Grok Code Fast 1: xAI's 'Price Butcher' Move into the Coding Model Race

> Note: This is a structured English rendering of a Chinese-language zhichai.net forum post. Claims below reflect the post's analysis; some figures (e.g., activated parameter counts) are industry estimates cited by the author, not official specifications.

Key points

• On August 29, 2026, xAI launched Grok Code Fast 1 simultaneously into seven mainstream AI coding clients: GitHub Copilot, Cursor, Cline, Kilo Code, Roo Code, opencode, and Windsurf, with limited-time free availability on each platform.
• Pricing: $0.20 per million input tokens, $1.50 per million output tokens — the cheapest entry yet at its performance tier.
• Benchmark scores (per materials reposted by the post): SWE-bench Verified 70.8%, LiveCodeBench 62.0%, Terminal-Bench 25.8%, Arena Code Elo 1164, IOI 4.3%. It trails only Codex-1 (72.3%) on SWE-bench Verified while undercutting competitors on price (Gemini 2.5 Pro: $1.25/M input, 63.2%; GPT-5: ~$1.25/M input).

The architecture: MoE + 256K context

• Grok Code Fast 1 uses a Mixture-of-Experts (MoE) architecture with a frequently cited 314B total parameters. Activated parameters are not officially disclosed; the author estimates 24B–40B based on throughput and peer models in the same tier.
• Context window: 256K input + 10K output (per xAI docs). Knowledge cutoff: June 30, 2024.
• The author calls it the first coding-specialized model to sit at the intersection of "SWE-bench 70%+" and "$0.20/M input price" — a price-performance frontier that previously did not exist.

PR-aware training

• Rather than generic SFT + RLHF, xAI reportedly built a coding-focused corpus and fine-tuned on a curated dataset of real, human-merged Pull Requests, plus real user usage feedback.
• The post links this to Anthropic engineers describing "PR-aware training" earlier in August: only code changes actually merged by human engineers signal *which AI-written code gets accepted*.
• This shifts the training signal from "does an annotator like it" to "would an engineer merge it" — consistent with Grok's stronger results on real engineering tasks (SWE-bench 70.8%) than on pure algorithmic problems (LiveCodeBench 62.0%, IOI 4.3%).

Prompt caching as a hidden cost weapon

• xAI reports a >90% prompt cache hit rate in collaborative coding scenarios. The author estimates an effective blended cost near $0.04/M tokens, since most context tokens (e.g., a large monorepo being refactored in Cursor) are re-read rather than recomputed.

Price comparison

| 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 | | SWE-bench Verified | 70.8% | 63.2% | 69.4% | | LiveCodeBench | 62.0% | 75.6% | 83.0% | | License | Proprietary | Proprietary | MIT (open source) |

• The post calculates Grok Code Fast 1 is ~6.25x cheaper on input and ~6.7x cheaper on output than Gemini 2.5 Pro, saving a mid-size SaaS spending on 1B tokens/month roughly significant six-figure annual input costs by switching.

Why seven platforms at once?

• The author frames the simultaneous free rollout as distribution-side terminal capture: developer decision points have shifted from "which model" to "which IDE/plugin." Building "switch to Grok in Cursor" muscle memory locks users into the Grok ecosystem regardless of default-model changes.
• It is the first time a single coding model went limited-free on seven independent clients at once; the closest prior example cited is DeepSeek V3.2's free push on three platforms in November 2025.
• Clients now face forced choices: permanently bundle Grok (with revenue sharing) or keep Claude and risk losing cost-sensitive users. The post argues clients gained real negotiating leverage for the first time in 2026 H2.

The verticalization inflection

• The post characterizes Grok Code Fast 1 as xAI's first move from a general-purpose "L4" model to a vertical "L5" coding model — while rumors point to a next-generation Grok 5 in training.
• After this launch, the author argues the definition of a competitive "coding model" becomes: MoE with ~24B–40B activated params + PR-aware data + 90%+ cache hit rate + SWE-bench 70%+ + $0.20/M input pricing. Entrants missing any of these will face immediate "why is it expensive or slow" scrutiny.

What to watch (author's outlook)

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

References cited in the post

1. xAI, "Grok Code Fast 1", https://docs.x.ai/developers/models/grok-code-fast-1, 2026-08-29. 2. 新智元, "Introducing Grok Code Fast 1", https://ima.qq.com/wiki/?shareId=daddb77b3579db6995ad71b308d08318020ebe6a9bc75e1c89127af949fe149d, 2026-08-29. 3. 行业研报·全球 AI 周报, "xAI 上线首个编码模型 Grok Code Fast 1", 2026-08-29. 4. LLM-Stats, "Grok Code Fast 1 vs MiniMax M2 / Gemini 2.5 Pro", https://llm-stats.com/models/compare/, 2026-08-29. 5. AnotherWrapper, "Grok Code Fast 1 Pricing", https://anotherboilerplate.com/tools/llm-pricing/grok-code-fast-1, 2026-08-27.