Blog • KI / Releasekalender

Wann kommt das nächste DeepSeek-Modell? LLM-Releasekalender 2026

Search “DeepSeek next model” and you will often get a specific calendar day. That day is usually not from DeepSeek.

As of 2026-09-05 DeepSeek has not announced a model named V5 and has not published a release date. The current generation is V4: DeepSeek-V4 Preview on April 24, a Flash API update in July, Pro GA on August 13, and an experimental Vision model on August 21. Those events match the DeepSeek API changelog.

OpenAI, Anthropic, and Google ran a different cadence this summer: a numbered family, then incremental drops a few weeks apart. GPT-5.6 reached GA on July 9, Claude Fable 5.1 on September 1, Gemini 3.8 Flash on September 2.

Dieser Artikel behandelt:

• Whether DeepSeek’s next model has an official date
• A confirmed 2026 timeline across four labs
• What the current flagship is for GPT, Claude, and Gemini
• How to separate confirmed facts, inference, and rumor
• How to pin the model field and validate JSON locally after a switch

Die kurze Antwort

DeepSeek’s next generation (often called V5) has no official ship date. What you can put in production is still the three V4 API names.

The flagship you should watch today is not the same question as “will the next generation arrive this month”:

If you just need a model that works, change the model field in config. Do not wait for a second-hand post that invents a calendar day.

Bestätigter Zeitplan 2026

This table only includes events that match an official blog, API changelog, or product page. Two labs on the same day is not a formatting error.

This table answers “what already happened,” not “which lab ships next Tuesday.” Mixing those two is where fake-precise dates come from.

DeepSeek: V4 wird noch iteriert, kein offizielles V5

The April 24 V4 Preview split the line into Pro (~1.6T / 49B active) and Flash (284B / 13B active). The next three months were not “wait for V5”—they pushed the same API names into production:

2026-07-24 retired deepseek-chat / deepseek-reasoner; 2026-07-31 shipped V4-Flash-0731; 2026-08-13 took V4-Pro-0813 to GA; 2026-08-21 added deepseek-v4-flash-vision-exp. The call path did not change: base_url stays https://api.deepseek.com, only model changes.

For Function Calling, JSON Mode, and thinking effort, see DeepSeek V4-Pro explained. This article only answers when the next generation arrives.

If you replay the ~four-month gap from V3.2 (2025-12) to V4 (2026-04), the earliest window lands in Q4 2026. That is interval arithmetic, not a commitment. V4 is still shipping Vision and Pro GA. Until DeepSeek names V5, pin V4 in production.

OpenAI: GPT-5.6-Familie und Astra

On 2026-07-09 OpenAI took the GPT-5.6 family to GA: Sol is the flagship, Terra is everyday work, Luna is cost and speed. In the API, the gpt-5.6 alias routes to Sol. On August 21 OpenAI cut Sol API prices, with the promo running at least through 2026-11-21.

• Sol: hard coding, research, computer use
• Terra: near prior-flagship quality, cheaper
• Luna: fastest and cheapest

Astra is not a rename of GPT-5.6. Path to Astra says it is the first model to hit the Preparedness Framework Critical cybersecurity threshold; parts of the release were delayed to strengthen safeguards. The wording is “available soon,” with the most advanced cyber capabilities on the Daybreak trusted path—not “it already replaced gpt-5.6 in your code.”

For developers: GPT-5.6 is what you can pin in CI today. Treat Astra’s public surface as whatever the next official product page says. Do not copy a review-site date into crontab.

Claude: Fable 5.1 und Mythos 5.1

Anthropic shipped fast this summer: Fable 5 around July 1 GA, Opus 5 on July 24, and Claude Fable 5.1 / Mythos 5.1 on September 1. Fable 5.1 is the public flagship; Mythos 5.1 is the same model with more permissive safeguards, limited to trusted cybersecurity and life-science work.

The API name is claude-fable-5-1. For structured output and tool calls, see the Claude Fable 5.1 JSON guide.

There is no official date for 5.2 or the next major. The July-to-September follow-up pace is a recent habit, not a published calendar.

Gemini: 3.8 Flash und Flash Cyber

Google shipped Gemini 3.8 Flash on 2026-09-02 and called it the third Flash in six weeks. 3.7 Flash landed on August 13. 3.8 targets reasoning and coding at an introductory $0.75 / $3.75 per million tokens through 2026-12-31; rates double on 2027-01-01.

