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/anthropicthat 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. Source: Kucoin – Deepseek V4 Launches In Mid July With Peak Valley Pricing
- Model mapping (Claude Code) - when Claude Code or the Claude Desktop app calls DeepSeek, names starting with
claude-opusroute todeepseek-v4-proand names starting withclaude-haikuorclaude-sonnetroute todeepseek-v4-flash. Source: DeepSeek – Claude Code
Latest Changes
Changes since the 2026-05 report.
- Pricing strategy change (announced, not yet live): On 2026-06-29 DeepSeek emailed developers and put a banner on the API platform stating that the official (non-preview) V4 release, due mid-July, will use peak/valley pricing: 2x the current rate during 09:00-12:00 and 14:00-18:00 Beijing Time (UTC+8), with the off-peak price unchanged. The public pricing page still shows flat rates as of this report. Source: Kucoin – Deepseek V4 Launches In Mid July With Peak Valley Pricing ; user confirmation at News – Item
- Price verification (no change): The live pricing page shows the same permanent-discount prices first announced 2026-05-22: 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
- New capability (consumer only): On 2026-06-18 DeepSeek shipped "Vision" on the chat app (chat.deepseek.com), letting the model understand the contents of an image. It is not an image generator and, as of this report, is not exposed on the API, which blocks full Claude Code / Agent SDK use because those SDKs require a vision-enabled model. Source: News – Item ; News – Item
- Deprecation timeline (unchanged, imminent): The
deepseek-chatanddeepseek-reasonermodel aliases will be fully retired on 2026-07-24 15:59 UTC. They currently route to V4-Flash non-thinking and thinking modes. The deadline is now 24 days away at report time. Source: DeepSeek – Pricing ; DeepSeek – Updates - Forward release: The non-preview ("official") V4 release is scheduled for mid-July, per the same announcement that introduced peak/valley pricing. Source: News – Item
- Regulatory event (status quo preserved): On 2026-06-17 Reuters reported the US government held off adding DeepSeek to a blacklist of more than 100 firms deemed security risks. Access was not disrupted, but the review keeps geopolitical and supply-chain risk live. Source: Reuters – Us Holds Off Blacklisting Chinas Deepseek More Than 100 Firms Deemed Security 2026 06 17
- Harness integrations expanded: DeepSeek's Agent Integrations index now lists 16 supported harnesses (Claude Code, GitHub Copilot, GitHub Copilot CLI, Kilo Code, WorkBuddy/CodeBuddy, OpenCode, Oh My Pi, OpenClaw, AstrBot, Deep Code, Hermes, nanobot, Crush, Pi, Reasonix, Langcli), up from the handful named in the May report (OpenCode, Claude Code, Pi, KiloCode). Source: DeepSeek – Claude Code
- Feature added (Claude Code integration): The DeepSeek API now natively supports the Web Search tool inside Claude Code, and documents a model-mapping scheme that also lets the Claude Desktop app bypass its model-name restrictions by repointing base_url and api_key. Source: DeepSeek – Claude Code
- Independent benchmark (contested): On 2026-06-07 RuntimeWire scored DeepSeek V4 Pro at 38.0 versus GPT-5.5 Pro at 33.0 across four on-the-fly precision tasks judged by grok-4-1-fast-non-reasoning. The methodology (single run, AI judge, four tasks) drew heavy criticism on HN, and a separate community benchmark (deepswe.datacurve.ai) ranked GPT-5.5 far higher. Source: Runtimewire – Deepseek V4 Pro Beats Gpt 5 5 Pro On Precision ; News – Item
- Compute-supply signal: On 2026-06-17 Tom's Hardware reported a Huawei-led team post-trained DeepSeek's 1.6T-parameter V4 Pro on roughly 1,000 Ascend 910C chips, a data point on DeepSeek's non-Nvidia training path. Source: Tomshardware – Huawei Led Team Claims It Post Trained Deepseeks 1 6 Trillion Parameter Models On Ascend 910C Chips
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
| Plan | Price | Included Usage | Key Details |
|---|---|---|---|
| API (pay-per-token) | Pay as you go | No usage caps; concurrency limits apply (see below) | Balance topped up manually; no monthly commitment |
Concurrency limits:
| Model | Concurrency Limit |
|---|---|
| deepseek-v4-pro | 500 |
| deepseek-v4-flash | 2,500 |
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 own hosted "Go" plan (opencode.ai/go) does not route to DeepSeek's China API. Its page states the plan "is designed primarily for international users, with models hosted in the US, EU, and Singapore for stable global access" and that "our providers follow a zero-retention policy and do not use your data for model training" (quoted via HN discussion). This matters for buyers who want DeepSeek weights without sending data to DeepSeek's own infrastructure. Source: News – Item
API Pricing
All prices per 1M tokens, verified live on 2026-06-30.
