DeepSeek V4 release date: April 24, 2026, still labeled preview.

Quick answer: the verified timeline.

DeepSeek announced V4 on April 24, 2026, in its own channels and API docs news, describing the release as a preview that was live and open-sourced the same day: weights for both variants published under MIT on Hugging Face, API access immediate. Mainstream coverage, including CNBC's same-day report, matches the date. Two qualifiers keep the answer honest. The launch came months after a widely rumored mid-February window tied to Lunar New Year, a reminder that this vendor's release dates firm up only when they happen. And as of July 2026 the release is still labeled a preview: the models serve production traffic and the vendor has scheduled its legacy ids for deprecation, but no date for a stable V4 has been announced, and anything you read about one is inference.

What shipped on April 24: verified specs.

The performance claims attached to the launch, Pro rivaling top closed-source models and leading open models on math, STEM, and coding evaluations, are the vendor's own. They have held up well enough in community testing to make V4 the default value family of mid-2026, but they are launch claims, and your own evaluation set outranks them for anything that matters.

• Two variants from day one: DeepSeek-V4-Pro at 1.6T total parameters with 49B active, and DeepSeek-V4-Flash at 284B total with 13B active, per the vendor's own announcement.
• MIT open weights for both, published on Hugging Face with the release.
• 1M-token context on both variants, framed by DeepSeek as "the era of cost-effective 1M context length", with output up to 384K tokens per the API docs.
• Thinking and non-thinking modes on both variants, plus JSON output, tool calls, and chat prefix completion; fill-in-the-middle restricted to non-thinking mode.
• Published pricing (July 2026): Flash at $0.14 per million input tokens and $0.28 output; Pro at $0.435 and $0.87; automatic context caching bills cache-hit input at $0.0028 and $0.003625 respectively.

From V3 to V4: what actually changed.

The V3 line carried DeepSeek from late 2024 through early 2026, with the V3.2 work on sparse attention closing it out; that attention research had visible influence beyond the company, and Z.ai's release notes credit DeepSeek Sparse Attention in GLM-5. V4 consolidates the lineage into a different shape. Three changes matter to integrators. First, the context leap: the V3 line served 128K-class windows, and V4 serves 1M on both variants at prices that make long context routine rather than premium. Second, the two-variant split: instead of one general model plus a separate reasoner, V4 ships as a scale pair, Flash for volume and Pro for depth, each supporting thinking and non-thinking modes. Third, the id consolidation: the old deepseek-chat and deepseek-reasoner names are deprecated as of July 24, 2026, per the vendor's docs, mapping to V4 Flash's non-thinking and thinking modes in the meantime. The upgrade path is explicit: new code targets deepseek-v4-flash and deepseek-v4-pro directly.

What "preview" means for building on V4.

DeepSeek has not attached a stability commitment or a finalization date to the preview label, so the working assumption is that model behavior can shift as the vendor iterates. In practice V4 has been the workhorse family of the value tier since spring, and the deprecation of the legacy ids signals the vendor treats it as the present, not an experiment. The engineering posture that fits: build on the V4 ids now, keep a small regression eval you rerun when behavior seems to move, and avoid coupling your pipeline to quirks (exact formatting habits, reasoning verbosity) that a preview-stage update could change. Routing through a multi-model endpoint adds one more hedge, since a temporary shift in one family leaves you a one-string fallback to GLM or Claude ids rather than an outage.

What carried over from V3, for teams that built on it.

The discontinuities are the ones already covered: a 1M context where V3 served 128K-class windows, the two-variant split, and the retirement of the separate reasoner id in favor of thinking modes. For most integrations the honest migration cost is an eval rerun and two string edits.

• The API surface: V4 speaks the same OpenAI-compatible chat completions shape the V3 line did, so existing client code migrates by changing the model string, not the integration.
• The pricing posture: the family stays the market's low-rate anchor, with the published V4 rates extending the pattern rather than resetting it, and automatic context caching continuing to discount repeated prefixes on the direct platform.
• The open-weight cadence: like its predecessors, V4 shipped weights with the announcement rather than after it, MIT-licensed on Hugging Face.
• The usage-field conventions: responses still report exact token usage per request, including the cache hit and miss split on the direct platform, so V3-era cost dashboards keep working against V4 traffic.

How to call DeepSeek V4 today.

Both released ids are live in the APIsRouter catalog and served with the 1M context window: deepseek-v4-flash and deepseek-v4-pro, standard OpenAI-compatible request shape, one key shared with the GLM, Claude, and GPT families. No DeepSeek platform account is required through the gateway, and the free starting credit means release-day curiosity costs nothing to satisfy.

import os
from openai import OpenAI

client = OpenAI(
    api_key=os.environ["APISROUTER_API_KEY"],
    base_url="https://api.apisrouter.com/v1",
)

resp = client.chat.completions.create(
    model="deepseek-v4-flash",   # released April 24, 2026
    messages=[{"role": "user", "content": "What changed between DeepSeek V3 and V4?"}],
    max_tokens=2000,
)
print(resp.choices[0].message.content)

Pay-as-you-go · transparent per-model pricing

Selected models are priced below official list prices. Exact input, output, cache, and per-request prices are shown for each model.

Why V4 release information stays confusing.

Most of the noise traces to three patterns. Pre-release rumor congealed into pseudo-fact: the mid-February window was repeated widely enough that some pages still cite it as a slipped official date, which it never was. The preview label reads differently to different audiences: "released as preview" and "not yet released" both circulate, and for an integrator the operational truth is simple, the ids are live and billable. And the legacy-id deprecation gets misread as a new release event rather than what it is, the old names being retired onto V4's modes. The check that cuts through all of it is the same one that works for every family: list the models your endpoint serves. If deepseek-v4-flash and deepseek-v4-pro appear in /v1/models, V4 is released for you, whatever label the vendor attaches.

curl -s https://api.apisrouter.com/v1/models \
  -H "Authorization: Bearer $APISROUTER_API_KEY" | grep deepseek-v4

FAQ