No sooner have companies begun integrating GPT-5.5, Claude Sonnet 5, or GLM-5.2 than a new competitor has already entered the race. SpaceXAI, Elon Musk’s artificial intelligence company, has just officially unveiled Grok 4.5, a next-generation model designed primarily for software development, complex reasoning, and agent-based tasks. Behind this announcement lies a much broader ambition: to offer artificial intelligence capable of carrying out complete business missions while reducing operating costs.
After several weeks of private beta testing, Grok 4.5 is now beginning to roll out to developers and businesses. SpaceXAI claims it is its most powerful model to date, trained on the Colossus supercomputer and capable of competing with the best models from Anthropic, OpenAI, and Google.1
SpaceXAI is now focusing on professional applications
The early versions of Grok had primarily attracted attention due to their integration with the social network X and their more conversational tone compared to other AI assistants. With Grok 4.5, SpaceXAI is clearly shifting its strategy. The goal is no longer just to offer a high-performing chatbot, but to develop a true enterprise-grade model capable of automating complex tasks.
The new model primarily targets three areas: software development, agent-based workflows, and knowledge work. Specifically, Grok 4.5 is capable of analyzing an IT project, writing code, using various tools, planning several successive steps, and then correcting some of its own errors before producing its final response. This autonomy is gradually bringing Grok closer to intelligent agents, which are now becoming the new battleground for competition among major artificial intelligence research labs.
According to Elon Musk, Grok 4.5 belongs to the “Opus” category of models—that is, state-of-the-art models designed for the most demanding applications. This development confirms that SpaceXAI now intends to compete directly with Claude Opus 4.8, GPT-5.5, and Google’s upcoming professional models.
A massive architecture to speed up reasoning
To achieve this level of performance, Grok 4.5 is based on an architecture with approximately 1,500 billion parameters, trained using the Colossus supercomputer, one of the largest computing infrastructures dedicated to artificial intelligence currently in operation.1
This computing power enables the model to handle particularly complex reasoning tasks while maintaining a high execution speed. SpaceXAI reports, in particular, a generation rate of up to approximately 80 tokens per second—a throughput that significantly improves fluidity during time-consuming tasks such as programming or document analysis.
Beyond speed, Grok 4.5 aims above all to optimize the use of its resources. The company claims that its model uses up to four times fewer tokens than Claude Opus 4.8 to perform certain comparable tasks. If this claim holds true in real-world use, the total cost of ownership could become one of SpaceXAI’s key selling points to businesses.
The benchmarks show that this model is very competitive
Like all artificial intelligence labs, SpaceXAI is accompanying its launch with a series of benchmarks designed to measure Grok 4.5's performance.
On Terminal-Bench 2.1, a metric that measures a model’s ability to operate directly in a command-line interface, Grok 4.5 achieved a score of 83.3%, compared to 78.9% for Claude Opus 4.8.2
The model also achieved a score of 62% on DeepSWE 1.0, a benchmark that evaluates the ability to solve complex software development tasks, compared to 55.8% for Claude Opus 4.8.
These results demonstrate significant progress, but they do not necessarily mean that Grok outperforms all of its competitors. Other independent evaluations show that Claude Opus 4.8 continues to lead, in particular, on SWE-Bench Multilingual and SWE-Bench Pro—two major benchmarks for measuring an AI’s ability to fix real bugs in complex software projects.3
This situation perfectly illustrates current market trends. The performance gaps between the top models are narrowing, and each lab is now seeking to differentiate itself through its costs, integration with professional tools, or agent-based capabilities rather than through a few extra points on a benchmark.
AI designed for developers… but not just for them
One of Grok 4.5's strengths is its rapid integration into the development environments that companies already use.
The model is already available on Grok Build, SpaceXAI’s development platform, as well as on Cursor, one of the most popular AI-assisted programming environments today.4 Developers can also use it via the dedicated console and with an API key that allows them to directly integrate Grok 4.5 into their own applications.
However, the international rollout is taking place gradually. Early adopters in the United States already have access to the full range of features, while the service is being rolled out to European Union countries in phases. For France, SpaceXAI plans a gradual rollout in mid-July, subject to ongoing regulatory approvals.4
This strategy is similar to the one recently adopted by Anthropic and OpenAI, which are now prioritizing phased rollouts in order to fine-tune their models before a full global launch.
An exceptionally competitive price-to-performance ratio
Another argument put forward by SpaceXAI concerns its pricing strategy.
API access is currently offered at $2 per million tokens for incoming transactions and $6 per million tokens for outgoing transactions, which is a particularly competitive price compared to the most advanced models on the market.5
Beyond the listed price, SpaceXAI places particular emphasis on the actual cost of the tasks performed. While Grok 4.5 does indeed use fewer tokens to solve the same problem, the final cost to businesses could be significantly lower than that of some competitors, even though their listed rates are similar.
This approach is specifically designed for organizations that process several hundred thousand AI queries daily and for which inference costs now account for a significant portion of their digital spending.
