DeepSeek

Key Terms

• DSA (DeepSeek Sparse Attention) - a V4 attention mechanism that uses token-wise compression to handle 1M-token contexts without linearly scaling compute. It is the architectural reason DeepSeek cites for its cost efficiency. Source: DeepSeek – News260424
• MLA (Multi-Head Latent Attention) - DeepSeek's compression technique that shrinks the KV cache versus standard attention, enabling cheaper inference and larger batch sizes. Source: DeepSeek – News260424
• Thinking / Non-Thinking mode - V4 models expose two modes through a single model name. Thinking mode produces extended chain-of-thought reasoning before the answer; non-thinking mode gives direct responses. Source: DeepSeek – Thinking Mode
• Anthropic-compatible API - an endpoint at api.deepseek.com/anthropic that accepts Anthropic-format requests, letting Claude Code and other Anthropic-SDK tools call DeepSeek directly. Source: DeepSeek – Claude Code
• Context Caching (KV Cache) - DeepSeek caches previously processed tokens on disk, so repeated prompts cost less. Cache hits are billed at roughly 1/50 of the cache-miss rate for V4-Pro ($0.003625 vs $0.435 per MTok input). Source: DeepSeek – Pricing
• Peak / valley pricing - a time-of-day price model where a higher rate applies during busy hours and a lower (here, unchanged) rate applies off-peak. DeepSeek's announced version charges 2x during 09:00-12:00 and 14:00-18:00 Beijing Time. As of this report, it is announced but NOT live. Source: DeepSeek – Pricing
• Responses API - OpenAI's newer API format (used by Codex) that combines tool use, reasoning, and structured outputs. DeepSeek-V4-Flash now supports it natively; V4-Pro support is expected early August 2026. Source: DeepSeek – Responses Api ; DeepSeek – Pricing
• Model mapping (Claude Code) - when Claude Code or the Claude Desktop app calls DeepSeek, names starting with claude-opus route to deepseek-v4-pro and names starting with claude-haiku or claude-sonnet route to deepseek-v4-flash. Source: DeepSeek – Claude Code
• DeepSeek Harness - DeepSeek's own in-house agentic coding harness (not yet publicly released), used to generate the official V4-Flash benchmark scores. Source: DeepSeek – Updates

Latest Changes

June watch-item verification

• Official V4 release mid-July with peak/valley pricing: PARTIAL. The V4-Flash official release shipped on July 31 (two weeks later than the mid-July target), but the V4-Pro official release is still pending with the changelog stating "The official release of DeepSeek-V4-Pro will follow soon." Peak/valley pricing did NOT go live. The pricing page footnote now reads: "The DeepSeek API service will soon adopt a peak/off-peak pricing policy... The effective date will be subject to the official announcement." No date given. Source: DeepSeek – Updates ; DeepSeek – Pricing
• deepseek-chat and deepseek-reasoner alias retirement July 24: CONFIRMED. The July 24 15:59 UTC deadline has passed. The pricing page now lists only deepseek-v4-flash and deepseek-v4-pro. Any integration still using the old aliases will have broken. Source: DeepSeek – Pricing ; DeepSeek – News260424
• Vision API availability: STILL PENDING. Image understanding remains chat-app only (chat.deepseek.com). It is not listed in the API pricing or features table, and the Claude Code / Agent SDK path still lacks a vision-enabled model. Source: DeepSeek – Pricing
• US blacklist status: STILL PENDING. No update to the June 17 Reuters report that the US held off adding DeepSeek to its blacklist. The review remains live but no action was taken in July. Source: Reuters – Us Holds Off Blacklisting Chinas Deepseek More Than 100 Firms Deemed Security 2026 06 17

