DEV Community

Cover image for DeepSeek API: deepseek-flash vs V4-Pro for New Tools
Shamuddin
Shamuddin

Posted on Originally published at aitecharchive.com

DeepSeek API: deepseek-flash vs V4-Pro for New Tools

On November 13, 2026, DeepSeek opened the features behind its consumer app to developers through the DeepSeek API: web search, file input, URL reading, safety checks, OCR, speech-to-text and text-to-speech (official announcement). The built-in tools run on deepseek-flash, and for tool-heavy work it beats deepseek-v4-pro: it is the model the tools are built for, it supports vision, and its output tokens cost $0.60 per 1M off-peak against $1.98 for V4-Pro (pricing page). Search costs $0.75 per 1,000 searches and OCR costs $0.225 per 1,000 pages, both billed on top of tokens per the tool fee table on the pricing page. Choose V4-Pro only for hard, text-only reasoning jobs: it carries the lower 500-request concurrency limit against Flash’s 2,500, and it cannot process images at all.

TL;DR

  • Five built-in tools (web search, deep thinking search, file input, URL reading, safety checks) now work on deepseek-flash across the Messages, Responses and Chat Completions APIs.
  • Standalone Search, Fetch, OCR, ASR and TTS APIs are also open.
  • Tool fees (pricing page): $0.75 per 1,000 searches, $0.75 per 1,000 fetches, $0.225 per 1,000 OCR pages, with no peak/off-peak split.
  • Token prices on Flash (pricing page): $0.15 input (cache miss) and $0.60 output off-peak; $0.30 and $1.20 at peak, about ₹14 and ₹113 per 1M tokens in India terms at Wise’s 94.28 INR per USD rate.
  • The underlying model, released October 17, 2026, improved on all 11 vendor-reported benchmarks (model card via AgentsPulse), but frontier models still lead on several.
  • Last verified: 15 November 2026.

What did the DeepSeek API add on 13 November?

DeepSeek moved from selling a model to selling a small platform: the tool calls and context assembly happen server-side, so you enable a tool by parameter instead of writing the loop yourself. The announcement lists five tools:

  1. Web search. Searches the web or images and returns citations, in non-thinking mode, via the deepseek_web_search parameter.
  2. Deep thinking search. With thinking enabled, the model searches and reads pages across several reasoning steps.
  3. File input. Parses PDFs, Office files and images into the conversation.
  4. URL reading. The deepseek_read_url tool detects links in a user message and fetches the pages.
  5. Safety checks. Screens inputs and outputs and returns a refusal indicator when it detects risk.

DeepSeek states that public-facing apps must enable safety checks, and that developers stay responsible for content moderation, model registration and AI-content labelling. Treat that as a compliance requirement, not an optional flag.

How do I call deepseek-flash with these tools?

Use the model name deepseek-flash, which serves DeepSeek-V4.1-Flash-1017. The older names deepseek-v4-flash and deepseek-v4-flash-vision-exp are still accepted and are routed to the new model at Flash pricing, per the pricing page.

  • OpenAI-style base URL: https://api.deepseek.com
  • Anthropic-style base URL: https://api.deepseek.com/anthropic
  • Context length: 1M tokens, with up to 384K tokens of output (pricing page)
  • Supported features: JSON output, tool calls, the Responses API, the Anthropic API and vision

For an existing OpenAI-compatible client, the change is usually the base URL, the model name and the tool parameter. Test with one long document and one URL before production.

How much does the DeepSeek API cost in dollars and rupees?

Prices below come from the official pricing page, checked on 15 November 2026. Rupee figures use 94.28 INR per USD, the Wise rate on the same date, and are rounded.

Item (deepseek-flash) Off-peak Peak Peak in INR
Input, cache hit, per 1M tokens $0.003 $0.006 under ₹1
Input, cache miss, per 1M tokens $0.15 $0.30 ₹28
Output, per 1M tokens $0.60 $1.20 ₹113

Tool fees do not change with the time of day:

Tool Price In INR
Search (text, image) $0.75 per 1,000 about ₹71
Web fetch $0.75 per 1,000 about ₹71
File parsing / OCR $0.225 per 1,000 pages about ₹21
TTS $0.45 per 10,000 characters (list), $0.225 (limited-time) ₹42 list, ₹21 promo
ASR $0.116 per hour of audio (list), $0.058 (limited-time) ₹11 list, ₹5.5 promo

The pricing page lists the TTS and ASR promotions as limited-time 50 percent off with no announced end date, so budget at list price for anything long-running. Tool fees are billed directly from your account balance.

Peak hours are 01:00 to 04:00 and 06:00 to 10:00 UTC, Monday to Friday, excluding Chinese public holidays — 06:30 to 09:30 and 11:30 to 15:30 in IST. Part of the Indian working day sits in the half-price window and part does not; evening and weekend batches get the lower rate.

Is deepseek-flash or deepseek-v4-pro better for tool use?