Gemini 3.8 Flash Cyber sits beside it for vulnerability work and patching, via Fairwind, for trusted governments, critical infrastructure, and maintainers—the same product pattern as Mythos and Astra’s trusted access.

There is no official Gemini 4 date. What you can put in config is 3.8 Flash. “Next generation” is unannounced here too.

Prognosen lesen: bestätigt, geschlossen, Gerücht

Release-calendar posts fail when they print three kinds of information as the same “ships on this day.”

Filter any “2026 release timeline” with these four steps:

1. 1
   Find a first-party link

   Official changelog, product blog, or API model page. Aggregators and social screenshots stay unconfirmed.

2. 2
   Separate Preview, GA, and retirement

   V4 Preview, V4-Pro GA, and retiring deepseek-chat are three different events. Calling a Preview “full launch” mis-times migrations.

3. 3
   An alias is not a new model

   gpt-5.6 routes to Sol; deepseek-chat pointed at V4-Flash until retirement. Changing an alias is not the next generation.

4. 4
   Trusted access is not public GA

   Mythos, Flash Cyber, and parts of Astra sit on allowlists. Check whether your account is on the list before you write an external SLA.

So verfolgen Entwickler Versionen

Model IDs will keep moving. Do not scatter the string across ten files. Keep one validatable JSON config, override it with an env var, and reject retired IDs.

{
  "routing": {
    "deepseek": {
      "default": "deepseek-v4-pro",
      "fast": "deepseek-v4-flash",
      "vision": "deepseek-v4-flash-vision-exp"
    },
    "openai": {
      "default": "gpt-5.6",
      "flagship": "gpt-5.6-sol"
    },
    "anthropic": {
      "default": "claude-fable-5-1"
    }
  },
  "retired": ["deepseek-chat", "deepseek-reasoner"],
  "pin_snapshots": true
}

Load that config at startup and refuse retired IDs:

import json
import os

def load_routing(path="models.json"):
    with open(path, encoding="utf-8") as f:
        return json.load(f)

cfg = load_routing()
model = os.environ.get("LLM_MODEL") or cfg["routing"]["deepseek"]["default"]
if model in cfg["retired"]:
    raise ValueError(f"retired model id: {model}")
print("using", model)

After a switch, pin one sample of tool arguments or structured output and re-run the schema. What breaks is usually field names and enums, not the prompt.

JSON nach dem Wechsel in JSONNote prüfen

Moving from V4-Flash to V4-Pro, or from GPT-5.6 to Claude Fable 5.1, is often slowest in the JSON you get back. JSONNote runs locally in the browser:

1. 1
   Format first

   Paste the response into JSON-Formatierer to see missing commas, truncation, and indent errors immediately.

2. 2
   Validate against the same schema

   Paste tool definitions and model arguments into JSON Schema. After a model switch, required and enum fail first.

3. 3
   Diff two calls

   Use JSON Diff to see what changed in structured output—missing fields versus type changes.

4. 4
   Share the debug case

   Use URL-Hash-Teilen to embed sample JSON in a link. Colleagues can reproduce it; nothing is uploaded.

FAQ

When does DeepSeek V5 release?

There is no official date. As of 2026-09-05 DeepSeek has not announced V5. The current generation is V4. Pages that name a specific day are not official.

Which DeepSeek model should I call now?

Use deepseek-v4-pro for hard work, deepseek-v4-flash for high QPS, and deepseek-v4-flash-vision-exp for images. Do not keep writing deepseek-chat.

Are GPT-5.6 and Astra the same product?

No. GPT-5.6 is GA. Astra is a separate line marked on 2026-09-01 as the first Critical cyber-capability model; it had not replaced the gpt-5.6 alias.

Can I trust a specific release date I saw online?

Treat dates without an official changelog or product blog as rumors. Only write “confirmed” when the date matches a first-party page.

Will structured output break after I switch models?

Yes—especially field names, required keys, and enums. Re-run the same JSON Schema after you change model. “OpenAI-compatible” does not mean identical fields.

Zusammenfassung

All four labs shipped this summer. The cadence was not the same.

DeepSeek’s next generation has no official date; V4 is what you can call. GPT-5.6, Claude Fable 5.1, and Gemini 3.8 Flash are already live.

Keep confirmed facts and forecasts apart. Keep model IDs in one JSON file. After a switch, format, schema-check, and diff locally. Names will keep changing. The contract layer is what you control.

← Zurück zum Blog