New July developments

• V4-Flash official release (July 31): DeepSeek released the official V4-Flash API in public beta. The model version is DeepSeek-V4-Flash-0731, which keeps the same architecture and size as the preview (284B total / 13B active) but was re-post-trained with significantly enhanced agent capabilities. Benchmark scores "far exceed V4-Pro-Preview": Terminal Bench 2.1: 82.7, NL2Repo: 54.2, Cybergym: 76.7, DeepSWE: 54.4, Toolathlon verified: 70.3, Agent Last Exam: 25.2, Automation Bench (Public): 25.1, DSBench-FullStack: 68.7, DSBench-Hard: 59.6. The API calling method is unchanged (model name deepseek-v4-flash). Source: DeepSeek – Updates
• Responses API and Codex integration (July 31): V4-Flash now natively supports the Responses API format and is specifically adapted for OpenAI Codex (CLI, desktop app, and VS Code extension). A dedicated integration page with a one-click setup script is available. V4-Pro Codex support is expected in early August 2026. Source: DeepSeek – Codex ; DeepSeek – Updates
• V4-Pro official release still pending: The changelog explicitly states "This update only upgrades the DeepSeek-V4-Flash API. The DeepSeek-V4-Pro API and the APP/WEB models are unchanged. The official release of DeepSeek-V4-Pro will follow soon." The pricing page model version still shows "DeepSeek-V4-Pro" without a date suffix, indicating it remains in preview. Source: DeepSeek – Updates ; DeepSeek – Pricing
• Peak/valley pricing not yet live (still announced): The pricing page footnote now reads: "The DeepSeek API service will soon adopt a peak/off-peak pricing policy. During peak hours, prices will be 2x the regular prices, applicable to all billing items. The effective date will be subject to the official announcement." Peak hours confirmed as 9:00-12:00 and 14:00-18:00 Beijing Time (UTC+8) daily. Source: DeepSeek – Pricing
• Price verification (no change): The live pricing page shows identical rates to June: V4-Pro $0.435/$0.87 per MTok (cache-miss input/output) and V4-Flash $0.14/$0.28 per MTok. Cache-hit prices are $0.003625 (Pro) and $0.0028 (Flash) per MTok. Source: DeepSeek – Pricing
• Agent integrations list updated: The Agent Integrations index now lists 6 named harnesses with dedicated documentation pages: Claude Code, Codex, OpenCode, OpenClaw, Hermes, Reasonix. (June listed 16 harnesses in the general index; the dedicated-docs subset has been restructured.) Source: DeepSeek – Claude Code
• Liang Wenfeng investor meeting transcript leaked (July 22-25): A four-hour audio-to-text transcript of founder Liang Wenfeng's investor meeting was leaked on GitHub. It revealed: (1) DeepSeek needs approximately 200,000 Huawei 950 GPUs to train a frontier-scale model but received only 16,000, (2) Liang framed the gap as a resource problem not a talent problem ("The biggest gap between us and the United States lies in resources"), (3) he expected the compute crunch to last at least three years, (4) he floated narrowing the gap with US labs to three to six months using a fraction of their computing. Bloomberg reported on July 25 that DeepSeek paused its fundraising round, reportedly in part due to Liang's frustration over the leak. Source: GitHub – Liang Wenfeng Investor Meeting 2026 7 22 ; Bloomberg – Deepseek Said To Tell Backers Of Funding Pause After Viral Posts ; News – Item
• Gigawatt-scale data center planned (July 30): Bloomberg reported DeepSeek is developing a massive gigawatt-scale AI data center in Inner Mongolia, signaling a push to address the compute constraints Liang described. Source: Bloomberg – Deepseek Is Developing Massive Ai Data Center In Inner Mongolia ; News – Item
• Distillation censorship study (July 30): CTGT published a Show HN demonstrating that distilling DeepSeek V4 Flash into GPT-OSS-120B does not transfer the teacher model's censorship characteristics. The teacher showed a +45.45-point gap on politically sensitive question pairs (7 standard deviations from chance), while the distilled student remained within 1 point of its American base model's behavior. They released the LineageEval framework and open weights. Source: Ctgt – Distillation Censorship Transfer ; News – Item

Plans

DeepSeek is an API-only provider. There are no subscription plans, tiers, or bundled coding agents. Users top up a balance and are billed per token consumed. Source: DeepSeek – Pricing

Concurrency limits:

Higher concurrency is available at no extra cost via a capacity-expansion request. Source: DeepSeek – Rate Limit

Terms explained:

• Concurrency limit - the number of simultaneous in-flight requests an account may have. Exceeding it returns HTTP 429. This is not a per-minute or per-day request cap. Source: DeepSeek – Rate Limit

Note on OpenCode Go (intermediary plan): OpenCode's hosted "Go" plan does not route to DeepSeek's China API. Its page states the plan is designed for international users with models hosted in the US, EU, and Singapore, with a zero-retention policy. This matters for buyers who want DeepSeek weights without sending data to DeepSeek's infrastructure. Source: Opencode – Go

API Pricing

All prices per 1M tokens, verified live on 2026-07-31. Prices unchanged from June.

