FULL STORY

DeepSeek V4 Pro: from leaks to launch and pricing backlash

After weeks of leaks, DeepSeek launched V4 Pro on Aug 13, only to pull a questionable 0813 build amid weak benchmarks, then sparked backlash with API price hikes of up to 1100% and new peak/off-peak pricing effective Aug 17.

2026-08-11 ~ 2026-08-18 · 7 episodes · 105 posts

Episode 1 · DeepSeek V4 Pro Launch Imminent, API and Timeline Leaked (2026-08-11, 5 posts)

Recent community and social media chatter suggests the imminent release of DeepSeek V4 Pro. Developer @dejavucoder claimed the official GA release is being prepared, and @jiayuanjy spotted an API link for deepseek-v4-pro-0813, corroborating the rumor. Multiple bloggers predict a launch as soon as August 12 or 13 evening.

Confirmed

  • API exposure: A user found the deepseek-v4-pro-0813 API link, confirming a version update.

Unconfirmed

  • Release timing: Rumored for August 12 or 13 evening, pending official announcement.
  • Performance expectations: Blogger @AlchainHust speculates V4 Pro could approach Opus 5 / Fable 5 levels based on V4 Flash, but this is subjective.

Why it matters

  • DeepSeek V3's success has raised high expectations for V4 Pro. If performance approaches or matches top competitors, it could reshape the LLM competitive landscape.

Episode 2 · DeepSeek V4 Specs and Pricing Leak Spark Debate (2026-08-11, 5 posts)

Leaked DeepSeek API docs and online rumors have revealed core specs and suspected pricing for the next-generation V4 series, drawing significant community attention. Model capabilities are largely clear, but pricing reports conflict sharply, and all information awaits official confirmation.

Confirmed

  • Multiple media and bloggers report API doc pages showing DeepSeek plans to release V4 series models: deepseek-v4-flash and deepseek-v4-pro.
  • Both models support up to 1M context length and 384K max output.
  • User @赛博禅心 notes that model name deepseek-v4-pro and endpoint DeepSeek-V4-Pro appear accessible via API, though no official announcement has been made.

Unconfirmed

  • Pricing rumors diverge: @teortaxesTex relays leaked prices of $2.40/M input, $4.80/M output, $0.20/M cache hit (about 5.5x current prices).
  • @haider1 posts benchmark and pricing data suggesting disruptive low prices: $0.435/M input, $0.87/M output, $0.07/M cache input (exact figures per original post).

Why it matters

  • If the 1M context spec is real, DeepSeek V4 would rank among the top in long-context capability. The final pricing, whether a sharp increase or continued low cost, will directly impact the API price war in the LLM industry.

Episode 3 · DeepSeek launches V4-Pro with peak pricing, API price hike sparks debate (2026-08-12, 62 posts)

On Aug 13, DeepSeek officially released DeepSeek-V4-Pro with 1M token context, FP8 quantization, and significant gains on agent benchmarks, while introducing peak/off-peak pricing that raised API costs, sparking developer backlash over unpredictability and high expenses.

Confirmed

  • Model release: DeepSeek announced V4-Pro and listed V4-Pro-0813 on API pricing page and Hugging Face. It supports 1M token context, uses FP8 and 8-bit quantization, and offers flexible reasoning intensity control. V4 Pro and V4 Flash are available on OpenRouter and other platforms.
  • Performance gains: According to official reports, V4-Pro showed large improvements over preview on agent benchmarks like Terminal-Bench 2.1, CyberGym, and DeepSWE. Some commenters say it rivals open-source SOTA and approaches top closed models.
  • Pricing strategy: Base input $0.435/M tokens, output $0.87/M. Peak/off-peak pricing goes into effect Aug 16. For V4-Flash, peak output rises to $1.32/M, off-peak halves.

Unconfirmed

  • Model weights: Hugging Face does not yet offer V4-Pro-0813 weights; no open-source plan announced.
  • Price increase disputes: Reports vary: some say off-peak base nearly doubled and peak doubles again; others claim 12x peak surge; one tweet says flagship rose 355% from $0.87 to $3.96.

