DeepSeek’s V4 lineup does not ship time-of-day pricing. Despite the framing attached to the April launch, every published V4 rate is a flat per-token price with a cache-tier split, and the lever that actually moves spend is a fiftyfold-or-greater discount on cached input. The figures in this article come from DeepSeek’s own API documentation and pricing pages, not from an independent audit, and that documentation is the only primary source for them as of 2026-07-07.
Does DeepSeek V4 actually have peak-valley, time-of-day pricing?
DeepSeek’s published V4 rates show no variance by time of day as of 2026-07-07. The Models & Pricing page lists three prices per model: a cache-hit input rate, a cache-miss input rate, and an output rate. There is no peak window, no valley window, no surcharge for daytime hours.
The “peak-valley pricing” framing that trails the launch conflates two different things. DeepSeek does tier its price by where the input token comes from (cache or fresh), and it does cap how many concurrent requests an account can run. Neither of those is a clock. A request at 02:00 UTC costs the same as one at 14:00 UTC on the same model and cache path.
The distinction shapes how teams budget. If inference were genuinely cheaper overnight, the rational move would be to queue batch jobs into a valley window. Under the pricing DeepSeek actually published, the optimization lever is cache reuse and model selection, not cron timing. The rest of this article covers what is documented rather than what was rumored.
What is the July 24, 2026 deadline?
DeepSeek is retiring its legacy deepseek-chat and deepseek-reasoner endpoints on July 24, 2026 at 15:59 UTC, after which they stop responding entirely. According to the V4 release notes, the legacy names currently route through to V4-Flash in its non-thinking and thinking modes, so most callers are already hitting V4 hardware under the old names. The deadline forces the model identifier to change, not necessarily the underlying weights.
The practical migration is smaller than the announcement sounds. Because the legacy identifiers already alias to V4-Flash, a team that switches its model string to the explicit V4 name and re-runs its eval suite has done most of the work. The one trap is mode: teams that leaned on deepseek-reasoner for chain-of-thought work need to confirm they are calling V4-Flash’s thinking mode (or stepping up to V4-Pro) rather than silently landing on a non-thinking path.
What are the V4 models, and how do Pro and Flash differ?
V4 ships two models, launched April 24, 2026 per the release notes. V4-Pro is the larger track: 1.6 trillion total parameters with 49 billion active per token. V4-Flash is the lighter track: 284 billion total parameters with 13 billion active per token. A total-to-active ratio this wide is the mixture-of-experts signature; the divergence is roughly an order of magnitude on Flash and over thirtyfold on Pro. Both default to a 1 million token context length.
The split mirrors the large-and-small pattern other labs have settled on, but DeepSeek’s framing is unusual in one respect: the legacy chat and reasoner identifiers both resolve to Flash, the smaller model. Teams that expected the default path to be the flagship will find V4-Pro is an explicit opt-in, not a fallback.
V4-Pro also carries DeepSeek’s strongest published capability claims. The release notes credit it with open-source state-of-the-art on agentic coding benchmarks and assert it beats all current open models on math, STEM, and coding tasks. Those are vendor-graded claims at medium confidence, with no independent replication cited in the brief as of 2026-07-07; treat the direction (Pro is the stronger model) as settled and the exact margin as unverified.
What does V4 cost, and how does the cache discount work?
V4’s published rates price input tokens at two tiers depending on whether DeepSeek’s context cache already holds the prefix, and the gap between those tiers is where the cost lives.
| Cache-hit input | Cache-miss input | Output | Concurrency | |
|---|---|---|---|---|
| V4-Pro | $0.003625 / 1M | $0.435 / 1M | $0.87 / 1M | 500 |
| V4-Flash | $0.0028 / 1M | $0.14 / 1M | $0.28 / 1M | 2,500 |
Figures per 1 million tokens, all DeepSeek-published, per the Models & Pricing page.
The discount on a cached token is large. On Flash, cache-miss input is $0.14 per 1M and cache-hit input is $0.0028 per 1M, a fiftyfold difference. On Pro it is steeper: $0.435 versus $0.003625 per 1M, roughly 120x. DeepSeek has been explicit that this is deliberate. The change log documents an August 2024 entry describing “context caching on disk technology” that the company said cut prices by “another order of magnitude” on top of prior reductions.
The economics push a specific workload shape. A system prompt, retrieved documents, or a long shared prefix that recurs across requests is cheap once it lands in cache; the first cold read is not. Batch workloads that replay the same large context, such as eval harnesses, document-grounded retrieval, and few-shot pipelines, collect the full discount on every repeat. One-shot interactive prompts over fresh context pay cache-miss rates on essentially every token.
Why do concurrency limits matter as much as the token price?
The concurrency caps are a hard ceiling on parallel work, and they shift which model a workload should use more than the token price does. Pro caps at 500 concurrent requests and Flash caps at 2,500, per the pricing page. Those are ceilings on how much parallel work a single account can sustain, independent of token volume or generation speed.
