Contains affiliate links. We may earn a commission — it never changes the verdict. Details
The Stack / Chat Assistants
DeepSeek
02 DecisionsYes with eyes open, and don't architect around the price. DeepSeek's chat is still free, and the cost floor moved twice this quarter: August's price hike more than doubled off-peak output pricing, and September 10 partly reversed it — V4.1-Flash cut off-peak input to $0.15/M and output to $0.60/M, with cached input down to $0.003/M. DeepSeek is still cheap next to the frontier labs, but a floor that moves twice in five weeks is not a floor — it's a variable you price into any workflow that depends on it. Two trust boundaries: the data question (a China-based service, so route work by sensitivity) and the cost question (rates that change by the month and by the hour).
The job in the stack
The cost-floor seat: a genuinely free consumer assistant and the cheapest credible API for volume workloads.
Pricing
| Plan | Price | Notes |
|---|---|---|
| Chat | $0 | Fully free — no paid consumer tier exists |
| API (V4.1-Flash) | $0.15/M input (off-peak) | $0.60/M output off-peak; cache-hit input $0.003/M. Peak doubles every rate: $0.30/M input, $1.20/M output, $0.006/M cache-hit. API id is now deepseek-flash, with vision, tool calls and thinking mode built in (thinking on by default). The old deepseek-v4-flash and vision-exp names still work and bill at Flash rates. Effective Sept 10, 2026, a cut from the Aug 16 rates ($0.22/$0.66/$0.007) |
| API (V4-Pro) | $0.66/M input (off-peak) | $1.98/M output off-peak; cache-hit input $0.022/M. Peak doubles to $1.32/$3.96/$0.044. DeepSeek's pricing page confirms peak is 01:00-04:00 and 06:00-10:00 UTC, Monday to Friday, excluding Chinese public holidays. Weekends and holidays are off-peak. No vision. Effective Aug 16, 2026, up from the prior $0.435/$0.87 |
| API Search & Fetch | $0.75 per 1,000 requests | Successful web search or page-fetch calls. Calls through DeepSeek Harness, Claude Code or Codex are free until per-request billing starts Nov 1, 2026 (Beijing time) |
Pricing verified 2026-10-24. Plans change — the source of truth is DeepSeek's own pricing page.
Models & versions
- DeepSeek V4.1-Flash Fast tier (API id deepseek-flash). 552B-parameter MoE with native vision, 1M-token context and up to 384K output. Replaced V4-Flash and absorbed the separate Vision-Exp model on Sept 10, 2026, at a lower rate card. The live snapshot is now V4.1-Flash-1017 (Oct 17), at unchanged rates
- DeepSeek V4-Pro Flagship reasoning tier (API id deepseek-v4-pro, currently the 0813 snapshot). 1M context and up to 384K output, text only. DeepSeek planned to route all V4-Pro traffic to V4.1-Flash on Sept 14, 2026, then reversed that after user feedback. V4-Pro stays live at unchanged pricing
The honest read
What works
- + Consumer chat is entirely free — no caps ladder, no upsell maze
- + API pricing an order of magnitude below Western frontier rates
- + Aggressive cache discounts reward well-structured prompting
- + Open-weight releases keep self-hosting on the table
What breaks
- − Data residency and governance: a China-based service is a real trust-boundary call for business data
- − Ecosystem and tooling depth trail OpenAI/Anthropic
- − Content policies differ — some topics return notably filtered answers
- − The Aug 6 price-increase warning landed for real on Aug 16 — off-peak output pricing more than doubled, and peak-window hours double it again
- − DeepSeek nearly force-migrated V4-Pro traffic onto V4.1-Flash with four days' notice, then reversed it after pushback — a platform that changes its rate card and its model roadmap this fast needs a swap-out plan, not trust
Use it for
- →High-volume API workloads where token math dominates
- →Cost-sensitive experimentation and prototyping
- →A capable free assistant for non-sensitive personal use
Skip it if
- ×Client or regulated data is involved — keep that inside your vetted seats
- ×You need the strongest reasoning/writing available — pay for the frontier
Questions operators ask
?DeepSeek's prices went up. Should I still build on it?
Build on it, don't marry it. The increase DeepSeek warned about on August 6 landed August 16 — output pricing more than doubled off-peak and doubles again in peak hours — and it's still cheap next to the frontier labs. The operator move is the same one that applies to any commodity seat: keep your prompts and routing behind an abstraction so swapping the model underneath is a config change, not a rewrite, and batch volume work into off-peak hours. If DeepSeek's price is the only reason a workflow is viable, that workflow was already fragile — this is the proof.
?Is DeepSeek safe for business use?
Define safe. The models are capable and the price is unbeatable — but data flows to a China-based service, which is disqualifying for client, health, financial, or regulated data in many operations. Use it where the data wouldn't hurt you published; keep everything else in your governed seats.