Why it matters

  • Post-training test: Commentators say V4 Pro's performance will define Chinese AI labs' post-training strength.
  • Cost vs demand: Disruptive base pricing plus extreme dynamic peak pricing is called a "demand-destroying" approach, signaling a shift away from API compute subsidies. Some third-party aggregators try to maintain old prices, but heavy API users face serious cost control challenges.

42 more related posts →

Episode 4 · DeepSeek V4 Pro Released and Potentially Withdrawn Due to Wrong Version (2026-08-13, 18 posts)

DeepSeek quietly released the V4 Pro 0813 version around August 13. However, it faced criticism for marginal or even regressed performance in general benchmarks and was allegedly withdrawn by officials due to a 'wrong version' error. The model showed impressive results in cybersecurity and coding domains, achieving top-tier recall rates at a very low cost, but the stark contrast between its mediocre general performance and specialized breakthroughs sparked widespread controversy in the community.

Confirmed

  • Marginal General Performance Improvement: Multiple developers pointed out that V4-0813 performed almost identically to V4-Flash-0731 in most benchmarks, with SciCode scores even regressing, and its Intelligence Index only 1 point higher than Flash. @haider1 added that while it beat Opus 4.8 at an extremely low cost, the score gap between frontier models is hard to perceive in daily use.
  • Code and Text Rankings: According to Chatbot Arena's AutoEval, DeepSeek-V4-Pro (Max) scored 1607 in the WebDev Code Arena, ranking 2nd among open-source models, and scored 1465 in the Text Arena, ranking 5th.
  • Outstanding Security but Specialized: Under pass@3 settings, V4 Pro rediscovered 87.5% of baseline CVEs, higher than Opus 5 and Qwen 3.8's 81.3%. @teortaxesTex tested it finding an RCE vulnerability in an open-source project in under 30 minutes but noted it ranked last in precision despite top recall. Tests also showed V4 Flash outperformed the Pro version in Canvas coding tasks.
  • Cost and Volume Disadvantages: The cost per task is only $0.14, 17 times cheaper than the higher-ranked Kimi K3. However, @scaling01 noted that its parameter volume reaches 1.6T, 5 times larger than the Flash version, leading to accusations of extremely low cost-effectiveness.
  • Suspected Wrong Version and Withdrawal: @orange and @op7418 indicated the version was likely withdrawn due to a 'wrong version'; while the API price page lists 0813, the update log does not mention it, and promotional banners were removed. @orange suggested that underwhelming performance might have led to the cancellation of a planned price hike.

Unconfirmed

  • Score Gaming Controversy: @teortaxesTex mentioned the model's 11-point surge in the Vals index might be inflated, accused of relying solely on pass@2 gaming, but officials have not responded.
  • Reason for Withdrawal: The withdrawal is currently only speculated as a 'wrong version'; there is no official notice confirming the specific cause.

Why it matters

The release and withdrawal of DeepSeek V4 Pro highlight the dilemma of 'marginal gains' in current general LLM benchmarks. However, its breakthrough performance and low cost in security and coding domains offer new ideas for vertical applications. The official's low-key promotion and withdrawal also reflect the complex game between frontier model iteration and evaluation expectations.

Episode 5 · DeepSeek Releases V4 Pro (2026-08-13, 5 posts)

DeepSeek officially released V4 Pro, featuring a MoE architecture with 1.6T to 1.7T total parameters and 49B active parameters. Its inference FLOPs are only 27% of V3.2. The model is open-sourced under the MIT license with API pricing about 1/31 of GPT-5.6. It scored close to GPT-5.6 in the LMArena Code Arena and is now available on Baseten.Confirmed- The model uses a MoE architecture with 1.6T-1.7T total parameters and 49B active parameters; inference cost is only 27% of V3.2.- The model is open-sourced under the MIT license and supports zero data retention (ZDR) by default.- It scored 1607 in the LMArena Code Arena WebDev AutoEval, only 15 points lower than GPT-5.6 Sol xHigh.- Baseten launched the model API on Day 0.- The model natively supports the Responses API and features advanced agent capabilities.Why it mattersThis model challenges closed-source flagships with extremely low pricing and an open-source strategy. Its performance approaches GPT-5.6, significantly reducing the inference cost of high-performance models.

