DeepSeek API Pricing and Cost Comparison With OpenAI

DeepSeek's API pricing is genuinely low. But the number a team will actually pay depends on three things a sticker price never shows: what time of day the calls run, how often the cache gets hit, and which model version is behind the API call. That variability is invisible on a pricing page screenshot.

The pricing itself keeps moving, which makes the problem bigger than it sounds. A registry maintained by benchlm.ai, updated September 25, 2026, lists at least seven distinct DeepSeek rate snapshots, and they stretch from January 2025 through September 2026. Several of those older snapshots still show up in guides published earlier this year, and teams are citing them as if they're current.

That creates a real budgeting hazard. Rates from the April 2026 V4 Pro launch, the May 2026 promotional cut, the July 2026 flat-rate period, and the August 2026 peak/off-peak introduction are each meaningfully different from what's billed since September 10, 2026. A team pulling numbers from a cached guide or a bookmarked blog post is working from the wrong numbers, possibly by a wide margin.

This piece lays out the rates actually in force right now, compares them against OpenAI's current lineup tier by tier, and walks through the structural factors, cache behavior, time-of-day billing, and model churn, that determine what a production workload really costs.

Current DeepSeek API rates, exactly as they stand in October 2026

As of October 2026, DeepSeek's hosted API runs two active models, each with two rate periods depending on time of day.

DeepSeek V4 Pro, in force since August 16, 2026, costs $1.32 per million input tokens and $3.96 per million output tokens at peak. Off-peak, those rates drop to $0.66 and $1.98. DeepSeek V4.1 Flash, served under the model name deepseek-flash since September 10, 2026, costs $0.30 per million input tokens and $1.20 per million output tokens at peak, and half that, $0.15 and $0.60, off-peak.

| Model | Input (peak) | Input (off-peak) | Output (peak) | Output (off-peak) | |---|---|---|---|---| | DeepSeek V4 Pro | $1.32/M | $0.66/M | $3.96/M | $1.98/M | | DeepSeek V4.1 Flash | $0.30/M | $0.15/M | $1.20/M | $0.60/M |

Peak hours run 01:00 to 04:00 UTC and 06:00 to 10:00 UTC, Monday through Friday, but they don't apply on Chinese public holidays. The practical effect is that teams need to map their own traffic against UTC, not assume "peak" means business hours anywhere in particular.

Cache state is the other lever on the bill: it is the biggest one. On deepseek-flash, a cache hit on input tokens costs roughly 50 times less than a cache miss at peak rates. Any workload that resends a stable system prompt or a repeated document context, an agent framework, a RAG pipeline answering questions against the same knowledge base, picks up most of that discount automatically, without any special configuration. The discount only applies if the prefix actually lands in DeepSeek's cache, and DeepSeek clears unused cache entries after a window of hours to a few days. A workload with gaps in traffic can lose its cached prefixes and quietly start paying full price again.

DeepSeek lists no Batch API discount, which teams used to OpenAI's structure should flag. That's a feature OpenAI customers often build cost models around, and it doesn't carry over.

Teams working from outdated guides are especially exposed when sticker prices change as often as DeepSeek's have. A unified gateway like Concentrate AI can log actual spend broken out by model, provider, and time period, which makes it possible to see whether a team is still being billed, or budgeting, against rates that are months out of date.

OpenAI's current API rates and model lineup

OpenAI's pricing spans a wide range depending on which model in the family a team is calling, so any comparison against DeepSeek has to specify a tier before it means anything.

At the flagship end, GPT-5.5 carries a higher input rate and a higher output rate than the rest of the lineup. GPT-5.4 runs $2.50 per million input tokens, and its output rate is higher still. Below that sit GPT-5.4-mini and GPT-5.4-nano, both priced lower on input and output than the 5.4 flagship, built for workloads where raw capability matters less than throughput and cost. At the very top, OpenAI's GPT-6 Astra flagship launched September 3, 2026, at $50.00 per million output tokens, a rate that puts it in a different cost category entirely from anything else in the lineup.