The cap interacts with model choice in a way the token price alone hides. Pro is roughly 3x the input cost of Flash on a cache miss ($0.435 versus $0.14) and 3x the output cost ($0.87 versus $0.28), but its concurrency ceiling is 5x tighter (500 versus 2,500). A workload that needs throughput more than quality, such as classification, extraction, or bulk summarization, will hit Pro’s concurrency wall long before its token budget strains, which makes Flash the rational default for anything that does not need the larger model’s reasoning. Interactive workloads with a few deep calls favor Pro; fan-out batch work favors Flash. The interesting cases are the ones in between, where the right answer is to shard across both and let model choice follow request shape.
How should an engineering team handle the migration?
The migration is a model-string change, a base_url repoint, and a cache-ordering audit, roughly in that order of risk.
First, find every hardcoded deepseek-chat and deepseek-reasoner string. The legacy names keep working until July 24, 2026, then stop. Config files, environment variables, SDK defaults, and prompt templates are the usual hiding spots, not the primary API client call.
Second, lean on the compatibility surface. DeepSeek’s API accepts both the OpenAI ChatCompletions format and the Anthropic message format, requiring only a base_url and model change to switch a client over, per the API overview. Teams already pointed at an OpenAI- or Anthropic-compatible endpoint can repoint without rewriting their request layer.
Third, design for the cache. The input discount, fiftyfold on Flash and larger on Pro, only fires when a prefix is reused, so the migration is a good moment to audit which parts of the request are stable (system prompts, tool definitions, shared context blocks) versus which vary per call. Stable prefixes should be ordered first in the message array so the cache key covers the maximum span.
How does this compare to OpenAI and Anthropic?
On two documented points, DeepSeek is structurally distinct. First, it folds the cache tier directly into its input price: cache-hit and cache-miss input are separate line items at fiftyfold-plus different rates, per the pricing page. The pattern at other major labs has been a base input rate with prompt caching applied as a discount on top, so the structural difference is where the discount lives in the price sheet, not whether one exists. Second, the API accepts both OpenAI ChatCompletions and Anthropic message formats with only a base_url and model change, per the API overview. A team already pointed at an OpenAI- or Anthropic-compatible client can treat DeepSeek as a second source without rewriting its request layer.
The compatibility point matters more than any per-token comparison, which is just as well. The brief does not contain current, comparable rates from OpenAI or Anthropic, so a direct price-per-token table is not made here and any specific competitor figure would be unsourced.
DeepSeek’s published bet, per the change log, is that the cache, not the clock, is where inference cost gets negotiated, a direction the 2024 disk-caching work set up. What the V4 launch does not do, despite the framing around it, is move the industry toward time-of-day pricing.
Frequently Asked Questions
When should teams migrate versus waiting until the deadline?
Migrating early lets you validate behavior differences and build cache warmth. The legacy endpoints already route to V4-Flash, so outputs should match closely, but subtle mode differences between thinking and non-thinking paths can appear. More importantly, the cache discount only activates after repeated requests on the same prefix, and that warm period takes days of production traffic to reach steady state. A late migration leaves no buffer to discover that your prompt structure prevents cache reuse.
What is the absolute cost floor with perfect caching?
With full cache hits on a 1 million token context, Flash costs $0.0028 per million input tokens and Pro costs $0.003625, plus output fees regardless of cache status. That represents roughly a 25x reduction from typical GPT-4-class pricing at launch. However, hitting this floor requires the entire stable prefix to repeat across requests. Inserting retrieved documents mid-context breaks the cache for everything after the insertion point, which is why most real workloads settle between the floor and the cache-miss rate.
How do I route requests between Pro and Flash in a hybrid setup?
Send classification, extraction, formatting, and translation tasks to Flash. Send multi-step reasoning, code generation, and math problems to Pro. The practical decision boundary is whether your evaluation set shows accuracy gains on Pro that justify the 3x input cost. Most teams run a representative test suite on both models, measure pass rates, and route each request type to the cheapest model that clears their accuracy threshold. V4-Pro is the only option if you need DeepSeek’s claimed agentic coding performance.
Do enterprise contracts increase the concurrency limits?
DeepSeek does not publish concurrency scaling in public documentation. Enterprise customers report higher limits after negotiation, but the uplift varies by use case, commit size, and region. Treat 500 for Pro and 2,500 for Flash as the standard account floor rather than a fixed ceiling. If your workload plan assumes higher parallelism, confirm actual limits with DeepSeek support before launch rather than assuming enterprise tier automatically grants headroom.
What error codes appear after the July 24 deadline?
Requests to the legacy deepseek-chat and deepseek-reasoner endpoints return 404 or 410 status codes after 15:59 UTC on July 24, 2026. DeepSeek has specified this is a hard cutover with no grace period or redirect behavior. Any client code still pinned to the legacy model identifiers breaks immediately at that moment. The migration is mostly a model string change, but verify SDK defaults and infrastructure-as-code templates aren’t silently baking the retired names into deployments.