| Model | Input (Cache Hit) | Input (Cache Miss) | Output | Context | Max Output |
|---|---|---|---|---|---|
| deepseek-v4-flash | $0.0028 | $0.14 | $0.28 | 1M | 384K |
| deepseek-v4-pro | $0.003625 | $0.435 | $0.87 | 1M | 384K |
Notes:
- V4-Pro prices reflect the permanent 75% discount announced 2026-05-22. The original prices ($1.74/$3.48 per MTok) are shown with strikethrough on the pricing page and are no longer charged.
- Cache-hit prices were reduced to 1/10 of the launch price effective 2026-04-26.
- 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.
- Prices may vary; DeepSeek reserves the right to adjust them (see the announced peak/valley change in Latest Changes).
Source: DeepSeek – Pricing
Model Performance / Benchmarks
DeepSeek's V4 launch page publishes benchmark results as images only, not as machine-readable numbers, so exact scores remain undisclosed. The full details are in the tech report PDF. Source: Hugging Face – Deepseek V4.Pdf
| Benchmark / Metric | Finding | Notes |
|---|---|---|
| Agentic Coding (vendor claim) | "Open-source SOTA" for V4-Pro | Exact score undisclosed in text. Source: DeepSeek – News260424 |
| NIST CAISI (May 2026) | "On par with GPT-5" | US NIST evaluation. Source: Nist – Caisi Evaluation Deepseek V4 Pro |
| RuntimeWire precision head-to-head (June 7) | V4 Pro 38.0 vs GPT-5.5 Pro 33.0 | 4 tasks, single run, judged by grok-4-1-fast-non-reasoning. Heavily criticized as statistically weak. Source: Runtimewire – Deepseek V4 Pro Beats Gpt 5 5 Pro On Precision |
| DeepSWE (community counter-benchmark) | GPT-5.5 well above DeepSeek | Commenter cited GPT-5.5 near 70% and DeepSeek near 8% on the deepswe.datacurve.ai benchmark, the opposite picture from RuntimeWire. Source: News – Item |
Community consensus on coding quality is consistent with May: V4-Pro is competitive with Sonnet-tier models for everyday coding at a fraction of the cost, but trails top frontier models on hard problems. User KronisLV: "for overall software development, DeepSeek V4 Pro (Max reasoning) is pretty fast and generally okay ... I'd compare it with maybe Sonnet." News – Item User rurban: "Precision yes, but depth of thinking not. I can use DeepSeek V4 Pro 90% of my time, but for very tricky problems I have to use GPT or Claude models." News – Item
Latest News
Peak/Valley Pricing Announced for Mid-July V4 Release (June 29, 2026)
A crypto-news flash reported that DeepSeek's official (non-preview) V4 launch in mid-July will introduce peak/valley pricing, charging 2x during 09:00-12:00 and 14:00-18:00 Beijing Time (UTC+8) and leaving the off-peak price unchanged. HN users confirmed receiving an email and seeing a banner on the API platform. The public pricing docs page had not been updated as of this report. Sources: Kucoin – Deepseek V4 Launches In Mid July With Peak Valley Pricing , News – Item , News – Item
DeepSeek Vision Ships on Chat App, Not API (June 18, 2026)
DeepSeek added image understanding ("Vision") to its consumer chat app, letting the model describe what is in a photo. It is not an image generator and is not yet on the API, which blocks agentic use because the Claude Code / Agent SDK path requires a vision-enabled model. Source: News – Item
US Holds Off Blacklisting DeepSeek (June 17, 2026)
Reuters reported the US government held off adding DeepSeek to a blacklist of more than 100 firms deemed security risks. API access was not disrupted, but the review leaves export-control and data-residency risk open for buyers. Source: Reuters – Us Holds Off Blacklisting Chinas Deepseek More Than 100 Firms Deemed Security 2026 06 17
Huawei-Led Team Post-Trained V4 Pro on Ascend 910C (June 17, 2026)
