Anvat
DeepSeek1M contextdeepseek-v4

DeepSeek V4 Pro

Open-weight flagship — 80.6% SWE-Bench at 1/20th the cost of Claude Opus 4.7

DeepSeek V4 Pro is the leading open-weight frontier model in 2026 — released April 24, 2026 as a 1.6T-parameter Mixture-of-Experts (49B active) with a 1M-token context window, MIT-licensed. It hits 80.6% on SWE-Bench Verified (within 0.2 points of Claude Opus 4.6) and 93.5 on LiveCodeBench (the highest of any model), at $1.74/$3.48 per MTok — roughly 1/20th of Opus 4.7's output price.

Pricing

RateList priceAnvat effectiveSavings
Input$1.74$1.2230%
Output$3.48$2.4430%
All prices per million tokens (MTok). List = provider direct. Anvat effective = 30% discount applied.

Pricing verified 2026-06 · See full Anvat pricing

Strengths

  • 80.6% SWE-Bench Verified — within 0.2 points of Claude Opus 4.6
  • 93.5 LiveCodeBench Pass@1 — highest of any model
  • 1M-token context window standard, no separate pricing tier
  • MIT-licensed weights — self-host on vLLM/SGLang/TGI
  • MoE architecture (1.6T total / 49B active) keeps inference cost low
  • FP4 + FP8 mixed precision halves memory footprint

Where it underperforms

  • Trails Opus 4.7 + GPT-5.5 by ~2-3 points on hardest reasoning (HLE, HMMT)
  • Trails GPT-5.4 by 7 points on Terminal-Bench 2.0
  • Text-only — no native multimodal (no vision, no audio)
  • Interleaved-thinking pattern requires agent framework updates
  • Tool ecosystem (LangChain etc.) still primarily Anthropic/OpenAI-shaped

Use cases this model is the right pick for

  • High-volume coding tasks where cost per request is the binding constraint
  • Open-weight requirements (regulatory, audit, on-prem option)
  • Long-context analysis at 500K-1M tokens (full codebase, large legal docs)
  • Pairing with closed frontiers in router patterns (V4-first, escalate on hard tasks)
  • Bulk generation (content drafting, summarisation, translation)

Benchmarks

  • coding

    SWE-Bench Verified

    80.6% (within 0.2pt of Opus 4.6)

  • coding

    LiveCodeBench Pass@1

    93.5 (highest of any model)

  • coding

    Codeforces Rating

    3206 (ahead of GPT-5.4 xHigh)

  • agentic

    MCPAtlas Public

    73.6 (#2 after Opus 4.6)

  • reasoning

    MRCR 1M long-context recall

    83.5

  • agentic

    Toolathlon

    51.8 (ahead of Gemini 3.1 Pro)

Benchmark numbers self-reported by provider; verify against the latest publisher documentation before quoting.

Quickstart

Same wire format as direct provider APIs — your existing SDK code keeps working. Point at api.anvat.app/v1 and use your Anvat key.

import OpenAI from "openai";

const client = new OpenAI({
  baseURL: "https://api.anvat.app/v1",
  apiKey: process.env.ANVAT_API_KEY,
});

const response = await client.chat.completions.create({
  model: "deepseek-v4",
  max_tokens: 4096,
  messages: [
    { role: "user", content: "Implement a Trie data structure in TypeScript with full test coverage." },
  ],
});

Try DeepSeek V4 Pro — 30% off list

Same model, same quality, same wire format — at the discounted Anvat effective rate. $2 free credit on signup, no card required.

Get a key →

Related