English static mirror for SEO/GEO · AI-assisted translation · Read Chinese original

DeepSeek V4 Pro 0813 and Grok 4.6 Launch on the Same Night — AI Coding Backends Split into "Cheap and Strong" Tracks

Forum topic · 小凯 · 2026-08-13

Summary

On August 12, 2026, DeepSeek V4 Pro 0813 and SpaceXAI's Grok 4.6 launched within two hours of each other, capping an August wave of AI coding backend upgrades. DeepSeek V4 Pro 0813 is a GA release with the same 1.6T MoE architecture (49B active) and 1M context as the April preview, but post-training drove dramatic gains: DeepSWE rose from 12.8 to 62.7, Terminal Bench 2.1 from 72.1 to 87.9, and Cybergym from 52.7 to 83.3. Pricing stays at $0.435/$0.87 per million input/output tokens — about 1/60 of Claude Fable 5 Max. Grok 4.6 keeps the 1.5T V9 base, expanded experts from 128 to 256 while keeping 8–16 active per token, and scores 61 on AA Intelligence Index versus Fable 5 Max's 62 at roughly half the price. Cursor offers 2× first-week usage. Both target mid-priced long-horizon agents, giving developers a third option beyond the cheap-vs-ecosystem tradeoff.

Key points

  • Same-night launches on August 12, 2026: DeepSeek V4 Pro 0813 and SpaceXAI's Grok 4.6 went live within two hours of each other, marking the closing segment of August's AI coding backend upgrade curve.
  • DeepSeek V4 Pro 0813: same 1.6T MoE / 49B active parameters and 1M context as the April preview. The release is purely post-training (GA stabilization + multi-turn agent RL). Benchmarks jumped sharply: DeepSWE 12.8 → 62.7, Terminal Bench 2.1 72.1 → 87.9, Cybergym 52.7 → 83.3. Pricing remains at $0.435 / $0.87 per million tokens (input/output) — roughly 1/60 of Claude Fable 5 Max.
  • Grok 4.6: built on the same 1.5T V9 base used since July, with experts expanded 128 → 256 while keeping 8–16 active per token (more "selectable capability," not higher per-inference cost). Scores 61 on AA Intelligence Index (vs. Fable 5 Max 62, GPT-5.6 Sol 61). DeepSWE v1.1: 65.9%. Pricing: $2 / $6 per million tokens.
  • API ecosystem reach: DeepSeek exposes OpenAI, Anthropic, and Responses API formats, plus Chat and Huawei Ascend support — meaning any LiteLLM or routing layer silently upgrades to V4 Pro 0813. Grok 4.6 is integrated with Cursor (2× first-week usage), Grok Build, OpenRouter, Vercel, and Cloudflare.
  • Long-horizon agent efficiency: Singularity.Kwi's AA-Briefcase comparison shows Grok 4.6 completes the same task in ~53 turns and ~0.5B input tokens vs. Claude Opus 5 (max) at ~103 turns and ~2B tokens — half the turns, one-quarter the tokens.
  • DeepSWE as the new battleground: full-codebase multi-file editing benchmark replacing HumanEval/MBPP. Top scores: GPT-5.6 Sol 73%, Fable 5 Max 70%, Grok 4.6 65.9%, DeepSeek V4 Pro 0813 62.7%.
  • Grok 4.6 pricing caveat: above 200K prompt tokens, input jumps to $4 and output to $12 per million — applied to the entire request, which can hurt accumulating-context agents.
  • Comparison table

    | Dimension | DeepSeek V4 Pro 0813 | Grok 4.6 | Claude Fable 5 Max | |---|---|---|---| | Total / active params | 1.6T MoE / 49B | 1.5T MoE / 8–16 active | undisclosed | | Context / max output | 1M / 384K | 256K (stable 128K) / 500K (if benchmark supports) | 200K | | AA Intelligence Index | — | 61 | 62 | | DeepSWE v1.1 | — | 65.9% | 70% | | Terminal-Bench v3.0 | — | 26% | 34.1% | | DeepSWE (preview → GA) | 12.8 → 62.7 | — | — | | Terminal Bench 2.1 (preview → GA) | 72.1 → 87.9 | — | — | | Cybergym (preview → GA) | 52.7 → 83.3 | — | — | | Price ($/MTok in / out) | 0.435 / 0.87 | 2 / 6 | 5 / 25 | | Platforms | API (OpenAI + Anthropic + Responses) / Chat / Huawei Ascend | Cursor / Grok Build / OpenRouter / Vercel / Cloudflare | Anthropic API / Bedrock / Vertex |

    Product strategy

    Both launches sit at the "mid-price + long-horizon agent" tier. Cursor's 2× first-week usage for Grok 4.6 and DeepSeek's free pre-price-hike migration window mean developers no longer face a binary choice between "cheap DeepSeek" and "integrated Grok ecosystem."

    Ranking by dimension: 1. Price-performance: DeepSeek V4 Pro 0813 wins (60× gap) 2. Multimodality: Grok 4.6 wins (text + image + audio + video + Grok Build) 3. Default agent harness coverage: DeepSeek wins (API protocol-level reach) 4. Long-horizon reasoning efficiency: Grok 4.6 wins (half the turns, 1/4 the tokens) 5. Long-context cost: Grok 4.6 may lose (200K+ automatic price jump)

    Limitations and unknowns

  • DeepSeek V4 Pro 0813 has no vision/audio/video — multimodal "full" version still undated as of August 13
  • Grok 4.6 200K+ long-context pricing jump real-world impact needs more deployment data
  • DeepSeek pricing is pre-increase — future hike magnitude and zero-cost switch window length undisclosed
  • Neither model has published independent third-party API latency / throughput data
  • Grok 4.6 quality above 200K context lacks public benchmarks
  • Why same-night timing matters

    The dual launch gives developers a unified "mid-price, long-context, strong-agent, cross-harness" option that didn't exist at the start of August. Domestic developers get a fully self-hostable Ascend path without NVIDIA dependency; international developers get a Cursor/Cloudflare/Vercel/OpenRouter-defaulted model with full agent chain capability at half the cost of Opus 5 or Fable 5 Max. August 12 effectively splits AI coding backends from "expensive but strong" into two independent tracks: "expensive but strong" and "cheap but strong."

    Sources

  • https://api-docs.deepseek.com/quickstart/pricing
  • https://weibo.com/5424490006/5331252565247196
  • https://k.sina.com.cn/article_7879923925_1d5ae18d506801mg5e.html
  • https://www.cnblogs.com/sing1ee/p/22436555
  • https://k.sina.com.cn/article_7879996424_1d5af340806801n246.html
  • http://x.ai/news/grok-4-6
  • https://tech.ifeng.com/c/8vX3XV7yIcU
  • https://singularity.kiwi/spacexai-grok-4-6-frontier-cheap-agents-2026
  • http://agentbreaking.com/blog/grok-4-6-deep-dive-benchmarks

Tags

#deepseek-v4-pro#grok-4-6#ai-coding-backend#long-horizon-agents#deepswe-benchmark#model-pricing#huawei-ascend#agent-infrastructure

This page is an English static mirror generated for search and AI citation. It may be a full translation or structured summary of the Chinese original. Canonical interactive discussion lives on the Chinese page: https://zhichai.net/topic/178633409