For most tool workloads, deepseek-flash wins. The comparison, from the pricing page:

deepseek-flash deepseek-v4-pro
Output, off-peak / peak per 1M $0.60 / $1.20 $1.98 / $3.96
Input cache miss, off-peak / peak per 1M $0.15 / $0.30 $0.66 / $1.32
Vision Yes No
Concurrency limit 2,500 500

Flash is cheaper by a factor of about 3.3 on output (V4-Pro’s $3.96 peak output versus Flash’s $1.20, from the pricing page) and it handles images, which matters because file input and OCR-style work often include scans and screenshots. V4-Pro makes sense for harder text-only reasoning tasks at small scale; Flash carries the higher 2,500-request concurrency tier. For a fuller breakdown of how the two models relate, see our V4-Pro vs V4.1-Flash retirement analysis.

What are the limits of the OCR, ASR and TTS APIs?

These limits come from secondary write-ups (Maccus, WebProNews); confirm them against the official docs before designing around them.

  • OCR: up to 1,000 pages and 100 MiB per file. Upload, recognition and response must finish within 180 seconds. It offers plain text recognition and a rich layout mode with per-region positions (Maccus).
  • ASR: 21 languages with automatic detection, which you cannot set manually. It is non-streaming, with a cap of 120 seconds of audio per request, so long recordings need chunking (WebProNews).
  • TTS: two voices and 10 languages, with WebSocket streaming.
  • URL reading: reportedly the first 10 distinct links in a message.

The 120-second ASR cap and the 180-second OCR window are the constraints most likely to shape your architecture. For dedicated Hindi voice output, compare Chariot-1, a multilingual Hindi text-to-speech API before committing.

How good is the model behind the DeepSeek API?

DeepSeek-V4.1-Flash-1017 was released on October 17, 2026. The scores below are vendor-reported model-card figures as relayed by AgentsPulse, compared with the September V4.1-Flash checkpoint:

Benchmark 1017 September Change
Terminal-Bench 4.0 42.3 31.2 +11.1
HLE (no tools) 42.6 36.8 +5.8
DeepSWE v1.1 (Pass@1) 77.0 74.2 +2.8

All 11 reported benchmarks improved, with gains between 2.0 and 11.1 points, per the model card relay. In the same vendor-reported table (AgentsPulse), Claude Opus 5 scores 56.3 on HLE and 51.8 on Terminal-Bench 4.0, so Flash is not the strongest reasoner available. Treat these as vendor numbers: harness, reasoning effort and benchmark version can all move a score. For a head-to-head against other frontier models, read DeepSeek V4.1 Flash vs Claude Opus 5, GPT-5.6 Sol and Kimi K3. If you build coding agents, the DSH open-source harness leaderboard shows how these models behave in practice, and Gemini 4 Hydrogen vs GPT-6.1 Terra covers the wider pricing picture.

What are the trade-offs before I adopt it?

  • Vendor lock-in. Server-side tool orchestration is convenient, but it ties your retrieval and parsing behaviour to one provider.
  • Compliance duty. Public apps must enable safety checks, and moderation and labelling duties remain yours.
  • Promotional pricing. Voice discounts have no end date, which also means no guarantee.
  • Unverified limits. Several OCR, ASR and TTS limits above are secondary-source figures.

The practical test before moving traffic: run a representative sample of your own documents and queries, record cost per task, and compare accuracy.

FAQ

Q: What is the DeepSeek API model name for the new tools?
A: deepseek-flash, which serves DeepSeek-V4.1-Flash-1017. The older deepseek-v4-flash and deepseek-v4-flash-vision-exp names route to it at Flash pricing.

Q: How much does DeepSeek API web search cost?
A: Search costs $0.75 per 1,000 searches, about ₹71 at 94.28 INR per USD, on top of token charges (DeepSeek pricing; Wise rate). Tool fees have no peak or off-peak split.

Q: Does the DeepSeek API support OCR?
A: Yes. File parsing and OCR cost $0.225 per 1,000 pages, about ₹21 (DeepSeek pricing). Reported limits (Maccus) are 1,000 pages and 100 MiB per file, with a 180-second processing window.

Q: How long can an audio file be for DeepSeek speech-to-text?
A: Reports (WebProNews) put the cap at 120 seconds per request, with 21 languages, automatic language detection and no streaming. Split longer recordings into chunks.

Q: Do I have to enable safety checks?
A: For public-facing apps, DeepSeek says yes. Moderation, model registration and AI-content labelling duties remain yours.

Q: Is deepseek-flash or deepseek-v4-pro cheaper?
A: Flash: $0.60 per 1M output tokens off-peak against $1.98 for V4-Pro (DeepSeek pricing), and Flash is the one that supports vision.

This article was prepared with AI assistance and checked against DeepSeek's published documentation on 15 November 2026. Corrections log: none yet.

Top comments (0)