Episode 6 · DeepSeek Hikes V4 API Prices Up to 1100% Against the Trend, Adds Peak/Off-Peak Billing (2026-08-15, 5 posts)

DeepSeek sharply raised V4 series API prices starting August 16 — by up to 1100% — and introduced peak/off-peak billing, with time-based pricing for V4 Pro effective August 17 at 16:00 UTC. An Artificial Analysis evaluation found V4 Pro's intelligence index up 8 points to 53 versus the April version while blended prices climbed 264%. Coming amid August price cuts by OpenAI and Google, the counter-trend move is attributed to compute shortages driven by surging V4 Flash demand.

Confirmed

  • New prices took effect August 16: V4-Flash and V4-Pro double in peak hours (UTC 01:00–04:00 and 06:00–10:00) and halve off-peak, with increases up to 1100% (@eyishazyer, @heypearlai)
  • V4 Pro moved to time-based pricing from 16:00 UTC on August 17: peak output at $3.96 per million tokens, off-peak at $1.98 (relayed by @AccBalanced)
  • Artificial Analysis evaluated V4 Pro 0813: intelligence index of 53, up 8 points from the April version, with blended prices up 264% (@ArtificialAnlys)
  • OpenCode Go updated its limits to reflect the new prices and teased more news on "Operation Cheepseek" (relayed by @gaganghotra)

Timeline

  • Aug 15: DeepSeek announced the price hike; Artificial Analysis published its V4 Pro 0813 evaluation
  • Aug 16: New prices and peak/off-peak billing took effect
  • Aug 17: V4 Pro time-based pricing detailed, effective 16:00 UTC that day

Why it matters

  • @eyishazyer argues that raising prices while OpenAI and Google cut theirs in the same month is essentially price leverage against compute shortages caused by surging V4 Flash demand
  • By Artificial Analysis's math, a 264% blended cost increase bought only an 8-point intelligence gain, visibly eroding DeepSeek's long-standing low-price, high-value positioning
  • Peak/off-peak billing ties cost to call timing, forcing heavy users to rethink job scheduling and cost structures

Unconfirmed

  • Specific V4 Flash price figures were not fully disclosed in the material, and details of "Operation Cheepseek" remain unclear

Episode 7 · DeepSeek Hikes V4 API Prices Up to 1100%, Introduces Peak/Off-Peak Pricing (2026-08-17, 5 posts)

DeepSeek's new API pricing took effect August 17 at 16:00 UTC, raising V4 series model prices by roughly 50% to over 1100% depending on model and time slot, and introducing peak/off-peak pricing for the first time. The move marks a clear departure from DeepSeek's long-standing ultra-low-price strategy.

Confirmed

  • Effective August 17 at 16:00 UTC, confirmed by APPSO and Zhidongxi; peak hours are 9:00–12:00 and 14:00–18:00 at full price, with the remaining 17 off-peak hours uniformly at 50% off.
  • V4-Flash peak output pricing rose from $0.28 to $1.32 per million tokens, roughly 4.7x.
  • V4-Pro peak input rose from 3 to 9 yuan and output from 6 to 27 yuan, with cached-input pricing also adjusted.
  • DeepSeek also launched V4 Pro, priced up to 14x higher than V4 Flash, creating a tiered lineup for different workloads and performance needs.

Why it matters

  • Commentators such as Emmanuel Vivier noted that DeepSeek's low pricing was its core selling point; the hike signals a strategy shift that could significantly raise costs for apps built on its API.
  • The peak/off-peak scheme, likened by Zhidongxi to utility time-of-use pricing, encourages load shifting and reflects peak compute supply pressure, a notable case of fine-grained pricing among Chinese LLM vendors.