Tom's Hardware reported a Huawei-led team post-trained DeepSeek's 1.6T-parameter V4 Pro on roughly 1,000 Ascend 910C chips, evidence of DeepSeek's non-Nvidia compute path. Source: Tomshardware – Huawei Led Team Claims It Post Trained Deepseeks 1 6 Trillion Parameter Models On Ascend 910C Chips
Independent Precision Head-to-Head (June 7, 2026)
RuntimeWire ran four on-the-fly precision tasks and scored DeepSeek V4 Pro 38.0 to GPT-5.5 Pro's 33.0, judged by grok-4-1-fast-non-reasoning. The single-run, AI-judged methodology drew strong criticism, and a separate community benchmark painted the opposite ranking. Sources: Runtimewire – Deepseek V4 Pro Beats Gpt 5 5 Pro On Precision , News – Item
Reminder: Alias Retirement on July 24, 2026
The deepseek-chat and deepseek-reasoner model names will stop working on 2026-07-24 15:59 UTC. They currently route to V4-Flash non-thinking and thinking modes. Any integration still using the old names will break. Source: DeepSeek – Updates
Community Signals
Peak/valley pricing: first retreat from flat ultra-low pricing (HN, June 29)
The "DeepSeek V4 Peak Valley Pricing Change" thread (54 points, 33 comments) worked out the timezone impact. The clearest technical read came from user bel8, who mapped the 2x windows into Western timezones and noted that intermediaries may shield end users: "So DeepSeek v4 will cost 2x between 9:00 to 12:00 and 14:00 to 18:00, Beijing Time ... Subscription plans like OpenCode Go might not even pass down this 2x cost since most customers are EU/US." News – Item User gruez clarified the framing: "The off-peak price is still the same price as today. In other words they hiked the price 2x for peak hours only." News – Item User phainopepla2 confirmed the source of the news: "I got an email from them with this information approximately six hours ago. There is also a banner about the price change on their API platform site." News – Item
US-blacklisting near-miss: access preserved, geopolitical risk debated (HN, June 17)
The Reuters story ("US holds off blacklisting DeepSeek", 537 points, 603 comments) split between developers who plan to keep using DeepSeek and those who flag the data risk. User apatheticonxon, who codes with it daily, summed up the value proposition: "I use DeepSeek every day (via VSCode Insiders and Zed Editor). It's very affordable and, while it's slightly behind Claude (not sure how far behind Fable), it suits my working style well." News – Item User l5870uoo9y put the price gap in numbers: "it is impressive by just how much DeepSeek undercuts OpenAI and Claude. DeepSeek charges $0.87 per million output [tokens]." News – Item
Cost-in-practice: dollars where OpenAI spends hundreds (HN, June 8)
On the "DeepSeek V4 Pro beats GPT-5.5 Pro on precision" thread (397 points, 225 comments), user SwellJoe gave a concrete cost comparison from a vulnerability-scanning benchmark: "I tried adding GPT 5.5 Pro to a vulnerability scanning benchmark I made ... and it blew through the $100 budget limit halfway through. DeepSeek V4 Pro cost about a dollar for the whole benchmark. GPT Pro cost an average of $22 per case." News – Item User psadauskas reported switching to OpenCode Go's $10/mo plan after Anthropic "banned Opencode", finding parity for daily work: "I've been using Deepseek through Opencode Go on the $10/mo plan, and I honestly can't really tell much difference." News – Item
Vision gap: API absence blocks Claude Code / Agent SDK (HN, June 18)
The "DeepSeek Introduces Vision" thread (498 points, 204 comments) made clear the new feature is consumer-only. User tornikeo explained why it matters for agentic coding: "I really need this as an API. Turns out, to use Claude Agents SDK, you need to have a vision enabled API. If Deepseek API could see, it can fully drive Claude Code and Claude Agents SDK." News – Item User jiehong clarified the scope: "this allows Deepseek to understand a picture ... but this is not an image generation system ... it's only on the chat web, not the API." News – Item