Notes:

• V4-Pro prices reflect the permanent 75% discount announced 2026-05-22 (original prices were $1.74/$3.48 per MTok).
• V4-Flash model version is now DeepSeek-V4-Flash-0731 (official release, July 31). V4-Pro remains in preview.
• Both models support thinking and non-thinking modes at the same price. Thinking mode may consume more tokens.
• FIM completion is available in non-thinking mode only.
• Deduction rule: expense = tokens x price. Fees come out of topped-up or granted balance, with granted balance consumed first.
• Peak/off-peak pricing is announced but not yet live (see Latest Changes).
• The Responses API currently supports only deepseek-v4-flash. Pro support expected early August 2026.

Source: DeepSeek – Pricing

Model Performance / Benchmarks

DeepSeek's V4 launch page publishes benchmark results as images only, not as machine-readable numbers. However, the July 31 changelog entry for V4-Flash-0731 includes specific numeric scores for agentic benchmarks, tested using the DeepSeek Harness minimal mode (not yet publicly released) with max effort level, topp=0.95, and temperature=1.0. Source: DeepSeek – Updates

V4-Flash-0731 official benchmarks (July 31)

Cross-model comparison

Community consensus on coding quality: V4-Flash has become the primary daily driver for many developers. User f311a: "I've been driving flash model for 90% of my tasks. It's better than pro (for unknown reasons), very cheap and fast. I try to keep changes under 1000 lines and drive architectural decisions myself, barely notice any difference compared to frontier models." News – Item User mordae noticed an immediate quality jump on release day: "It started reasoning more extensively from nowhere and precision went up a lot. It also changed its prose style for the better." News – Item

Latest News

V4-Flash Official Release (July 31, 2026)

DeepSeek released the official V4-Flash API in public beta as DeepSeek-V4-Flash-0731. The model keeps the same architecture and size as the preview (284B total / 13B active) but was re-post-trained with significantly enhanced agent capabilities. Benchmark results far exceed V4-Pro-Preview across nine agentic benchmarks. The model natively supports the Responses API format and is adapted for OpenAI Codex. V4-Pro is unchanged and its official release is still pending. Source: DeepSeek – Updates

Codex Integration Launched (July 31, 2026)

DeepSeek published a dedicated Codex integration page with a one-click setup script. Only deepseek-v4-flash currently supports Codex. The integration covers Codex CLI, the ChatGPT desktop app, and the Codex IDE extension for VS Code. deepseek-v4-pro is expected to support Codex in early August 2026. Source: DeepSeek – Codex

Liang Wenfeng Investor Meeting Transcript Leaked (July 22-25, 2026)

A four-hour audio-to-text transcript of founder Liang Wenfeng's investor meeting was leaked on GitHub. It revealed a severe compute constraint: Liang stated he needs approximately 200,000 Huawei 950 GPUs to train a frontier-scale model but received only 16,000, expects the crunch to last at least three years, and framed the gap as purely resource-driven ("The biggest gap between us and the United States lies in resources"). Bloomberg reported on July 25 that DeepSeek paused its fundraising round, reportedly in part due to Liang's frustration over the leak. Sources: GitHub – Liang Wenfeng Investor Meeting 2026 7 22 ; Bloomberg – Deepseek Said To Tell Backers Of Funding Pause After Viral Posts ; Fortune – Deepseek Liang Wenfeng Backers Fundraising Pause Viral Posts Investors

DeepSeek Plans Gigawatt-Scale Data Center in Inner Mongolia (July 30, 2026)

Bloomberg reported DeepSeek is developing a massive gigawatt-scale AI data center in Inner Mongolia, a direct response to the compute constraints described in the leaked investor meeting. Source: Bloomberg – Deepseek Is Developing Massive Ai Data Center In Inner Mongolia

Distillation Does Not Transfer Censorship (July 30, 2026)

CTGT published research using DeepSeek V4 Flash as a teacher model for distillation into GPT-OSS-120B. The teacher showed a +45.45-point gap on politically sensitive question pairs (7 standard deviations from chance), but the distilled student model's behavior remained within 1 point of its American base model. They released the LineageEval evaluation framework and 20B open weights. Source: Ctgt – Distillation Censorship Transfer ; News – Item

Alias Retirement Completed (July 24, 2026)

The deepseek-chat and deepseek-reasoner model names were retired as scheduled on 2026-07-24 15:59 UTC. The pricing page now lists only deepseek-v4-flash and deepseek-v4-pro. Source: DeepSeek – Pricing ; DeepSeek – News260424

