Is DeepSeek Cheaper Than Claude for Agent Workloads?
DeepSeek-V4-Pro costs $0.66 per million fresh input tokens and $1.98 per million output tokens off-peak, against $2.00 and $10.00 for Claude Sonnet 5. On output, the line that agent loops bill hardest, that is roughly 5x cheaper.
The headline rate is not the interesting part. DeepSeek is the only major vendor that charges by time of day, and almost nobody explains how that interacts with prompt caching. Get both wrong and the same tokens on the same model cost 60x what they should.
AI Perks tracks $7.7M in credits across 194 companies. This page covers what credits do not: the per-token bill underneath.
The Mechanic Nobody Explains: DeepSeek Charges by Time of Day
DeepSeek peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Everything outside those two windows is off-peak and costs exactly half.
| DeepSeek-V4-Pro-0813 | Off-peak | Peak |
|---|---|---|
| Input, cache hit | $0.022 / 1M | $0.044 / 1M |
| Input, cache miss | $0.66 / 1M | $1.32 / 1M |
| Output | $1.98 / 1M | $3.96 / 1M |
Three things follow from those windows that the pricing page does not spell out.
Peak is only about 21% of the week. Seven hours a day across five weekdays is 35 hours out of 168 (derived). Half price is the default state of the DeepSeek API, not a discount you negotiate.
The peak windows are the Chinese working day. 01:00 to 10:00 UTC is 09:00 to 18:00 in UTC+8 (derived). DeepSeek is pricing its own domestic demand curve, which is why the windows look arbitrary from anywhere else.
A US West Coast team is already off-peak. A 09:00 to 18:00 Pacific workday maps to 16:00 to 01:00 UTC, which sits outside both windows (derived). A European team working 09:00 to 17:00 local spends its entire morning inside the second peak window and pays double.
For batch jobs, evals, nightly reindexing and anything queued, scheduling is free money. Shifting a job out of 06:00-10:00 UTC halves its bill with no code change.
Cache Hits Stack With Off-Peak for a 60x Spread
A cache hit on DeepSeek-V4-Pro input costs $0.022 per million tokens off-peak. A cache miss at peak costs $1.32. Same model, same tokens, 60x apart.
Two multipliers stack, and together they are the whole story of DeepSeek's cost. A cache hit is roughly 30x cheaper than a cache miss on V4-Pro. Off-peak is 2x cheaper than peak.
Agent loops bite hardest here. An agent resends a growing transcript every turn, so by turn twenty the input is mostly repeated context. Cached, the resend is nearly free. Break the prefix with a timestamp or a reordered tool list and every turn pays cache-miss rates, twice over during peak hours.
DeepSeek-V4.1-Flash sits a tier below and is cheaper still: $0.003 per million cached input tokens off-peak, $0.15 per million on a cache miss. DeepSeek has not published an output rate for Flash, so do not budget against a number quoted elsewhere without checking it.
DeepSeek vs Claude: What an Agent Month Actually Costs
On a workload of 50M input tokens at a 90% cache hit rate plus 5M output tokens, DeepSeek-V4-Pro run entirely off-peak lands near $14, Claude Sonnet 5 near $69, and Claude Opus 5 near $173.
| Model | Cached input | Fresh input | Output | Month total |
|---|---|---|---|---|
| DeepSeek-V4-Pro, off-peak | $0.99 | $3.30 | $9.90 | $14.19 |
| DeepSeek-V4-Pro, all peak | $1.98 | $6.60 | $19.80 | $28.38 |
| Claude Sonnet 5 | $9.00 | $10.00 | $50.00 | $69.00 |
| Claude Opus 5 | $22.50 | $25.00 | $125.00 | $172.50 |
All four rows are derived from published list rates. Claude cache reads are estimated at roughly a tenth of the base input rate, and cache-write surcharges are excluded, so the Claude rows are conservative rather than exact. Claude Haiku 4.5 at $1.00 and $5.00 per million sits between the DeepSeek and Sonnet rows.
The spread is real but smaller than the "10x cheaper" line that circulates. Once caching is switched on for both vendors, a well-built agent on Sonnet 5 costs about 5x a DeepSeek agent, not 20x. Badly built, either one costs far more than the table says.
What Changed in DeepSeek's 2026 Lineup
Anything written about DeepSeek before 16 August 2026 quotes a flat rate that no longer exists. The peak and off-peak split replaced single-price billing, so older comparisons describe a model that is gone.
The sequence worth knowing:
- 24 April 2026: V4-Pro and V4-Flash preview, with a 1M-token context window, up to 384K output, and switchable thinking and non-thinking modes (medium confidence).
- 31 July 2026: DeepSeek-V4-Flash-0731 stabilises the Flash line. Last known pre-change pricing was $0.14 per million cache-miss input and $0.28 per million output. Historical context only.
