Grok 4.5: SpaceXAI and Cursor's Jointly Trained Coding Agent — 500K Context, $2/MTok, #1 Agentic Tool Use

On July 8, 2026, SpaceXAI released Grok 4.5 — its first model co-developed with Cursor, trained on real developer session data from the Cursor IDE. The release was broken by Axios and covered in depth by MarkTechPost and DataCamp. This is a research-based guide; we reviewed those sources and independent benchmark data from BenchLM and Artificial Analysis. We did not run the model ourselves.

What Grok 4.5 is

Grok 4.5 is SpaceXAI’s targeted model for coding, agentic tool calling, and knowledge work. It is not a general-purpose frontier model or a reasoning flagship — it is purpose-built for developer workflows and multi-step agent tasks, specifically the kind that happen inside an IDE over an extended session.

The model runs with a 500,000-token context window, accepts text and image inputs, and supports configurable reasoning effort (low, medium, or high — high is the default). Knowledge cutoff is February 1, 2026.

Grok 4.5 is built on SpaceXAI’s V9 foundation — the same 1.5-trillion-parameter base as Grok v9 — with additional training specifically for coding and agent tasks. SpaceXAI trained it across tens of thousands of NVIDIA GB300 GPUs on coding, science, engineering, and mathematics data, with reinforcement learning covering “hundreds of thousands of tasks, including agent runs that lasted for hours.”

The Cursor data advantage

The distinguishing fact about Grok 4.5 is where part of its training data came from. Per SpaceXAI’s launch announcement, the model was trained on real Cursor session data — actual interaction logs from developers working inside live codebases. That data does not exist anywhere outside of Cursor’s servers. No standalone model lab has it.

This matters in a specific, narrow way: the model’s training distribution includes what engineers actually do in long IDE sessions, not what researchers construct as proxy tasks. BenchLM’s evaluation puts Grok 4.5 at #1 on agentic tool use on the Artificial Analysis Intelligence Index — the one benchmark category most directly correlated with that training signal.

Whether Cursor-sourced training data provides durable advantage over larger-scale synthetic approaches is not settled. But the benchmark result exists, and the mechanism is plausible.

Benchmark results

Based on BenchLM’s independent benchmarks and DataCamp’s comparative analysis:

DataCamp’s analysis summarizes the picture accurately: Grok 4.5 leads on SWE Marathon and tops the agentic tool use ranking, but Claude Fable 5 leads on DeepSWE 1.0, DeepSWE 1.1, and SWE-Bench Pro. Grok 4.5 is #1 at the task type most resembling a real multi-step agent run; it trails on isolated repository-editing tasks where Fable 5 is optimized.

On speed, SpaceXAI claims approximately 80 tokens per second. Artificial Analysis’ independent measurement puts it closer to 120 tokens per second. Both readings put it at or above the frontier median.

Pricing

Standard pricing per SpaceXAI’s official page and DataCamp:

A fast-serving variant is available at approximately $4/$18 per MTok input/output for latency-sensitive paths, per launch coverage. Cursor offered 50% off usage at launch.

For comparison: Claude Opus 4.8 lists at $15/$75 per MTok (input/output). At $2/$6, Grok 4.5 is roughly 7–13× cheaper per token for most workloads. The comparison is not apples-to-apples — Opus 4.8 has a broader capability profile — but for coding and agentic tasks specifically, the cost gap is material.

Where it is available

Per SpaceXAI’s announcement and DataCamp:

• SpaceXAI API console — model ID: grok-4.5 — OpenAI-compatible endpoints
• Grok Build — default model as of launch
• Cursor — all plans, including free
• Model gateways — OpenRouter, Vercel AI SDK, Cloudflare, Snowflake, Databricks Mosaic
• Office add-ins — SpaceXAI’s enterprise integrations

EU availability: Not available as of launch day. SpaceXAI cited EU AI Act compliance as the reason. Expected later in July 2026 per their updated developer guide.

Builder decision guide

Use Grok 4.5 when:

• You are building multi-step agent workflows where tool calling and action chains are the primary cost driver
• Your tasks look like long IDE sessions — multi-file edits, repo exploration, iterative debugging — rather than one-shot code generation
• Token cost is a constraint and you need something closer to Opus-class quality at $2 input (not $15)
• You are already in the Cursor ecosystem and the 50% launch discount reduces risk

Consider alternatives when:

• You need maximum accuracy on isolated repository-editing benchmarks (SWE-Bench Pro, DeepSWE 1.0/1.1) — Fable 5 leads there
• Your users are in the EU — Grok 4.5 is not available there yet
• You need longer context than 500K tokens — the Inkling 1M context or Grok 4.20’s 2M window may fit better for very large codebases
• Knowledge cutoff matters: February 1, 2026 means events from the past five months are not in its training

The $2/$6 pricing opens Grok 4.5 to workloads where cost previously ruled out a frontier model. The agentic tool use advantage is real and documented. The benchmark shortfall versus Fable 5 on static coding tasks is also real. Neither story is wrong — they describe the same model at different points in the benchmark surface.