DeepSeek V4 Rumors vs Reality: What Actually Shipped
Lifecycle update — September 10, 2026: DeepSeek released V4.1 Flash. On DeepSeek's direct API,deepseek-v4-flashanddeepseek-v4-flash-vision-expnow forward to V4.1 Flash, anddeepseek-v4-prois scheduled to follow on September 14, 2026 at 12:00 Beijing time (04:00 UTC). On EvoLink,deepseek-v4-flashanddeepseek-v4-proare not affected and continue to serve DeepSeek V4 Flash and V4 Pro;deepseek-v4-flash-vision-expnow redirects to DeepSeek V4.1 Flash. See the official update, the V4.1 Flash model page and the migration guide.
This page was first published before DeepSeek V4 shipped. It now preserves the pre-release claims as a fact-check baseline. For the earlier August update, use the Pro 0813 release page. For the current transition, use the V4.1 Flash migration guide. For model selection, use the Pro vs Flash guide.
deepseek-v4-flash and deepseek-v4-pro unchanged. DeepSeek Transparency Center DeepSeek Models & PricingThe useful reason to keep this URL is not to repeat launch hype. It is to show which widely circulated claims were confirmed, which changed, and which still should not be used for production decisions.
Quick verdict on the old DeepSeek V4 claims
| Pre-release claim | What happened | Status |
|---|---|---|
| DeepSeek V4 would launch around February 17, 2026 | The V4 family was released on April 24, 2026 | Incorrect date |
| V4 would be one next-generation coding model | DeepSeek launched separate Flash and Pro tiers | Incorrect product shape |
| V4 would have million-token context | Both current API tiers list 1M context | Confirmed, with a defined 1M limit |
| V4 would beat Claude and GPT at coding | Independent evidence does not support a universal winner claim | Still unproven |
| Projected pricing could be used for budgets | Actual prices and later changes differed materially | Unsafe assumption |
| Engram research described the shipping architecture | The research was real; treating it as the complete released-model specification was not justified | Unverified attribution |
| Developers would need a new dated API name for each update | Callable IDs stayed stable while dated versions changed behind them | Incorrect API assumption |
The actual DeepSeek V4 release timeline
January and February 2026: rumor demand outran evidence
Pre-release coverage converged on a February launch window, a coding-first narrative, and claims that V4 would surpass Claude and GPT. Some reports also connected DeepSeek's Engram research directly to the unreleased model and published estimated API prices.
Those reports were useful as monitoring signals, not as production facts. There was no public V4 API model ID, final price table, dated release note, or independent result that justified building a budget or migration plan around them.
April 24, 2026: V4 shipped as a two-tier family
- Flash for lower-cost, higher-throughput workloads
- Pro for harder workloads that may justify a premium route
- 1M context and up to 384K output for both API tiers
- thinking and non-thinking modes
- tool calls and structured output support
The April release changed the correct user question from “when will V4 launch?” to “which V4 tier should this workload use?”
July 31, 2026: Flash became a much stronger baseline
This matters because many early articles compared the April Pro preview with a much weaker picture of Flash. Those comparisons should not be used to choose between the current builds.
August 13, 2026: Pro moved to 0813
deepseek-v4-pro request ID. DeepSeek also documents Responses API, Anthropic API, tool calls, and coding-agent integration routes. DeepSeek Models & Pricing DeepSeek Codex integrationWhat the February 17 rumor got wrong
The main failure was not that the reported date slipped. It was that a speculative date was repeated as though it were an official release schedule.
A production team should require at least one of these before treating a model launch as confirmed:
- an official vendor announcement or dated release note
- a public API catalog entry
- a documented model ID and endpoint
- a price or commercial availability page
- a reproducible successful request on the intended provider
Community posts, source-based news, leaked screenshots, and search demand can justify a watch page. They cannot justify “available now,” executable integration code, or a migration deadline.
What happened to the coding-dominance claim?
The strongest rumor-era claim was that V4 would beat leading Claude and GPT models at coding. That statement was never safe as a universal fact because model quality depends on the benchmark harness, reasoning effort, tools, timeout, retry policy, and task distribution.
The current independent picture is more measured:
- Pro 0813 scores 53 and Flash 0731 scores 52 on the Artificial Analysis Intelligence Index
- both score 78.65% on Terminal-Bench v2.1 in that evaluation
- Pro has a small aggregate Agentic Index lead
- Flash has higher measured output speed and lower direct token pricing
Engram: real research, unsafe product inference
DeepSeek and Peking University published Engram research on conditional memory. The mistake in early coverage was not discussing the paper; it was turning research proximity into a confirmed specification for an unreleased product.
Unless the model card or official technical report attributes a released model's behavior to a specific architecture, keep the wording precise:
- safe: “DeepSeek published Engram research before V4”
- unsafe: “V4 uses Engram and therefore has a specific cost or benchmark advantage”
Architecture speculation should not drive API selection. Observable route behavior, supported interfaces, task success, latency, and cost are the production inputs.
Pricing: why projected numbers aged badly
DeepSeek's current direct-provider prices per 1M tokens are:
| Model | Cache-hit input | Cache-miss input | Output |
|---|---|---|---|
| Flash 0731 | $0.0028 | $0.14 | $0.28 |
| Pro 0813 | $0.003625 | $0.435 | $0.87 |
What developers should do now
The launch-watch work is over. The remaining job is route evaluation:
- Use Flash 0731 as the low-cost baseline.
- Test Pro 0813 on planning, review, ambiguous debugging, and high-error-cost tasks.
- Measure cost per accepted task, including retries and human review.
- Keep the stable model aliases out of assumptions about stable behavior.
- Preserve a fallback route and rerun regressions after version changes.
EvoLink's unified API gateway is useful here because the integration can stay consistent while the model policy changes by workload. The gateway should not be used to blur upstream facts: direct-provider capabilities, EvoLink endpoint compatibility, and EvoLink pricing must remain separately verified.
FAQ
Is DeepSeek V4 released?
Yes. DeepSeek's transparency page dates the V4 release to April 24, 2026, and its current API pricing page lists Flash 0731 and Pro 0813.
Did DeepSeek V4 launch on February 17, 2026?
No. That date was a widely circulated pre-release expectation, not the actual release date.
Is DeepSeek V4 one model?
No. The current API family has Flash and Pro request IDs, with different prices, concurrency limits, and intended routing roles.
Did DeepSeek V4 beat Claude and GPT at coding?
No universal conclusion is supported. Current independent results make V4 competitive, but rankings vary by task and harness. Teams should test their own repositories and agent tools.
Does the 0813 update require a new model ID?
deepseek-v4-pro; 0813 is the dated underlying version documented by DeepSeek.
