Grok 4.5 announced: the first Grok built specifically for coding and agents

What it is (这是什么)

The announcement sweeping X right now introduces Grok 4.5 as the team's first model "trained specifically for coding and agents." Two details carry the weight. First, the model was trained with Cursor, the AI code editor — suggesting it was shaped by real agent-and-human development workflows rather than web-scraped code alone. Second, the post claims "frontier intelligence" delivered at "leading speeds and cost efficiency."

Just as notable is what the post leaves out: no parameter count, no benchmark table, no token pricing, no context window. For a frontier-model launch, that restraint is unusual — so treat the capability claims as vendor claims until independent results appear.

Why it matters (为什么重要)

Grok 4.5 arrives in a market whose ground rules changed within the past week. DeepSeek released V4 Flash on July 31, 2026 with "substantially enhanced agentic capabilities," 304 billion parameters, and pricing of $0.14 per million input tokens and $0.27 per million output tokens. Simon Willison calls it "the best value-per-intelligence model out there," noting Artificial Analysis ranks it ahead of MiniMax M3, a 428B model. The price of agentic intelligence is collapsing — and every model vendor is responding to that pressure.

The category shift matters just as much. A model trained specifically for coding and agents, rather than a general-purpose chat model with code bolted on, tracks where actual demand sits: tool-use reliability, agent-loop completion, and editing speed. Training with Cursor is a data strategy as much as a product decision — the model learns from traces of real assistant-driven development instead of static code dumps.

Timing reinforces the point. The Model Context Protocol shipped its new MCP 2.0 spec on 2026-07-28, which Willison calls the most significant change to the spec since it launched. The agent tooling ecosystem is resetting right now, and a dedicated coding-and-agents model arrives just as that infrastructure layer standardizes.

Key features or specs (核心特性与规格)

• First model in its line trained specifically for coding and agents.
• Trained with Cursor.
• Claimed frontier intelligence at leading speeds.
• Claimed cost efficiency.
• Not disclosed: parameters, context window, token pricing, benchmarks.

How it compares (对比)

There is no direct public comparison yet, because Grok 4.5 has no published numbers. What we can anchor against is the current floor for agentic coding: DeepSeek V4 Flash is a 304B-parameter model — 167GB of weights on Hugging Face — that ranks above MiniMax M3 (428B) on Artificial Analysis and prices at $0.14 / $0.27 per million tokens. If "leading cost efficiency" means undercutting that, it is a high bar. If it means matching that value with different quality characteristics, the market just became genuinely two-sided: a specialized, Cursor-trained proprietary model against a cheap open-weights generalist that "punches well above its weight."

Who should use it (适合谁)

• Developers who live inside Cursor and other coding agents — you are both the training data and the target user.
• Agent builders shipping MCP-based tooling; the MCP 2.0 rollout is the moment to test a new model class.
• Teams that track cost-per-task rather than raw benchmark scores — and that will wait for independent verification before switching.

FAQ (常见问题)

What is Grok 4.5? Per the announcement, it is the first model in its line trained specifically for coding and agents. It was trained with Cursor and is claimed to deliver frontier intelligence at leading speeds and cost efficiency. The post does not disclose parameters, benchmarks, context window, or pricing.

Does Grok 4.5 ship inside Cursor? The announcement says the model was trained with Cursor — not that it is bundled into Cursor. Availability, access, and pricing are not disclosed.

How does it compare with DeepSeek V4 Flash? Not directly — Grok 4.5 has no published benchmarks or pricing yet. For reference, DeepSeek V4 Flash is a 304B-parameter model at $0.14/M input and $0.27/M output tokens, described by Simon Willison as the best value-per-intelligence model currently available.

FAQ

1. q

a

1. q

a

1. q

a