xAI Grok Code Fast 1: The 'Dirt-Cheap Rocket' Hitting the Coding Model Track at $0.20/M Tokens

Forum topic · 小凯 · 2026-08-27

Summary

On August 29, 2026, xAI launched Grok Code Fast 1, a coding-specialized model scoring 70.8% on SWE-bench Verified while pricing input tokens at $0.20 per million and output at $1.50 per million — roughly 6x cheaper than Gemini 2.5 Pro on both dimensions. The model uses a Mixture-of-Experts architecture (reportedly ~314B total parameters with an inferred 24B–40B active), a 256K input context window, and was trained with a 'PR-aware' pipeline built on real, human-merged Pull Requests rather than conventional RLHF-only annotation. xAI also claims over 90% prompt-cache hit rates in collaborative coding, potentially driving effective blended input costs down to around $0.04/M tokens. The model was rolled out simultaneously with limited-time free access across seven mainstream coding clients: GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf. This article analyzes the pricing economics versus GPT-5, Gemini 2.5 Pro, and open-source M2, the shift from benchmark-oriented to production-ready training signals, and what the seven-platform distribution strategy means for the AI coding market in late 2026.

Key points

Launch date & availability: xAI released Grok Code Fast 1 on August 29, 2026, simultaneously offering limited-time free access through seven coding clients: GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf.
Pricing: $0.20 per million input tokens and $1.50 per million output tokens — the cheapest entry point to date at the 70%+ SWE-bench Verified performance tier.
Benchmarks: 70.8% on SWE-bench Verified (vs. Codex-1 at 72.3%, Claude 4 Opus at 69.1%, M2 at 69.4%, Gemini 2.5 Pro at 63.2%); 62.0% on LiveCodeBench; 25.8% on Terminal-Bench; 4.3% on IOI; Arena Code Elo of 1164.
Architecture: MoE with an industry-reported ~314B total parameters; active parameter count is not officially disclosed but inferred at 24B–40B. Context window is 256K input + 10K output, with a knowledge cutoff of June 30, 2024 (per xAI documentation).
Training approach: Instead of generic SFT + RLHF, xAI built a coding-focused corpus and fine-tuned on a curated dataset containing real Pull Requests — an approach Anthropic engineers have called 'PR-aware training.' Training signals thus reflect whether human engineers actually merge AI-generated code, not just annotator preference.
Caching economics: xAI reports 90%+ prompt-cache hit rates in collaborative coding scenarios, which could reduce effective blended input cost to roughly $0.04/M tokens, since repeated code context is served from cache.

Price comparison table

| Dimension | Grok Code Fast 1 | Gemini 2.5 Pro | M2 | |---|---|---|---| | Input / M tokens | $0.20 | $1.25 | $0.30 | | Output / M tokens | $1.50 | $10.00 | $1.20 | | Context (input) | 256K | 1.05M | 1.00M | | Output cap | 10K | 65K | Unlimited | | SWE-bench Verified | 70.8% | 63.2% | 69.4% | | LiveCodeBench | 62.0% | 75.6% | 83.0% | | License | Proprietary | Proprietary | MIT (open source) |

That makes Grok Code Fast 1 about 6.25x cheaper on input and 6.7x cheaper on output than Gemini 2.5 Pro, despite scoring higher on SWE-bench Verified. A mid-sized SaaS company running 1 billion coding tokens per month could reportedly save on the order of $800K–$1.26M annually in input costs by switching.

Analysis highlights

First to combine two previously separate properties: Grok Code Fast 1 is described as the first coding-specialized model to simultaneously offer SWE-bench 70%+ performance and a $0.20/M input price — a 'value frontier' that did not previously exist.
Real-engineering strength vs. contest weakness: The 8.8-point gap between SWE-bench (70.8%) and LiveCodeBench (62.0%), plus the low IOI score (4.3%), suggests the model is stronger on real multi-file engineering tasks than on pure algorithmic contest problems — consistent with its PR-based training data.
Distribution strategy: The simultaneous seven-platform free rollout is framed as a 'terminal capture' play. As user decision points shift from 'which model' to 'which IDE/plugin,' embedding Grok into already-open tools builds switching-cost moats.
A three-part moat: Reproducing this position reportedly requires all of: (1) a 24B–40B active MoE coding model, (2) access to real PR datasets, and (3) 90%+ inference cache hit rates. As of the announcement, only xAI publicly checks all three boxes.

What to watch in the next 6–12 months

1. Q4 2026: Whether Anthropic's rumored coding-focused 'Mythos Preview' matches Grok's price at 70%+ SWE-bench. 2. Q4 2026: Whether OpenAI adjusts GPT-5 pricing (currently around $1.25/M input) in response. 3. Q1 2027: Revenue-sharing terms if Cursor or Windsurf permanently bundle Grok — the first public sample of a client × model business model. 4. Q2 2027: Whether MIT-licensed M2 (69.4% SWE-bench, $0.30/M) closes the gap to 70.8%.

References

1. xAI, "Grok Code Fast 1", https://docs.x.ai/developers/models/grok-code-fast-1, 2026-08-29. 2. LLM-Stats, "Grok Code Fast 1 vs MiniMax M2 / Gemini 2.5 Pro", https://llm-stats.com/models/compare/, 2026-08-29. 3. AnotherWrapper, "Grok Code Fast 1 Pricing", https://anotherboilerplate.com/tools/llm-pricing/grok-code-fast-1, 2026-08-27.