SpaceXAI and Cursor ship Grok 4.5, an "Opus-class" coding model at a fraction of the price

incrementalDeveloperEnterprise

Thursday, July 9, 2026

Confidence

Medium · — vendor-published benchmarks with an admitted contamination flag; no independent third-party eval yet

Evidence

lab announcement + Cursor co-published post + Axios/Reuters/Engadget reporting

SpaceXAI released Grok 4.5 on Tuesday as its first model since the SpaceX-xAI merger and Cursor acquisition — a mixture-of-experts model trained jointly with Cursor and priced well under Anthropic's flagship.

The details

Cursor confirms the model was trained jointly with SpaceXAI on trillions of tokens of Cursor data capturing user interactions with codebases and developer-agent interactions , and ships in Cursor across desktop, web, iOS, CLI, and its SDK at $2/M input and $6/M output tokens.
Axios reports Musk pitched it as "an Opus-class model, but faster, more token-efficient and lower cost," with a chart claiming it outperforms Opus 4.8 on several benchmarks . Reuters notes Anthropic's Opus 4.8 is priced at $5/M input and $25/M output — roughly a 4x price gap.
The self-published sheet is mixed: on DeepSWE 1.0 and Terminal-Bench 2.1 Grok 4.5 leads Opus 4.8; on DeepSWE 1.1 and SWE-Bench Pro it trails; Anthropic's Fable 5 leads several charts . Cursor also disclosed that an earlier snapshot of the Cursor codebase was accidentally included in training, contaminating its CursorBench score .

What it means

This is a pricing story dressed as a capability story. Grok 4.5 is competitive but does not clearly beat the frontier — it beats the frontier's price. That matters for engineering leads whose coding-agent bill scales with tokens: at 4x cheaper input and roughly 4x cheaper output than Opus 4.8, the model is defensible as a default even if it lands second on any given benchmark. The CursorBench contamination admission is the honest note buried in the launch; treat every self-reported number here as an upper bound, not a floor. The bigger structural signal is vertical bundling — a coding-tool vendor and a frontier lab co-trained a model, and it now ships as Cursor's first-party option. Expect Anthropic and OpenAI to counter with tighter IDE-native tuning rather than more benchmarks.

Cursor confirms the model was trained jointly with SpaceXAI on trillions of tokens of Cursor data capturing user interactions with codebases and developer-agent interactions , and ships in Cursor across desktop, web, iOS, CLI, and its SDK at $2/M input and $6/M output tokens.
Axios reports Musk pitched it as "an Opus-class model, but faster, more token-efficient and lower cost," with a chart claiming it outperforms Opus 4.8 on several benchmarks . Reuters notes Anthropic's Opus 4.8 is priced at $5/M input and $25/M output — roughly a 4x price gap.
The self-published sheet is mixed: on DeepSWE 1.0 and Terminal-Bench 2.1 Grok 4.5 leads Opus 4.8; on DeepSWE 1.1 and SWE-Bench Pro it trails; Anthropic's Fable 5 leads several charts . Cursor also disclosed that an earlier snapshot of the Cursor codebase was accidentally included in training, contaminating its CursorBench score .

Sources

SpaceXAI and Cursor ship Grok 4.5, an "Opus-class" coding model at a fraction of the price

Turn a Raw Expense Report Batch into an AI-Fraud Risk Memo

Gemini API Managed Agents

One daily brief. Every story gets a hype verdict.