FULL STORY

Grok 4.6: From Early Cursor Leak to Hands-On Reviews

After briefly surfacing in Cursor ahead of launch, xAI's Grok 4.6 shipped with benchmark-leading scores and unchanged pricing; hands-on tests put its coding on par with top open-source models, though Opus-level claims remain debated.

2026-08-11 ~ 2026-08-15 · 5 episodes · 129 posts

Episode 1 · Grok 4.6 Spotted in Cursor Then Pulled, Unconfirmed by xAI (2026-08-11, 7 posts)

On Aug 11, multiple users spotted Grok 4.6 in Cursor's model picker, but it was quickly pulled, leaving only a non-clickable option. The listing showed 256k context and a 'strongest model' tag, yet xAI has released no official info. The incident fuels speculation about rapid iteration.

Confirmed

  • Grok 4.6 briefly appeared in Cursor's model picker on Aug 11, showing 256k context and a 'strongest model' tag, according to user reports.
  • The model was subsequently rolled back, leaving only a non-clickable option.

Unconfirmed

  • Whether the model is fully developed and imminent. @eyishazyer found no official info, and rumored specs like 1.5T parameters and 1M context lack official basis.
  • Whether this was an accidental early deployment by Cursor or a small-scale test by xAI remains undetermined.

Why it matters

  • If Grok 4.6 launches within weeks, it would mark a major acceleration in xAI's iteration pace, less than a month after Grok 4.5.
  • Developers have praised Grok 4.5's capability and efficiency, raising expectations for the new version.

Episode 2 · Grok 4.6 Released with Multi-Tool Coding Benchmark (2026-08-11, 4 posts)

xAI released the Grok 4.6 model, prompting developers to conduct comparative tests across mainstream coding tools like Cursor and Zed, evaluating features like planning modes and multimodal support.

Episode 3 · xAI Releases Grok 4.6 with Top Performance and Same Pricing (2026-08-12, 92 posts)

xAI officially released Grok 4.6 on August 12. The model achieved leading results on multiple benchmarks while maintaining the same pricing as the previous generation, making it a market focus for its high cost-performance. Musk and NVIDIA officially praised it, and the model is fully integrated into Grok Build, Cursor, Grok Bot, and the xAI API.

Confirmed

  • Benchmark performance: Grok 4.6 ranked first on the GDPVal-AA v2 benchmark with an ELO of 1753, surpassing Fable 5 Max and GPT-5.6 Sol Max. It scored 61 on the Artificial Analysis Intelligence Index, tying with GPT-5.6 Sol, behind only Claude Opus 5 and Claude Fable 5. It ranked first in the $5 budget workload test. It scored 94.9% on GPQA Diamond, beating GPT-5.6 and Gemini 3. On the AA-Briefcase agentic knowledge work benchmark, it scored on par with Claude Fable 5 but at one-fifth the cost per task.
  • Pricing strategy: Pricing is unchanged from Grok 4.5: $2 per million input tokens, $6 per million output tokens, and $0.50 for cached input. Analyst Gavin SB Baker noted that its performance is comparable to Fable 5 Max but at 85% lower overall cost (80% cheaper input, 88% cheaper output), even cheaper than Anthropic's Sonnet 5.
  • Platform integration: Designed for long-horizon agent tasks, it is integrated into Grok Build, Cursor, Grok Bot, and the xAI API. The official announcement said that for one week after release, Cursor and Grok Build users get double token credits.
  • Hardware support: NVIDIA officially confirmed that Grok 4.6 is trained and run on the GB300 NVL72 system, leveraging NVLink technology for superior performance and extremely low token cost.

Why it matters

In the current intense competition among large models, Grok 4.6 achieves what Musk hinted at as "Pareto gold." It not only demonstrates frontier intelligence in complex scenarios like real-world tasks and agentic knowledge work, but also dramatically lowers the barrier to top-tier AI models through aggressive pricing. This cost-performance approach directly challenges the market positioning of OpenAI and Anthropic, offering developers and enterprise users a highly cost-effective SOTA option.

72 more related posts →

Episode 4 · Grok 4.6 Benchmarks: Coding on Par with Top Open Models, Debate on Opus Level (2026-08-13, 16 posts)

Within days of Grok 4.6's release, multiple independent developers ran intensive tests. Consensus: its coding ability rivals top open models like Kimi K3, sufficient to take over Claude Sonnet 4.5's daily workloads, with faster speed and lower cost; Minecraft spatial building and video generation also improved. Dispute: whether it reaches Claude Opus level, with testers reaching opposite conclusions.

