Grok 4.5 Matches Opus 4.7 While Cutting Token Use to One‑Quarter

In June 2026 xAI announced the launch of Grok 4.5, its first flagship model after integrating SpaceX and acquiring the AI‑coding tool Cursor. The model is positioned as a direct competitor to Claude’s Opus series, emphasizing faster inference and lower token costs.

Grok 4.5 is a 1.5 T‑parameter mixture‑of‑experts (MoE) model, three times larger than its predecessor Grok 4.3. It supports up to 500 k context tokens, visual inputs, and is targeted at programming, agent, and knowledge‑work scenarios.

Performance claims include matching Opus 4.7 on the DeepSWE Bench 1.0 benchmark and surpassing Opus 4.8 on the same suite. The model reaches up to 80 TPS inference speed and is priced at $2 per million input tokens and $6 per million output tokens, more than 60 % cheaper than Claude’s Opus series and GPT‑5.5. On the SWE Bench Pro, Grok 4.5 consumes less than one‑quarter of the tokens required by Opus 4.8 Max for the same tasks.

Official benchmark results show scores of 83.3 % (DeepSWE 1.0), 78 % (DeepSWE 1.1), 62 % (Terminal Bench 2.1) and 64.7 % (SWE Bench Pro), placing the model close to GPT‑5.5 and Claude Opus while offering markedly lower inference cost. Third‑party evaluations from Artificial Analysis rank Grok 4.5 in the second tier for overall intelligence but at the forefront for speed, token efficiency, and cost‑performance.

Demo examples include a generated solar‑system image, high‑quality PPT slides, a weather‑app UI, game UI mock‑ups, a Mario‑style game, and an interactive web page with liquid‑glass effects. Community members report that Grok 4.5 feels comparable to Fable at roughly one‑seventeenth of its cost.

The efficiency gains stem from a redesign of the training pipeline. Grok 4.5 was trained on tens of thousands of NVIDIA GB300 GPUs with extensive stability optimizations. Cursor data underwent large‑scale deduplication, quality filtering, and domain‑specific selection, adding tens of trillions of tokens that capture not only code but also developer‑tool‑agent interactions. Reinforcement learning was expanded to cover hundreds of thousands of real software‑engineering and knowledge‑work tasks, requiring multi‑step reasoning, tool invocation, error recovery, and result verification. An asynchronous agent‑training system allows the model to continue learning while executing long‑running complex tasks.

Future releases are planned: a million‑context version next week and a 2 T‑parameter model by the end of the month. The current Grok 4.5 does not yet use xAI’s internally optimized C/C++ inference stack; once integrated, Musk predicts the runtime speed could double again.

The article notes a broader industry shift: as model capabilities converge, competitive advantage increasingly depends on system‑level engineering—training frameworks, inference engines, and hardware co‑design. Grok 4.5’s roadmap signals xAI’s intent to join this engineering race.

Overall community feedback is positive, highlighting Grok 4.5’s notable improvements in speed, token efficiency, and cost while approaching Opus‑level performance.