GPT-4o deserves a specific note, because it causes real confusion. OpenAI pulled GPT-4o from the ChatGPT app in February 2026, and plenty of people read that as a retirement. It wasn't retired: the model stayed available through the API, and the chatgpt-4o-latest snapshot was retired around February 16 to 17, 2026, but other GPT-4o API variants kept running. GPT-4o's API rate has actually dropped since launch: the output rate is now $10.00 per million tokens, down from its original price.

The GPT-4.1 family, which launched in April 2025 with GPT-4.1, GPT-4.1 mini, and GPT-4.1 nano, rounds out the lineup at the lower end, and gives the clearest comparison point against DeepSeek's efficiency tier.

There's no single "OpenAI price" to put next to DeepSeek's. The comparison only works model by model, tier by tier.

Head-to-head: what DeepSeek costs compared to OpenAI at equivalent tiers

The honest answer to "is DeepSeek cheaper than OpenAI" is that it depends on which models you're putting side by side, what time your traffic runs, and how much of your input is cached. At the flagship tier the gap is large. At the efficiency tier it narrows, and sometimes it flips.

| Comparison | DeepSeek rate period | Result | |---|---|---| | DeepSeek V4 Pro vs. GPT-4o | Peak | DeepSeek materially cheaper on both input and output | | DeepSeek V4 Pro vs. GPT-5.5 | Peak | Gap widens substantially, especially output-heavy workloads | | DeepSeek V4.1 Flash vs. GPT-4.1 Nano | Off-peak | OpenAI Nano slightly cheaper at standard rates | | DeepSeek V4.1 Flash vs. GPT-4.1 Mini | Peak | Flash still cheaper, but by a smaller margin than headline comparisons suggest |

At the flagship tier, DeepSeek V4 Pro beats GPT-4o on both input and output cost by a wide margin when you use peak pricing. Against GPT-5.5, the gap gets even bigger, especially for agentic workloads that generate long responses, because output tokens pile up fast in that kind of use case.

The efficiency tier tells a more complicated story. DeepSeek V4.1 Flash at off-peak rates against GPT-4.1 Nano actually comes out slightly behind: OpenAI's Nano is a touch cheaper at standard rates. Flash against GPT-4.1 Mini at peak still comes out ahead for DeepSeek, but the margin is smaller than the headline "DeepSeek is 10x cheaper" comparisons that circulate online.

Timing is what makes or breaks the efficiency-tier comparison. A team whose traffic clusters inside DeepSeek's UTC peak windows can end up paying rates that erase most of the apparent advantage over OpenAI's mini-tier models. A team whose traffic happens to fall off-peak gets a very different bill for identical usage.

Cache behavior adds another variable. DeepSeek's cache-hit discount is more aggressive than OpenAI's, so workloads built around stable, repeated prefixes, RAG pipelines, support agents, code assistants carrying large system prompts, can push DeepSeek's effective input cost down close to zero. OpenAI's own cache discount exists but doesn't go nearly as far.

Output cost is where DeepSeek's advantage holds up most consistently. Across comparable tiers, DeepSeek's output rates run lower than OpenAI's, and because output tokens dominate the bill on agentic and long-form generation work, that's where you see the savings compound the most.

None of this produces a single winner. The cost gap between DeepSeek and OpenAI shifts hard depending on peak timing and cache behavior, so a team choosing between the two needs to know its own traffic pattern before the comparison means anything. A gateway that routes across both providers and logs actual cost per workload, which is what Concentrate AI is built to do, can show the real winner for a specific traffic pattern instead of forcing a decision based on sticker price alone.

Why DeepSeek's pricing has changed so many times

DeepSeek's pricing history is a string of fast moves, each with a short window to adapt, so you can't plan a budget around it, and that creates ongoing work for any team running production traffic against the API.