What changed
- Freshness sweep: API rates hold for both tiers. deepseek-flash now serves the V4.1-Flash-1017 snapshot, and thinking mode is on by default. New on the pricing page: Search & Fetch tool calls at $0.75 per 1,000 successful requests. Billing starts Nov 1, 2026, and calls via Harness, Claude Code or Codex are free until then. Added as a pricing row. Chat stays free.
- Freshness sweep: rates unchanged. Two fixes against DeepSeek's pricing page. (1) Our peak-window note said 'unconfirmed at source.' The docs now confirm peak is Mon-Fri 01:00-04:00 and 06:00-10:00 UTC, excluding Chinese public holidays, so weekends and holidays are off-peak. (2) V4-Flash-Vision-Exp is no longer listed. Vision is built into deepseek-flash (V4.1-Flash), so we removed the separate entry. V4-Pro now runs the 0813 snapshot.
- NEWS SCAN: DeepSeek released V4.1-Flash (Sept 10, 2026, 04:00 UTC), a 552B-parameter MoE model with native vision and the same 1M-token context window, replacing V4-Flash on a cheaper rate card — off-peak input cut from $0.22/M to $0.15/M, output from $0.66/M to $0.60/M, cached input from $0.007/M to $0.003/M (peak still exactly doubles every figure). Separately, DeepSeek had announced routing all V4-Pro traffic onto V4.1-Flash starting Sept 14, then reversed that plan after user feedback — V4-Pro remains available at its existing Aug 16 rates. Updated pricing table, model list, verdict, and cons; added a con about the near-miss forced migration. V4-Flash-Vision-Exp still runs on the prior (pre-cut) rate card pending its own update — not verified as changed this run.
- NEWS SCAN: the Aug 6 price-increase warning is now confirmed with real numbers. DeepSeek's peak/off-peak scheme took effect Aug 16, 2026: V4-Flash off-peak is $0.22/M input (cache-hit $0.007/M) and $0.66/M output — up from the flat $0.14/$0.28 that held through July. Peak hours (01:00-04:00 and 06:00-10:00 UTC) double every figure again. V4-Pro moved the same way: off-peak $0.66/M input, $1.98/M output, up from $0.435/$0.87; peak doubles that too. Separately, DeepSeek shipped V4-Flash-Vision-Exp on Aug 21 — a multimodal Flash variant on the same rate card. Verdict and cons rewritten: the price floor moved, it's not a rumor anymore.
- News scan follow-up to the August 6 announcement. The increase DeepSeek pre-announced has taken a shape: time-based pricing, with rates rising during two daily peak windows (01:00-04:00 and 06:00-10:00 UTC) and a lower rate every other hour. DeepSeek's own rate card has documented those windows for weeks while marking the surcharge inactive. Multiple secondary sources report it going live around August 16 with a higher base rate as well, but they disagree with each other and with DeepSeek's published card on the actual numbers, so our listed prices stay at the last figures we could verify at source: $0.14/M in and $0.28/M out on V4-Flash, $0.435/$0.87 on V4-Pro. Flagged to David for a manual check on platform.deepseek.com, which needs a login we do not have. The planning point stands regardless of the final figures: DeepSeek's price is now a function of what hour your jobs run, so batch work belongs off-peak.
- Material change. DeepSeek published a notice in its API docs stating it plans to raise overall API pricing in the near future, with a significant increase expected — reported the same day by Bloomberg, TechNode, and others. Current rates are unchanged as of today ($0.14/$0.28 Flash, $0.435/$0.87 Pro), and no new rate card or effective date has been published. Verdict and cons updated: treat the cost floor as temporary and keep a migration path. Reports of a peak/off-peak 2x policy circulated in July but appear nowhere in the official docs and are not treated as fact here.
- Re-verified. The V4-Pro 75%-off promo (was set to expire May 31, 2026) was made permanent — $0.435/M input, $0.87/M output is now the standing list price, not a promo rate. Note for API users: legacy model names deepseek-chat/deepseek-reasoner deprecate July 24, 2026 — new integrations should target V4-Flash/V4-Pro directly.
- Added to the stack (cost-floor seat). Verified: chat free, V4-Flash API $0.14/$0.28 per M with 98% cache discount, V4-Pro promo ~$0.44/$0.87.
Give DeepSeek a job
The cost-floor seat: a genuinely free consumer assistant and the cheapest credible API for volume workloads.
Try DeepSeek →Alternatives with different jobs
Claude
02 DecisionsThe thinking seat: long documents, hard reasoning, code, and anything where the words have to be right the first time.
ChatGPT
02 DecisionsThe generalist seat: quick answers, brainstorming, voice mode, and the widest plugin/app ecosystem when you need breadth over depth.
Google Gemini
02 DecisionsThe Google-side seat: massive context windows, Deep Research reports, and native reach into Gmail, Docs, and Drive.