Confirmed

  • Value: scaling01 says Grok 4.6 slightly outperforms Kimi-K3 in both performance and efficiency despite being nearly half the size; MiaAIlab's 8-hour test in 'low' effort mode shows it matches Kimi K3 on kernel and modding tasks, with faster speed and lower token usage.
  • Coding positioning: bindureddy believes 4.6 reaches Sonnet level and surpasses Gemini overall; his team has migrated some Sonnet workloads. Its scores approach the top tier, and due to speed it is more practical than K3.
  • Spatial & multimodal: adonissingh found significant improvement in spatial judgment and aesthetics over 4.5 in a Minecraft 'Klein bottle on 3D axes' task; Aizkmusic built the Golden Gate Bridge in MC Bench in just two interactions, with the only error being a wrong block ID, and also produced a coherent flythrough video via two-shot generation.
  • Iteration & flaws: karminski3 measured backend Agentic Coding benchmark improvement under 5% over 4.5; frontend results pending. Maziyar Panahi notes a clear feel difference from 4.5, crediting xAI's mid-training and post-training. Aizkmusic warns coding is not yet at Fable's level, not recommended for production, and occasionally uses rare words like 'proscenium'.

Unconfirmed

  • Whether it reaches Opus level: sergeykarayev's Superconductor blog review places Grok 4.6 alongside Opus 4.8/5, Fable 5, and Kimi K3 on the 'pedestal' and as a daily driver; kevinnbass says it is nearly indistinguishable from Opus 5 in data mining workflows, with lower cost and faster speed, even better in some aspects; but bindureddy explicitly states it is far from Opus level, with capability ceiling below top closed models.

Why it matters

  • If the 'half size, slightly better performance' comparison holds, Grok 4.6 will directly challenge Sonnet-tier commercial workloads with speed and cost advantages, and at least one team has already migrated.
  • The disagreement among reviewers on the 'top tier' boundary shows that gaps among top models are now at the perceptual level, making selection more dependent on workflow-specific testing than benchmark scores.

Episode 5 · Grok 4.6 tops real-world coding benchmarks with standout cost efficiency (2026-08-14, 10 posts)

xAI's newly released Grok 4.6 puts in a strong showing across coding and reasoning benchmarks: Cursor's team said it ranks first on the CursorBench 3.2 real-world coding test; developer aidenybai measured it fifth overall on ReactBench and second on React repair; and a research team found its biology-benchmark accuracy roughly matches Opus 5 / GPT-5.6-Sol at significantly lower inference cost. Near-frontier intelligence combined with standout cost efficiency is the common theme in early feedback.

Confirmed

  • Cursor team members Lauren and Roshan said on a livestream that Grok 4.6 ranks first on CursorBench 3.2, with cost effectiveness far beyond Fable 5 Max.
  • Developer aidenybai reported Grok 4.6 ranks fifth on ReactBench for both writing and repair, showing strong React semantic understanding and a good Pareto curve with better effort separation than 4.5; it ranks second on the repair track specifically.
  • a16z partner Martin Casado, after several hours of use, said it is fast for its intelligence level, describes complex code cleanly, and completed hard tasks such as building a RAM probe.
  • User vikvang1 said Grok 4.6 offers Fable-like intelligence at Sonnet-like speed for a fraction of the cost, and no longer needs elaborate skill files because the model figures things out on its own.
  • A research team evaluated it on a short-term biology benchmark of 1716 trajectories, finding accuracy roughly on par with Opus 5 / GPT-5.6-Sol at significantly lower inference cost.

Unconfirmed

  • Several posts mention model names such as Claude Fable 5, Fable 5 Max and GPT-5.6 Sol, including a parody leaderboard from Perplexity claiming Grok 4.6 matched Claude Fable 5's 0.496 score on a WANDR benchmark at $7.58 per task; with fictional names mixed in, official results should be separated from informal comparisons.

Why it matters

  • Grok 4.6 suggests xAI is closing in on the frontier in code generation and complex reasoning, and its price-performance could reshape developers' model choices; feedback that heavy skill wrapping is no longer needed hints that stronger model autonomy may simplify agentic workflows.