Introduction
Recently, DeepSeek released the open‑source tool DeepSeek Harness, betting on low prices and developer access. Almost simultaneously, SpaceX AI released Grok Bot, betting on real‑time information and agent execution. Claude Code and Codex, as established strong players in the same field, remain fiercely competitive.
Competition Has Shifted from Comparing Models to Capturing Entry Points
The competition among large AI models is no longer just about leaderboards. Today, OpenAI and Anthropic still hold high‑end capabilities, but they also have to compete for developer entry points, tool permissions, and long‑term tasks. DeepSeek focuses on low prices, long contexts, and interface compatibility, while Grok bundles real‑time information, the X ecosystem, and agent functions.
DeepSeek’s recent updates have been intensive. On July 31, the official API documentation announced the public beta of V4 Flash, and on August 13, V4 Pro was released. Both versions provide 1M context and integrate with the Responses API and Codex. The official announcement also stated that the API will be billed according to peak/off‑peak periods starting at 16:00 Beijing time on August 16. But for developers, cheap models are only the first step; whether they can be stably integrated into existing workflows is more important.
Musk’s xAI, on the other hand, seems to be expanding product entry points. After releasing Grok 4.5 on July 16, Grok successively entered web, X, and mobile platforms, and also launched Workflows and BuildMode. Regarding the latest model, online media reported on August 13 about the latest developments of Grok 4.6, but as of this writing, publicly available comparable evaluations are still mainly based on Grok 4.5.
The Four Products Are Not in the Same Race
As we all know, Claude Code and Codex are pioneers in AI coding tools, but Claude Code is closer to a local terminal, while Codex is more like an agent that executes tasks remotely. DeepSeek Harness is an open‑source chassis, requiring you to combine terminals, file systems, web, and sub‑agents yourself. Grok Bot combines real‑time information and automated execution.
Pricing strategies are also quite telling given their different product forms.
Claude Code’s individual plans: Pro at $20/month, Max 5x at $100/month, Max 20x at $200/month.
ChatGPT Plus is $20/month, Pro is $200/month.
Grok Bot makes a splash; it is not included in the regular Grok chat bundle, nor does it include paid xAI API. Public access information shows that Grok Bot is currently available through SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium, with reference prices of $300/month for individuals, and $200/month or $120/seat for enterprises and teams; there is no free tier. The previously available SuperGrok at $30/month and SuperGrok Plus at $100/month are regular Grok chat subscriptions and do not qualify for the Grok Bot pricing tier.
Model Leaderboard Rankings
In the public snapshot as of August 12, the Arena Agent rankings are:
Claude Opus 5 High – #1 with net improvement score 12.01
GPT 5.6 Sol xHigh – #4 with 10.80
Grok 4.5 – #15 with 6.00
DeepSeek V4 Flash High – #18 with 3.89
V4 Pro – #27 with no score yet
This ranking measures model‑plus‑configuration performance on real tasks and cannot be directly treated as an overall ranking of the four products.
In the Arena Text open dialogue rankings:
Claude Opus 5 High – #7 with 1494 points
GPT 5.6 Sol xHigh – #18 with 1481 points
Grok 4.5 – #34 with 1469 points
DeepSeek V4 Pro – #49 with 1458 points
In the same SWE‑bench Verified snapshot, Claude Opus 4.8 scored 88.6%, DeepSeek V4 Pro 80.6%, GPT 5.4 80.0%, and Grok 4 75.0%. Scores indicate a model’s position in certain tasks, but they cannot save consumers from considering factors such as permission management, testing, and user experience when choosing.
Claude Code and Codex: One On‑Premise, One in the Cloud
As the current leader among AI coding tools, Claude Code’s strength lies in continuous collaboration: reading files, modifying code, running tests, and then continuing to modify based on results. It is suitable for developers who work long‑term in a single repository, but the cost is that you need to manage terminal permissions, context, and usage yourself.
Codex places tasks in an isolated environment, suitable for diagnosing issues, writing patches, and running verification, then returning to check results. It is more like a remote colleague; the clearer the task breakdown, the more stable the delivery.
Neither is a black box that delivers at the press of a button. Online configuration, permission changes, and data operations still require developer confirmation.
DeepSeek Harness: Low Price, but You Build the Base Yourself
The appeal of DeepSeek Harness lies in its low cost, long context, and replaceable models. V4 Pro scored 80.6% on SWE‑bench Verified, but ranked 27th in the public Arena Agent snapshot. This gap shows that offline fixed‑repair scores are not the same as real task experience. It is suitable for highly customizable users who are willing to install, configure, and troubleshoot themselves, not for beginners or those who treat it as out‑of‑the‑box software.
Grok Bot: Not a Chat Window
According to the official product description, Grok Bot is currently in Early Beta. It can log into user tools and execute persistent tasks in the cloud, effectively providing a cloud development environment. Workflows, Automations, and BuildMode have already moved Grok from real‑time search toward application execution, but the scope of permissions and actual accessible tools still determine its upper limit. This is also why Grok Bot’s price is higher than ordinary Grok subscriptions. SuperGrok Heavy, Cursor Ultra, and Cursor Teams Premium are reference access bundles, and they are completely different from previous chat and API plans.
How to Choose? Depends on Your Work
If you primarily write code in a local repository, Claude Code is more convenient. If you need to asynchronously handle a batch of tasks, Codex is more suitable. If you want to build your own agent with low‑cost models, you can try DeepSeek Harness, but set aside time for configuration and maintenance. If you need to track real‑time information and then hand tasks to the cloud for execution, Grok Bot’s direction is more aligned.
It is hard to pick an overall champion among the four. My recommended combination: Grok for finding real‑time information, DeepSeek for low‑cost experimentation, and Claude Code or Codex for implementation and verification.