Community Signals

V4-Flash release: excitement about quality leap at ultra-low cost (HN, July 31)

The "DeepSeek-V4-Flash Update" thread (204 points, 84 comments) drew immediate positive reactions. User lionkor shared concrete cost-in-practice data: "I use deepseek for a lot of my personal day-to-day agent needs... last 30 days: Cost: $4.55USD, API requests: 3,467, Tokens: 323,183,886. And as an engineer who leads a small team, I have very high standards for quality... It has not disappointed at all for coding or review tasks." News – Item User kmarc described a multi-model workflow: "Essentially I'm running everything on flash now inside pi. With the correct set of MCP servers, context reducer tooling and skills it can implement any task I throw at it. Some sessions take 30+ turns, but it's fast and cheap; all this in an hour, with ~$0.5 cost." News – Item User mordae noticed the quality change in real time: "I was just using it when it landed. It started reasoning more extensively from nowhere and precision went up a lot. It also changed its prose style for the better." News – Item User baalimago anticipated the V4-Pro release: "Looks like they will release an updated version of deepseek-v4-pro soon, which most likely will beat kimi k3 at both intelligence and cost (judging by the vast improvements to dsv4 flash)." News – Item

Benchmark comparison: Flash trades blows with GPT-5.6 Terra (HN, July 31)

User dnhkng compiled a side-by-side comparison of V4-Flash-0731 against GPT-5.6 Terra: "DeepSeek V4 Flash (Preview -> 0731): Terminal Bench: 56.9 -> 82.7 (+25.8), Toolathlon: 51.8 -> 70.3 (+18.5). Compared to GPT-5.6 Terra: Terminal Bench: Flash 82.7 vs Terra 78.4, Toolathlon: Flash 70.3 vs Terra 53.1, DeepSWE: Flash 54.4 vs Terra 69.6, Agents' Last Exam: Flash 25.2 vs Terra 50.4. Trading blows with Terra, which is pretty interesting." News – Item User villish corrected the Terra Terminal Bench score to 87.4: "Terra 87.4" per OpenAI's own page. News – Item User Iolaum flagged the harness caveat: "Since they did this with their own harness I'm not sure it's apples to apples comparison." News – Item

Pricing comparison: cache-read cost remains orders of magnitude cheaper (HN, July 31)

User dudisubekti put the caching advantage in numbers: "Gpt 5.6 Luna cache read is $0.02 per mtok. V4 flash cache read is $0.0028 per mtok. That's not 'a bit cheaper', just saying." News – Item User nickandbro noted the vision gap as a trade-off: "DeepSeek v4 flash even though a bit cheaper, does not have vision abilities, which is a big draw for agentic tasks." News – Item User minraws confirmed peak/valley pricing did not go live: "They haven't raised prices though the plan was to increase it with peak hour usage for V4 Pro GA release, they didn't do that, so no price increases there I believe." News – Item

Liang Wenfeng investor meeting leak: compute gap and fundraising pause (HN, July 25)

The "DeepSeek pause fundraise after comments on compute gap to US leaked" thread (251 points, 200 comments) was one of the biggest DeepSeek stories of July. User credit_guy quoted key passages from the transcript: "There is certainly no shortage of funds or resources... Within our financial capacity, it's undoubtedly true that the more cards are always better... In reality, spending such a large sum is no easy task: you can't obtain enough cards, they're hard to come by... our only concern is whether we can obtain enough cards." News – Item User SyneRyder quoted the compute-gap framing: "The biggest gap between us and the United States lies in resources." Liang reportedly said he needs 200,000 Huawei 950 chips but received 16,000, and expects the crunch to last at least three years. News – Item User PeterHolzwarth summarized from a secondary report: "He also floated narrowing the gap with US labs to three to six months using a fraction of their computing." News – Item User credit_guy clarified the fundraising pause was about the leak, not the compute gap: "he's pausing because of the leak. the transcript content has nothing to do with it." News – Item

Distillation censorship study: DeepSeek censorship does not transfer (HN, July 30)

The "Show HN: Distilling DeepSeek into GPT-OSS doesn't transfer censorship" thread (123 points, 64 comments) provided a data point for buyers concerned about using Chinese-origin models. User cgorlla (the author) summarized: "The teacher's gap on the core political set of pairs was +45.45 points, ~7 standard deviations from chance, and every distilled student was within 1 point of its base." News – Item

Enterprise Readiness

Source: DeepSeek – Rate Limit ; DeepSeek

Transparency Gaps