Benchmark skepticism (HN, June 8)
Several commenters rejected the RuntimeWire precision result as too thin to be meaningful. User hit8run pointed to a conflicting benchmark: "This benchmark draws a very different picture having GPT5.5 on the very top with 70% and DeepSeek at 8% Deepswe" News – Item User shenberg also argued the AI judge got the first task backwards on email-regex boundaries. News – Item
Enterprise Readiness
| Feature | Available? | Details |
|---|---|---|
| SSO (SAML/OIDC) | No | Not mentioned in any documentation. API uses key-based auth only. Source: DeepSeek |
| SCIM | No | No user-provisioning features. |
| Audit logs | No | No audit logging mentioned. |
| IP indemnity | No | No IP indemnity or copyright commitment. |
| Data residency | No | Data processed in China. No regional options. Subject to China's National Intelligence Law. |
| HIPAA | No | No HIPAA compliance mentioned. |
| Air-gapped / On-prem | Partial | Models are open-weight; self-hosting is possible but V4-Pro is 1.6T params and needs major GPU resources. Source: Hugging Face – Deepseek V4 Pro |
| SLA | No | No SLA documented. |
| Admin controls (RBAC) | No | No admin controls. Per-account API keys only. The user_id parameter gives basic multi-user isolation. Source: DeepSeek – Rate Limit |
| Vision (multimodal) | No (API) | Image understanding shipped June 18 on the chat app only, not the API. Source: News – Item |
Source: DeepSeek – Rate Limit ; DeepSeek
Transparency Gaps
| Gap | Details | Severity |
|---|---|---|
| Peak/valley pricing not on docs page | The 2x peak surcharge for mid-July was announced by email and a platform banner (June 29) but is absent from the public pricing page as of this report. Buyers cannot yet read the official windows or rates from the docs. | High |
| Official V4 release date vague | "Mid-July" is the only stated window for the non-preview release. No exact date. | Medium |
| No training opt-out | DeepSeek's API terms still provide no way to opt out of training on submitted data. OpenRouter with ZDR, or OpenCode Go's zero-retention hosted plan, are workarounds, not DeepSeek features. Source: News – Item | High |
| Vision not on API | Image understanding is chat-app-only with no stated API roadmap, which blocks Claude Code / Agent SDK flows that need vision. Source: News – Item | Medium |
| Exact benchmark scores unpublished | V4-Pro's launch page shows benchmarks as images only; no machine-readable numbers. The tech report PDF holds detail. | Medium |
| No RPM/TPM limits published | Only concurrency limits (500 Pro, 2,500 Flash) are documented. There are no published requests-per-minute or tokens-per-minute caps. | Medium |
| Pricing sustainability undisclosed | The 75% discount was made permanent in May and a 2x peak surcharge is coming in July. DeepSeek has not said whether current rates sit above or below inference cost. | Medium |
| Data retention vague | The privacy policy references "Parties with Other Legal Rights" as a data-recipient category, which community members flag as concerning under China's National Intelligence Law. Source: Cdn – Deepseek Privacy Policy | High |
| Blacklist-review status unstable | The US held off blacklisting DeepSeek on June 17, but the review is ongoing and the criteria are not public, so future access is not guaranteed. Source: Reuters – Us Holds Off Blacklisting Chinas Deepseek More Than 100 Firms Deemed Security 2026 06 17 | Medium |