onto a production build — which is the good outcome.

What did change is the label. "Preview" was your excuse to keep a fallback path warm; it is gone now, which makes this a sensible week to re-run the spot-checks from section 02 against the 0813 build and delete the fallback if it passes.

One open item, one now closed. Still open: the Hugging Face repositories host the April preview weights — DeepSeek has not published 0813 weights or said whether the GA build differs from preview beyond post-training. If you self-host, you are not running what the API is running. Now closed: the "significant overall API price increase" notice has landed with numbers and a date — see the next section.

04Two things that do change your code and your bill

Reasoning effort is now a three-way switch. Thinking mode on both deepseek-v4-pro and deepseek-v4-flash accepts low, high and max. DeepSeek's own guidance: low for simple tasks, high for day-to-day agent work, max for genuinely complex scenarios. If you were running everything at one setting, this is the cheapest quality-and-cost lever you have — and with the price change below, the one worth tuning first.

The API now speaks the OpenAI Responses format natively and is specifically adapted for Codex, with a one-click configuration script in DeepSeek's docs. If you were maintaining a translation shim between the Responses API and DeepSeek's chat-completions endpoint, you can probably delete it.

The price rise is dated: 16:00 UTC, August 16, 2026. DeepSeek is replacing flat rates with peak/off-peak billing, off-peak at half of peak, with peak hours running 01:00–04:00 and 06:00–10:00 UTC. V4 Pro output goes from a flat $0.87 per million tokens to $1.98 off-peak and $3.96 at peak; the cache-hit input rate goes from $0.003625 to $0.022/$0.044. Full rate card and the batch-scheduling workaround: what DeepSeek costs.
If you pin versions

Pinning to the dated string deepseek-v4-pro-0813 rather than the floating deepseek-v4-pro alias buys you the same protection the legacy-alias retirement taught everyone in July: you decide when behavior changes, not the calendar. The trade is that you have to actually watch the changelog.

Frequently asked

DeepSeek — your questions, answered

How do I migrate to DeepSeek V4?
Replace retired legacy model aliases with current V4 names (Flash or Pro) everywhere in your code, then re-run your tests. The endpoint and message format stay the same.
Will my old DeepSeek model name keep working?
No. The legacy aliases deepseek-chat and deepseek-reasoner were retired on July 24, 2026 and no longer work. Move to the explicit V4 names, deepseek-v4-flash or deepseek-v4-pro.
Does V4 behave differently from older models?
It can format and reason slightly differently, so re-run your evaluations or spot-checks after switching rather than assuming identical output.
Do I need to change my code beyond the model name?
Usually not — the compatible API keeps the same shape. Just update model strings and verify reasoning vs non-reasoning variants.
Is DeepSeek V4 Pro still in preview?
No. DeepSeek shipped the general-availability build DeepSeek-V4-Pro-0813 on August 13, 2026, ending a preview that ran nearly four months. The deepseek-v4-pro endpoint now serves the 0813 build, with the same 1M-token context window and 384K maximum output.
Do I need to change anything for the V4 Pro GA build?
No, if you already migrated to the explicit V4 names. The endpoint, context window, output ceiling and API rates all carried over from preview. It is still worth re-running your spot-checks against the 0813 build before you delete any fallback path.
Keep going
← Back to all DeepSeek lessons