- 13 August 2026: DeepSeek-V4-Pro-0813 reaches general availability as the flagship reasoning and agent model.
- Around 10 September 2026: DeepSeek-V4.1-Flash ships, listed in the docs as DeepSeek-Flash.
V4.1-Flash also quietly retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names. Tutorials and copied code samples still reference them, and they will fail. If a DeepSeek integration broke in September without a deploy, that rename is the first thing to check.
Where Claude Still Wins for Agent and Coding Work
Price per token is one input to the decision, and for agent work it is not the only one that moves the bill.
Anthropic's own 50% discount is the Batch API, which trades latency for half price. That is the direct counterpart to DeepSeek's off-peak window: one vendor discounts patience, the other discounts timing. A team that can queue work takes 50% off either side.
The rest of the case for Claude is surface area rather than tokens: effort controls that trade thinking depth against spend, server-side tools, context editing and compaction for long loops, and a tool-calling contract agent frameworks target first. None of that shows up in a price table, and all of it changes how many turns a task takes.
The honest framing is routing, not loyalty. High-volume classification, summarisation and bulk transformation go to DeepSeek off-peak. Long-horizon agent runs, where a failed trajectory costs more than the tokens, stay on Claude. Most 2026 stacks run both, which is why credits from one provider subsidise experiments on the other. AI Perks makes that second part free.
Neither Vendor Gives Agent Workloads Away
The DeepSeek chat app is free. The DeepSeek API is not, and there is no stated free tier on the official pricing page.
Most comparison articles get this wrong. Claims that new DeepSeek accounts receive free API credits are inconsistently reported and not confirmed on the official pricing page. Budget as though they do not exist.
More importantly for founders: DeepSeek runs no dedicated startup credit program. Cheap per-token pricing is the whole offer. That is a real gap, because most providers it competes against do run one, and a credit grant beats a 5x cheaper token when you are pre-revenue.
| Route to DeepSeek | What it changes | Where to look |
|---|---|---|
| DeepSeek direct API | Cheapest rates, peak and off-peak billing | Official pricing page |
| Third-party hosts | Routing, uptime and region control, higher rates | Compare before committing |
| Provider credit programs | Removes the bill entirely, DeepSeek excluded | AI Perks |
On third-party hosting, Baseten lists DeepSeek V4.1 Flash input at $0.30 per million and carries a standing $30 new-workspace trial credit (verify before budgeting against it). OpenRouter, Together and Fireworks also serve DeepSeek models. Treat that layer as a routing and reliability decision, not a pricing one.
$7.7M in credits across 194 companies is tracked at getaiperks.com, covering the providers that do run a program.
Frequently Asked Questions
Is DeepSeek cheaper than Claude?
Yes, on list price. DeepSeek-V4-Pro off-peak is $0.66 per million fresh input tokens and $1.98 per million output, against $2.00 and $10.00 for Claude Sonnet 5. That is roughly 5x on output. With caching enabled on both sides the practical gap narrows, and free credits close it entirely. See AI Perks.
What are DeepSeek peak hours?
Peak hours are 01:00-04:00 and 06:00-10:00 UTC on weekdays. Calls inside those windows cost double. Everything else, including all weekend traffic, is off-peak at half price. That is about 21% of the week at peak rates (derived), so most workloads are already paying the lower price by default.
Does DeepSeek have a free API tier?
No. The DeepSeek chat app is free to use, but the official pricing page states no API free tier. Reports of free credits for new API accounts are inconsistent and unconfirmed. DeepSeek also runs no startup credit program, unlike most of its competitors, which are tracked at getaiperks.com.
Why did my DeepSeek API calls suddenly fail?
DeepSeek-V4.1-Flash retired the deepseek-v4-flash and deepseek-v4-flash-vision-exp endpoint names around 10 September 2026. Older tutorials and copied code samples still reference them and will fail against the current API with no deploy on your side. Check the model string first, before debugging keys, quotas or networking.
Should I use DeepSeek or Claude for coding agents?
Route by task. High-volume completion, refactors and bulk transformation run cheaply on DeepSeek off-peak. Long-horizon agent runs, where one failed trajectory costs more than the tokens saved, are where Claude's tool-calling and context handling earn their price. Most teams run both and fund the expensive half with credits from AI Perks.
How much can prompt caching actually save on DeepSeek?
A cache hit on V4-Pro input is roughly 30x cheaper than a cache miss. Combined with off-peak billing, a cached off-peak token at $0.022 per million against a peak cache-miss at $1.32 per million is a 60x spread on identical tokens. Caching is the single largest lever on any DeepSeek bill.
Route the cheap work to DeepSeek. Let someone else pay for the rest.