The battle is no longer just about the models
The launch of Grok 4.5 marks a significant shift in the artificial intelligence industry. For a long time, research labs focused primarily on developing the highest-performing model. Today, competition also centers on platforms, development tools, autonomous agents, and usage costs.
In this context, Grok 4.5 is not just a new language model. It is one of the building blocks of a much larger ecosystem that includes Cursor, Grok Build, X, the Colossus infrastructure, and, eventually, the future AI agents developed by SpaceXAI.
The company is thus joining a trend already seen at Anthropic with Claude Code and Claude Tag, at OpenAI with Codex and Daybreak, and at Google with Gemini Enterprise.
Ethical Issues: Ever-Greater Autonomy, but Under Control
Like other frontier models, Grok 4.5 raises several ethical questions. The more capable artificial intelligence systems become at using tools, performing complex tasks, or making intermediate decisions, the more essential oversight mechanisms become.
SpaceXAI claims to have strengthened safeguards against malicious use, particularly in sensitive areas related to cybersecurity. Nevertheless, as recent debates surrounding Claude Fable 5 and GPT-5.5-Cyber have shown, the most powerful models are now the subject of close scrutiny by public authorities.
Beyond security, the widespread adoption of these agent-based models also raises the issue of accountability. When artificial intelligence autonomously executes a complete sequence of actions, it becomes essential to maintain a record of decisions, the tools used, and human interventions in order to ensure responsible use in professional environments.
SpaceXAI Confirms Its Ambitions
With Grok 4.5, SpaceXAI is demonstrating that it no longer wants to merely participate in the race for artificial intelligence. The company now aims to establish itself as a major player in the professional market, capable of offering a credible alternative to models from Anthropic, OpenAI, or Google.
This new generation of models also confirms a strong trend: the future of artificial intelligence will no longer hinge solely on raw performance. Companies will choose their models based on a balance between the quality of reasoning, autonomy, ease of integration, speed of execution, and actual cost of use. In this regard, Grok 4.5 brings particularly strong arguments to the table.
How does Grok 4.5 work?
Grok 4.5 is a language model developed by SpaceXAI, designed to meet the needs of developers, businesses, and future agent-based artificial intelligence systems. Unlike earlier generations of conversational models, which were primarily focused on text generation, Grok 4.5 has been trained to solve complex tasks involving programming, logical reasoning, and process automation. It is part of the new generation of so-called “frontier” models, capable of using tools, planning multiple work steps, and executing tasks much more autonomously.
The model is based on a very large-scale architecture comprising approximately 1,500 billion parameters, trained on Colossus, the supercomputer developed by SpaceXAI specifically for next-generation artificial intelligence models. This infrastructure enables the model to analyze very large volumes of information, perform multiple lines of reasoning in parallel, and generate responses quickly, even for complex tasks such as software development or technical analysis.
One of the key new features of Grok 4.5 is its agent-based operation. When a user submits a request, the model no longer simply generates an immediate response. It automatically breaks down the task into several subtasks, selects the appropriate tools, performs various operations, and then verifies some of its own results before generating the final response. Optimized for development environments, it can also analyze a code repository, understand a project’s dependencies, identify the source of a bug, suggest fixes, and generate new code in multiple programming languages.
- Advanced Reasoning: Solving Complex Problems Through Multi-Step Planning
- Software development: code generation, analysis, correction, and optimization in multiple languages
- Agent-based AI: autonomous execution of action sequences using various digital tools
- Using Development Environments: Integration with Grok Build, Cursor, and Enterprise APIs
- Fast execution: can generate up to approximately 80 tokens per second, according to SpaceXAI
- Cost Optimization: Reducing token consumption to limit the actual cost of transactions
- Very Large-Scale Architecture: Training on the Colossus supercomputer to improve performance on complex tasks
- The best performance is primarily seen in programming tasks and professional applications
- Some independent benchmarks show that Claude Opus 4.8 maintains an advantage in several specialized evaluations
- International rollout continues to proceed gradually, depending on the region and regulatory constraints
- Agent-based capabilities still require human oversight for sensitive operations
- Like all large language models, Grok 4.5 may produce hallucinations or errors in reasoning in certain complex situations
- Uses related to cybersecurity or critical infrastructure remain subject to specific security policies
Learn more
The launch of Grok 4.5 confirms that the battle among large artificial intelligence models is now shifting toward professional applications and autonomous agents. On a related topic, check out our article “Anthropic Makes a Big Splash: Claude Sonnet 5 Offers Nearly the Power of Opus at Half the Price , ” which analyzes how Anthropic, too, is seeking to make state-of-the-art models more accessible to businesses.
References
1. SpaceXAI. (2026). Introducing Grok 4.5.
https://spacex.ai
2. Terminal-Bench. (2026). Terminal-Bench 2.1 Leaderboard.
https://terminalbench.ai
3. SWE-Bench. (2026). SWE-Bench Leaderboard.
https://www.swebench.com
4. Cursor. (2026). Grok 4.5 Integration Announcement.
https://cursor.com
5. SpaceXAI Developers. (2026). Grok API Pricing.
https://developers.spacex.ai