Grok 4.8 Development delayed; Musk says benchmark scores are not the goal

Article

Grok 4.8 Still not released; Musk says he does not care about benchmark scores, only about being genuinely useful

Beating AI Flash, Musk October 30 acknowledged that Grok 4.8 has taken longer to develop than expected, but assured that the new model “will not disappoint.” He said the team is not optimizing for benchmark tests, and that he does not care about benchmark scores, placing greater importance on the model’s practical utility.September 14 Musk previously disclosed that Grok 4.8 has 2500000000000 parameters and uses a completely new C++ training system. Pre-training was expected to conclude by the end of that week, followed by reinforcement learning. Now, more than 6 weeks have passed, and the new model has still not been officially released.

Musk emphasized that he does not care about benchmark scores because the performance of Grok 4.7 is simply not worth boasting about. Independent evaluation firm Artificial Analysis found that, at the same xhigh reasoning setting, the Grok 4.7 composite intelligence index was only 4.6 points higher than 2(44→46); meanwhile, average output per question for Token increased from approximately 38000 to 81000. Users in the Cursor community also complained that 4.7 consumes usage credits faster, while its practical improvements are not obvious.September During an interview with China Central Television, Musk had already acknowledged that Grok 4.7 lags behind Claude Opus 5.5, and estimated that it may not catch up with the leading tier until 2027.October He also proposed having Grok Bot call external models, including Claude, depending on the task, and plans to have the high-speed version of Grok 4.8 handle large volumes of simple requests.

Original link https://m.theblockbeats.info/flash/374615

Key points

Musk October 30 said that Grok 4.8 has taken longer than expected to develop and that the new model will not disappoint.
Musk said the team does not optimize for benchmark tests and places greater importance on the model’s practical utility.
Musk previously September 14 disclosed that Grok 4.8 has 2500000000000 parameters and uses a completely new C++ training system; it was expected to enter reinforcement learning after pre-training concluded by the end of that week.
The evidence states that Grok 4.8 still had not been officially released more than 6 weeks after the related information was published.
Artificial Analysis found that, at the same xhigh reasoning setting,Grok 4.7’s composite intelligence index was 4.6 points higher than 2, rising from 44 points to 46 points; average output per question Token increased from approximately 38000 to 81000。
The evidence states that users in the Cursor community complained that Grok 4.7 consumes usage credits faster, but its practical improvements are not obvious.
Musk previously acknowledged that Grok 4.7 lags behind Claude Opus 5.5, and estimated that 2027 may not catch up with the leading tier until that year.
Musk proposed having Grok Bot call external models, including Claude, depending on the task, and plans to have Grok 4.8’s high-speed version handle large volumes of simple requests.

AI-assisted interpretation

The following is analysis, separate from reported facts. Verify important claims independently.

This news concerns Grok 4.8’s development progress and product direction. Musk said development has been delayed and that the team does not regard benchmark scores as its primary objective, instead prioritizing actual performance in use.

Why it matters to readers

Grok 4.8 has not yet been officially released; therefore, its actual capabilities, speed, cost, and availability still require confirmation through subsequent information and evaluations.

Benchmark scores are a model’s results on standardized tests; parameter count and Token count are commonly used to describe model scale and output length, respectively, but neither alone can prove that a model is more useful.

Risks and unknowns

The evidence comes from media reporting and does not include Musk’s original remarks or an official release.
Grok 4.8’s official release date, model specifications, and availability have not been confirmed in the evidence.
2500000000000 Information such as parameter count, the training system, and pre-training progress has not been independently verified in the evidence.
The complete methodology and samples for the Artificial Analysis evaluation, Cursor community feedback, and related comparisons were not provided.
Grok Whether the plan to call external models will be implemented, and the specific partnership arrangements, are unknown.

Grok 4.8 Development delayed; Musk says benchmark scores are not the goal

Article

AI-assisted interpretation

Why it matters to readers

Risks and unknowns

Related Developments

Related concepts

Keep following

Security Center