DeepSeek V4 1M Context: max_tokens, Billing, and Concurrency

DeepSeek V4's 1M context window is a shared input-and-output budget. It does not mean you can send 1M input tokens and then generate another 384K. The 384K figure is the maximum output, while max_tokens is the output ceiling you choose for one request. Actual output is also limited by the space left in the 1M window^[1][2].

This guide answers the practical questions behind that specification: how to budget input and output, how to set max_tokens, what thinking tokens do to usage, how cache billing works, and why Flash and Pro have different concurrency limits.

DeepSeek V4 limits at a glance

DeepSeek released the V4 preview on April 24, 2026 with Flash and Pro variants and a 1M context window^[3]. The official pricing page lists the same 384K output cap for both models, together with JSON Output, tool calls, prefix completion, and FIM completion^[1].

Does 1M mean input only?

No. DeepSeek's Chat Completions reference says the total length of input tokens and generated tokens is limited by the model context length^[2].

Use this mental model:

input tokens + generated tokens <= 1M context
generated tokens <= max_tokens
generated tokens <= 384K model output limit

If the prompt nearly fills the context window, the model cannot also return a 384K answer. Reserve room for system instructions, tool schemas, conversation history, and the completion.

A large window also does not remove the value of retrieval and chunking. Sending less, better-organized context often improves latency, cost, and answer focus.

How to set max_tokens

max_tokens is the maximum completion length for one request. A finish_reason of length can mean the request reached this cap or exhausted the available context^[2].

Reasonable starting points:

These are engineering starting points, not official requirements. Measure finish_reason, actual completion tokens, latency, and useful-output rate before raising the cap.

from openai import OpenAI

client = OpenAI(
    api_key="YOUR_REAPI_KEY",
    base_url="https://api.reapi.ai/v1",
)

response = client.chat.completions.create(
    model="deepseek-v4-pro",
    messages=[
        {
            "role": "user",
            "content": "Review this repository inventory and produce a risk-ranked migration plan.",
        }
    ],
    max_tokens=16_384,
    reasoning_effort="high",
    extra_body={
        "thinking": {"type": "enabled"},
        "group": "default",
    },
)

See the DeepSeek V4 API reference for reAPI request fields, and the DeepSeek V4 model page for the playground and live gateway prices.

Thinking tokens and completion usage

Thinking mode is enabled by default. DeepSeek returns reasoning in reasoning_content, and the usage object can report it separately as completion_tokens_details.reasoning_tokens^[4][2].

Plan completion capacity around reasoning as well as the visible final answer. In thinking mode, temperature, top_p, presence_penalty, and frequency_penalty are accepted for compatibility but do not take effect^[4].

For deterministic extraction or low-latency work, test Flash with thinking disabled. Keep thinking enabled for planning, coding, and multi-step tool use where the additional reasoning is useful.

How billing works

DeepSeek Direct bills three token buckets:

1. cached input;
2. uncached input;
3. output.

The official USD rates on July 30, 2026 were^[1]:

The general formula is:

cost =
  cached_input / 1M × cached_rate
  + uncached_input / 1M × uncached_rate
  + output / 1M × output_rate

These are DeepSeek's direct prices. reAPI is a separate gateway with its own live USD rates, so use the pricing card on the reAPI DeepSeek V4 page for gateway billing.

Getting value from context caching

DeepSeek's disk cache is enabled automatically. A later request receives a cache hit only when it fully matches a stored prefix unit; similar wording is not enough^[5].

Put stable content first:

stable system prompt
+ stable tool schemas
+ unchanged repository or document corpus
+ the question that changes per request

Use the response fields prompt_cache_hit_tokens and prompt_cache_miss_tokens to calculate the real hit rate. Do not estimate savings from prompt similarity alone.

Flash and Pro concurrency

DeepSeek Direct publishes account-level concurrency limits of 2,500 for Flash and 500 for Pro. A request occupies one concurrency slot until its response finishes. The limit is per account, not per API key, and excess requests receive HTTP 429^[6].

The optional user_id can isolate safety, cache, and scheduling behavior, but ordinary accounts still share the total account concurrency across all user IDs. Creating more keys or user IDs does not increase capacity.

Those figures apply to DeepSeek Direct. Gateway concurrency and queueing can differ; when using reAPI, follow the limits and responses published by reAPI.

Production checklist for 1M-context requests

• Count tokens instead of treating characters as tokens.
• Reserve context for output, tool schemas, and follow-up turns.
• Set max_tokens to the task, not automatically to 384K.
• Record finish_reason and distinguish truncation from natural stops.
• Keep reusable prefixes stable and track cache hit tokens.
• Measure reasoning usage separately from visible answer length.
• Retry 429 responses with exponential backoff and jitter.
• Route high-volume routine work to Flash and reserve Pro for harder tasks.
• Use the current V4 model IDs; DeepSeek's announced deprecation date for deepseek-chat and deepseek-reasoner was July 24, 2026^[3].

FAQ

Does DeepSeek V4 Flash support 1M context?

Yes. Both Flash and Pro have a 1M context window and a 384K maximum output^[1].

Is the 1M context all input?

No. Input and generated output share the model context limit^[2].

Can I set max_tokens to 384K?

384K is the model output ceiling, but the request must also have enough space left in the 1M context window. The setting does not guarantee a 384K completion.

Do reasoning tokens count toward completion usage?

Capacity and billing plans should include them. DeepSeek reports reasoning tokens within completion usage details^[2].

Are concurrency limits per API key?

No. DeepSeek Direct limits concurrency per account: 2,500 for Flash and 500 for Pro^[6].

Do I need to enable prompt caching?

No. The context disk cache is automatic. Keep shared prefixes stable and inspect the cache hit fields in usage^[5].

Further reading

• DeepSeek V4 model and live pricing
• DeepSeek V4 API reference
• Browse reAPI chat models

References

1. DeepSeek. Models & Pricing — V4 context, output, features, prices, and concurrency. Retrieved July 30, 2026 from api-docs.deepseek.com/quick_start/pricing
2. DeepSeek. Create Chat Completion — max_tokens, context limits, finish reasons, and usage. Retrieved July 30, 2026 from api-docs.deepseek.com/api/create-chat-completion
3. DeepSeek. DeepSeek-V4 preview release. April 24, 2026. api-docs.deepseek.com/news/news260424
4. DeepSeek. Thinking Mode — controls, reasoning effort, and reasoning content. Retrieved July 30, 2026 from api-docs.deepseek.com/guides/thinking_mode
5. DeepSeek. Context Caching — prefix matching and usage fields. Retrieved July 30, 2026 from api-docs.deepseek.com/guides/kv_cache
6. DeepSeek. Rate Limit & Isolation — account-level concurrency and user_id. Retrieved July 30, 2026 from api-docs.deepseek.com/quick_start/rate_limit