On July 8, 2026, SpaceXAI and Cursor jointly released Grok 4.5, a new frontier model that directly targets enterprise coding and agentic workloads. The key claim from Elon Musk: “an Opus-class model, but faster, more token-efficient and lower cost.” The key number: $2 per million input tokens, $6 per million output tokens, against Anthropic Opus 4.8’s $5 and $25.
This is SpaceXAI’s first major model release since going public and completing its acquisition of Cursor, the AI coding editor valued at $60 billion. The release is also the first product to visibly combine SpaceXAI’s compute infrastructure with Cursor’s developer-interaction training data. For enterprise teams choosing between frontier models, it changes the pricing arithmetic for agentic workloads.
What Grok 4.5 Is: Architecture and Training
Grok 4.5 is a mixture-of-experts (MoE) model trained jointly across SpaceXAI’s Memphis data centers on tens of thousands of NVIDIA GB300 GPUs. The training drew on two major data sources: curated STEM, science, and engineering corpora from SpaceXAI, and trillions of tokens of Cursor interaction data capturing how developers actually work inside codebases.
That second source is the unusual part. Most frontier models learn from static code repositories. Grok 4.5’s Cursor data includes records of developer-agent interactions: the sequences of edits, tool calls, verifications, and recoveries that occur during real software work. Cursor’s team explains this trains the model to “investigate problems, use tools, recover from mistakes, and verify results” rather than simply predict the next token in a code file.
Reinforcement learning ran on hundreds of thousands of tasks spanning software engineering and knowledge work. SpaceXAI used a distributed agent system to generate these training environments at scale, with earlier model versions used to construct problems difficult enough to push the next generation. This self-reinforcing loop is now common across frontier labs, but the Cursor codebase and workflow data gives Grok 4.5 a different distribution from models trained on public repositories alone.
One note on benchmarks: Cursor disclosed that an earlier snapshot of its own codebase was accidentally included in Grok 4.5 training, creating a data leakage issue on CursorBench. The team removed that data for future models and is rebuilding the benchmark. Scores on CursorBench from this release should be disregarded.
How Grok 4.5 Performs: Benchmark Snapshot
SpaceXAI published benchmark comparisons at launch using each provider’s own harness, so direct cross-lab comparisons have caveats. That said, the numbers give useful directional signal for enterprise buyers.
| Benchmark | Grok 4.5 | Fable 5 Max | GPT-5.5 xhigh | Opus 4.8 Max |
|---|---|---|---|---|
| DeepSWE 1.0 (pass@1) | 62.0% | 66.1% | 64.31% | 55.75% |
| Terminal Bench 2.1 | 83.3% | n/a | n/a | n/a |
| SWE Bench Pro | 64.7% | n/a | n/a | n/a |
On DeepSWE 1.0, Grok 4.5 lands above Opus 4.8 and below Fable 5. That benchmark was created by Datacurve and run with each provider’s harness, so the methodology is consistent enough for a rough comparison even if not perfectly controlled.
The more actionable figure for budgeting teams is token efficiency. SpaceXAI reports Grok 4.5 uses 4.2 times fewer output tokens than Opus 4.8 on SWE Bench Pro tasks and delivers output at 80 tokens per second. For agentic loops, where a single task can generate thousands of output tokens across multiple steps, a 4x efficiency difference compounds quickly in cost.
The Pricing Case for Enterprise Teams
The base pricing table for current Opus-class and near-Opus models is now:
| Model | Input (per 1M tokens) | Output (per 1M tokens) |
|---|---|---|
| Grok 4.5 | $2.00 | $6.00 |
| Grok 4.5 Fast | $4.00 | $18.00 |
| Claude Sonnet 5 (standard) | $3.00 | $15.00 |
| GPT-5.6 Luna | $1.00 | $6.00 |
| Claude Opus 4.8 | $5.00 | $25.00 |
At a simple level, Grok 4.5 costs 40% less than Opus 4.8 on input and 76% less on output. Apply the reported 4.2x token efficiency advantage and the effective cost per completed SWE Bench Pro task would be substantially lower still, though teams should verify this on their actual workloads rather than accepting benchmark extrapolations directly.
The comparison relevant to teams already using Claude Sonnet 5 is different. Sonnet 5 at standard pricing is $3 input and $15 output, and it sits close to Opus 4.8 on many agentic benchmarks. Grok 4.5 is cheaper on input but comparable on output at the standard tier. The deciding variable becomes which model completes the actual tasks your agents run with fewer tokens and fewer steps.
The Cursor Distribution Flywheel
What makes this release strategically significant beyond the benchmark numbers is the distribution model. Grok 4.5 is the default model in Cursor, available on all plans (individual and team) starting July 8, with usage doubled through July 15.
Cursor is one of the most widely adopted AI coding tools in production engineering teams. When SpaceXAI acquired Cursor, it acquired a channel where developers evaluate models inside real projects every day. That channel now defaults to Grok 4.5. Without requiring any decision from a developer or IT team, SpaceXAI gains comparative production usage at scale.
This mirrors a pattern other labs have chased through partnership rather than acquisition. Anthropic distributes Claude through enterprise workspace integrations like Slack and desktop tools. Microsoft routes its MAI family and OpenAI models through M365 Copilot and Azure. SpaceXAI is taking a different path: owning the coding interface directly, then routing model traffic through it.
For enterprise buyers, this matters because Cursor integrations are often approved at the team or department level rather than through central IT procurement. Grok 4.5 may gain wide internal usage before it appears on any official vendor list.
The model is also available across major infrastructure gateways: OpenRouter, Vercel, Cloudflare, Snowflake, and Databricks Mosaic, as well as Word, PowerPoint, and Excel Office add-ins. This coverage means teams can access Grok 4.5 through existing AI gateway contracts without a new vendor relationship.
What Enterprise Teams Should Know Now
On access: Grok 4.5 is live via the SpaceXAI API console and Cursor today. EU teams will need to wait for mid-July availability. API keys are available from the SpaceXAI console with no waitlist.
On evaluation: The most meaningful test is a cost-per-task comparison on your actual agent workflows, not benchmark leaderboard position. SpaceXAI’s 4.2x token efficiency claim is compelling but should be validated against the specific task types your agents run. Loop length, tool call frequency, and context reuse patterns all affect effective cost substantially.
On governance: SpaceXAI added cybersecurity safeguards for this release and Cursor notes new controls for its cybersecurity capabilities. For regulated environments, review the system card and confirm EU compliance before any production deployment.
On the competitive context: This release lands the same week as the full public launch of OpenAI’s GPT-5.6 family. The frontier model market now has at least four providers shipping Opus-class or near-Opus performance at prices ranging from $1 to $5 per million input tokens. The enterprise AI adoption challenge has shifted from “can we afford frontier models?” toward “how do we evaluate and govern multiple competitive options running in parallel?”
Enterprises building serious agentic workflows in mid-2026 should be running parallel evaluations rather than committing to a single-provider strategy. Grok 4.5 gives those evaluations a new, credible, competitively priced option, particularly for teams that are already in the Cursor ecosystem.
Sources: SpaceXAI Grok 4.5 announcement (July 8, 2026); Cursor Grok 4.5 blog post (July 8, 2026); SpaceXAI developer docs; TechCrunch: SpaceXAI releases Grok 4.5 (July 8, 2026); The Next Web: SpaceXAI launches Grok 4.5 with Cursor (July 8, 2026).