Since V4 launched on April 24, 2026, the changes have come fast, one after another. On July 24, 2026, DeepSeek retired the deepseek-chat and deepseek-reasoner aliases outright, and calls using those names started returning errors instead of routing to a model. On July 31, 2026, the deepseek-v4-flash route was upgraded in place to V4-Flash-0731, a change that altered what a team was actually calling without changing the endpoint name. On August 16, 2026, DeepSeek introduced peak and off-peak billing, and for teams running heavy traffic during what turned out to be peak hours, some line items jumped sharply. On September 10, 2026, V4.1-Flash went live under yet another new name, deepseek-flash, the older Flash aliases were retired, and the V4 Pro rate card stayed unchanged from the August 16 structure. Then, on September 10 and 11, 2026, a planned retirement of V4 Pro, originally scheduled for September 14, got reversed after user demand pushed back, and billing continued unchanged.

Most of these moves, taken on their own, brought prices down. But the model name changes and alias retirements cost teams real engineering time, because a missed deadline on an alias retirement doesn't just mean a slightly different price. The old model name stops resolving and starts throwing errors, so it means a broken integration.

The August 16 peak/off-peak shift stands apart from the rest of these changes. It wasn't a price change in the usual sense. It changed how billing works in the first place; a team can't just update a number in a budget spreadsheet. Traffic scheduling itself becomes a cost lever, requiring teams to rethink when jobs run, not just what they cost.

OpenAI's pattern looks different. GPT-4o's API rate has dropped since launch, landing at $10.00 per million output tokens today, and the GPT-4.1 family, launched in April 2025, has held stable pricing since. OpenAI tends to add new models alongside old ones instead of retiring aliases and turning them into errors.

None of this means DeepSeek is unreliable as a provider. It means a provider moving this fast, with this much market pressure behind it, needs more active monitoring than a team might be used to from OpenAI's slower release cadence. That's a planning fact, not a knock on the company. Spend visibility that's granular by model, provider, and time period, the kind a tool like Concentrate AI provides, matters more here than it would with a provider whose pricing sits still for a year at a time, because the cache-hit discounts and peak/off-peak rate periods both shift the real cost of a workload independent of the sticker price, and only actual logged spend shows which rate period a team's traffic is really landing in.

The compliance and data sovereignty problem that pricing comparisons leave out

For any team handling personal, financial, or health data, the hosted DeepSeek API carries a documented data sovereignty problem that no price advantage makes up for.

DeepSeek's own privacy policy states that personal data is stored in the People's Republic of China. The hosted API comes with no enterprise Data Processing Addendum, no Business Associate Agreement, and no option for EU or US data residency. If your team operates under GDPR, HIPAA, or similar frameworks, that absence alone rules the hosted API out for a wide range of use cases before cost even enters the conversation.

The incidents on record back up the concern. Security firm Wiz found two DeepSeek ClickHouse databases sitting open with no authentication at all, exposing more than a million log entries that included plaintext chat history, API keys, and backend secrets. South Korea's Personal Information Protection Commission separately found that DeepSeek had transferred user prompts to Beijing Volcano Engine Technology and other Chinese companies without user consent.

Regulators and large employers have responded accordingly. Italy pulled DeepSeek from its app stores over GDPR concerns. Australia, Taiwan, the US Navy, and NASA have restricted its use on official systems. Microsoft has prohibited employees from using it, and Toyota and Mitsubishi Heavy Industries have put the same restriction in place for their own staff.

The open-weight models, R1 and V3, sit outside this problem entirely, and the distinction matters for anyone weighing the hosted API against other options. When a team self-hosts those weights on infrastructure it controls, the data never crosses into the vendor's own servers in the first place, so the transfer issue that triggered the security findings and the regulator's ruling doesn't apply. That route takes real infrastructure investment and operational capacity though, more than most teams calling the hosted API currently have in place.

Taken together, DeepSeek's price advantage is real at the API level, and this article has documented it across every tier. For teams bound by data sovereignty rules, that advantage runs into a limit that no rate card changes.