According to CNYES and Inside, DeepSeek's API documentation quietly switched to DeepSeek-V4-Pro-0813 on August 12, adding a 1,000,000-token context and 384,000-token output. Unofficial scores shared in DeepSeek's community group reportedly put agent performance close to Fable 5, but DeepSeek has not published these results officially.
When did DeepSeek V4 Pro 0813 launch, and what are its specs?
On the night of August 12, DeepSeek's official API documentation quietly switched its model listing to "DeepSeek-V4-Pro-0813," according to CNYES. The report states this change indicates the official V4 Pro version "very likely" has already gone live via API.
Both CNYES and Inside independently confirm the same technical specifications: the API's "models and pricing" page now lists DeepSeek-V4-Pro-0813 with a 1,000,000-token context window and up to 384,000-token maximum output, along with support for JSON Output, Tool Calls, the Responses API, the Anthropic API, and FIM (Fill-in-the-Middle).
- Fact: Model listing switched to DeepSeek-V4-Pro-0813 on the night of August 12 [Primary-source confirmed] — CNYES, Inside
- Fact: 1,000,000-token context, up to 384,000-token output, plus JSON Output/Tool Calls/Responses API/Anthropic API/FIM support [Primary-source confirmed] — CNYES, Inside
How credible are the reports that V4 Pro's agent scores approach Fable 5?
CNYES and Inside both report that multiple tech outlets, citing material posted in DeepSeek's official community group, say DeepSeek-V4-Pro-0813 scores in several tests are now close to Fable 5 — a marked improvement over the earlier V4 Pro preview version, with agent-related tests drawing the most attention.
However, both outlets flag the same caveat: these new scores are currently circulating only through DeepSeek's official group and community channels. DeepSeek has not yet published a complete set of V4-Pro-0813 benchmark results in its official changelog, so the figures remain unofficial and "cannot be fully equated with an official benchmark release."
- Fact: V4-Pro-0813 test results reportedly approach Fable 5, a marked improvement over the V4 Pro preview [Primary-source confirmed] — CNYES, Inside
- Fact: Scores are unofficial, sourced from DeepSeek's community group rather than an official changelog entry [Primary-source confirmed] — CNYES
What is DeepSeek's technical path from the V4 preview to V4 Pro?
Per CNYES, DeepSeek first introduced the V4 preview in April, at the time emphasizing that V4 Pro's Agentic Coding evaluation reached what was then a leading level among open-source models, with optimization for mainstream agent products including Claude Code, OpenClaw, OpenCode, and CodeBuddy, alongside support for a 1,000,000-token context and a thinking mode.
By late July, the V4 Flash official release further strengthened agent capability: DeepSeek published results on Terminal Bench 2.1, NL2Repo, DeepSWE, and Toolathlon verified, and stated that the official version's agent capability had substantially surpassed the V4-Pro-Preview.
- Fact: April V4 preview claimed leading open-source-level Agentic Coding results, with optimization for Claude Code, OpenClaw, OpenCode, CodeBuddy and 1,000,000-token context plus thinking mode [Primary-source confirmed] — CNYES
- Fact: Late-July V4 Flash release published Terminal Bench 2.1, NL2Repo, DeepSWE, and Toolathlon verified scores, claimed to surpass V4-Pro-Preview [Primary-source confirmed] — CNYES
Why hasn't DeepSeek's official changelog reflected the V4 Pro update yet?
Both CNYES and Inside report that DeepSeek's official update page still shows July 31 as its most recent announcement — the V4 Flash official release. At that time, DeepSeek explicitly stated only the V4 Flash API was being upgraded, while the V4 Pro API and the App/Web-side models remained unchanged. The same July 31 announcement pre-announced that "the official DeepSeek-V4-Pro version will be released as soon as possible."
This leaves a gap between the API's quiet August 12 model switch and the changelog's last confirmed entry from July 31.
- Fact: Official update page's latest post is dated July 31, covering the V4 Flash release only [Primary-source confirmed] — CNYES, Inside
- Fact: July 31 announcement pre-announced an official V4 Pro release "as soon as possible" [Primary-source confirmed] — CNYES, Inside
How does DeepSeek's pending price hike compare with Grok 4.6's pricing?
CNYES and Inside both report that DeepSeek has posted a notice on its pricing page stating it plans to raise overall API service prices in the near future, describing the expected increase as "substantial," though the actual new pricing has not yet been announced.
By contrast, both outlets cite Grok 4.6's published API pricing: $2 per million input tokens and $6 per million output tokens, which the reports describe as positioned to undercut other frontier models on cost.
- Fact: DeepSeek pre-announced a coming price increase described as "substantial," with no figures yet disclosed [Primary-source confirmed] — CNYES, Inside
- Fact: Grok 4.6 API pricing: $2/million input tokens, $6/million output tokens [Primary-source confirmed] — CNYES, Inside
Where does Grok 4.6 rank in agent-focused benchmarks?
According to CNYES and Inside, Grok 4.6 achieved 1,753 Elo in Artificial Analysis's GDPVal-AA v2 agent evaluation, placing first, and scored 61 on the Artificial Analysis Intelligence Index.
- Fact: Grok 4.6 scored 1,753 Elo on GDPVal-AA v2, ranking first [Primary-source confirmed] — CNYES, Inside
- Fact: Grok 4.6 scored 61 on the Artificial Analysis Intelligence Index [Primary-source confirmed] — CNYES, Inside
Benchmark and pricing figures at a glance
| Metric | DeepSeek-V4-Pro-0813 | Grok 4.6 |
|---|
| Context window | 1,000,000 tokens | — |
| Max output | 384,000 tokens | — |
| GDPVal-AA v2 (Elo) | Not disclosed officially | 1,753 (rank 1) |
| Artificial Analysis Intelligence Index | Not disclosed | 61 |
| API pricing (per 1M tokens) | Increase pre-announced, described as "substantial"; exact figures not yet released | $2 input / $6 output |
| Latest official changelog entry | July 31 (V4 Flash) | — |
What this means
The timeline laid out by CNYES and Inside shows a pattern of DeepSeek's product moving ahead of its own documentation: the V4 preview in April, the V4 Flash official release with published Terminal Bench 2.1, NL2Repo, DeepSWE, and Toolathlon verified scores in late July, and then a quiet API-side switch to V4-Pro-0813 on August 12 — three weeks after the changelog's last entry, which itself had only promised a V4 Pro release "as soon as possible." The agent scores said to approach Fable 5 come from the same unofficial, community-group channel that both outlets flag as unverified against a formal benchmark report. Meanwhile, the specific numbers available for comparison sit on opposite sides of the disclosure line: Grok 4.6 has published Elo and Intelligence Index scores plus concrete per-token pricing, while DeepSeek's V4 Pro has published only architecture figures (context and output length) and a qualitative promise of a "substantial" price increase, with performance scores and new pricing both still pending official release.