Grok 4.6 Matches GPT-4 and Claude 3 Opus Performance at a Much Lower Cost

Code Factory emphasized that Grok can implement projects at a very low cost. It said that the fees for implementing all of the projects it worked on used only 7% of the total budget, and added that this cost is hard for users to believe.

To evaluate Grok 4.6's performance, Code Factory set up a range of test items. It focused on how well the model implemented UI elements, how accurately it replicated functionality, and how well it handled complex components such as physics engines. Through this, it sought to verify Grok 4.6's overall development capability.

On cost, it has a very low pricing policy of $2 per million input tokens and $6 per million output tokens. These prices are highly competitive compared with other high-performance AI models.

Grok 4.6 shows dramatically lower costs than competing models. Fable is $10 input and $50 output, and Opus is $5 input and $25 output. GPT-4 Solo is $5 input and $30 output, Gemini Terra is $2 input and $12 output, and Luna is $0.2 input and $1.2 output. Compared with these, Grok 4.6 at $2 input and $6 output is confirmed to be one of the most economical options.

Opus sits in the highest tier, but Grok 4.6 falls in the 60-plus point range and can effectively be rated as a Tier 1 model. This means Grok 4.6 can stand alongside high-performance AI models.

This suggests Grok 4.6 is on par with these models in performance, and while it keeps pace with Fable Solo in particular, its cost is incomparably lower.

All projects were requested to use only native HTML, JavaScript, and CSS, and the model was instructed to build the physics engine from scratch. As a result, the cube's rotation and texture looked very natural, and the physics engine was also rated as feeling natural and well implemented. This project is a good example of Grok 4.6's ability to implement complex UI and physics engines.

Grok 4.6 also ran a test building a scroll-based website. The finished site shows a sundial concept in which the sun moves and time changes as you scroll. Interactive features were also implemented, such as text appearing or elements moving when the mouse moves. This shows Grok 4.6's proficiency in building dynamic web pages and user interactions.

All projects tested were implemented with a single prompt, meaning one request. This proves that Grok 4.6 can understand even complex instructions at once and produce high-quality results, demonstrating work efficiency.

Grok showed a strength in verifying its own output very diligently. However, it also revealed a limitation in that it could not fully catch certain errors, such as overlapping issues. Overall alignment was good and animations worked well, but it was assessed that there is room for improvement in overlooking minor errors.

All images used were generated by Grok itself. The project was completed in about 10 minutes, and its level of completion was high enough to make one think the real Airbnb had been copied directly, proving its cloning ability.

This was the project that took the longest among those Code Factory tested, about 40 minutes. The JavaScript file defining the physics engine was revised for a long time, but it was not implemented to a fully satisfactory level. This suggests Grok 4.6 still has weaknesses in implementing complex physics engines.

The next test was implementing a maze escape game controlled by mouse presses. This was also done with a single prompt, and the model quickly produced a game at the level YouTube BJs enjoy playing. Grok 4.6 received a positive assessment for being very capable at implementing a maze game of this difficulty.

Code Factory ran five projects simultaneously using all-in-one cards. Across all of this work, usage amounted to only 7%, because it used the $30 SuperGrok plan. Code Factory assessed that this $30 plan, in practice, provides roughly the same amount as the $100 plan. It stressed that although Grok is a top-tier model, its usage barely decreased, allowing users to make use of very large volumes.

Code Factory recommended Grok as a secondary account for users who are dissatisfied with using AI models such as GPT, Claude, Fable, and Opus, or whose budgets are tight. It further assessed that Grok's problem-solving and coding abilities are strong enough to serve as a primary account, and that it is fast. It strongly recommended Grok as a definite Tier 1 model for users who want to use many frontier models at a good price.