Grok 4.6

Scroll sideways for more →

Grok 4.6 is a proprietary large language model and reasoning model in the Grok family, developed by SpaceXAI and released jointly with Cursor through the xAI API on August 12, 2026. It succeeds Grok 4.5 at the same $2 and $6 per-million-token prices, keeps that model's 500,000-token context window and February 2026 knowledge cutoff, and is positioned by both companies for coding, agentic tasks, and knowledge work, with what the launch post calls "a particular focus on long-running agents and more ambitious interactive and visual work."^[1][3][4][8]

The release arrived five weeks after Grok 4.5 and closed a publicly telegraphed schedule: Elon Musk had posted the model's timing, claimed parameter count, and training approach on X in late July, then repeated the schedule on SpaceX's first earnings call.^[10][11][12] On Artificial Analysis' Intelligence Index, the most-cited independent measure, Grok 4.6 at high reasoning effort scored 61 on launch day, up five points on Grok 4.5 and tied with OpenAI's GPT-5.6 Sol at maximum effort, behind only Anthropic's Claude Opus 5 and Claude Fable 5 configurations.^[14][15]

Announcement and release

The launch was unusual in how much of it Musk narrated in advance. On July 24, 2026, he posted a chart claiming that "Grok 4.5 and Opus 5 are alone on Pareto frontier," and replied to his own post: "Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks."^[10] Four days later he added specifics: "Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL. Grok 4.7 will be the 2.1T model released a few weeks later."^[11] On SpaceX's first earnings call as a public company, held August 4, 2026, he updated the schedule again, telling investors that Grok 4.6 would arrive the following week, that Grok 4.7 would follow three to four weeks after the call, and that Grok 5, expected around the end of the year, would incorporate SpaceX's own engineering data.^[12][13]

The "around August 7" estimate passed without a release. Grok 4.6 shipped on August 12, 2026, announced at 15:32 UTC in a post from the @SpaceXAI account: "Introducing Grok 4.6. It delivers frontier intelligence and is a significant improvement over Grok 4.5 at the same price."^[9] SpaceXAI's release notes record API availability the same day, and Cursor published a matching launch post, opening "Today we are releasing Grok 4.6 together with SpaceXAI."^[3][8] As with Grok 4.5, the model reached developer surfaces first: the announcement lists the xAI API, Grok Build (where it became the default model), Cursor on all plans, and the OpenRouter, Vercel, and Cloudflare gateways.^[1][4] Both companies offered double the included usage inside Grok Build and Cursor for the first week.^[1][8]

The joint release continues the arrangement behind Grok 4.5, which Cursor and SpaceXAI trained together after SpaceX agreed to acquire Cursor's parent company Anysphere; the corporate background is covered at Grok 4.5 and xAI. Cursor's Grok 4.6 post is published under the Cursor Team byline and its site still carried an "Anysphere, Inc." copyright notice on launch day.^[8]

Model and training

SpaceXAI has not formally published an architecture description or parameter count. The 1.5 trillion figure comes from Musk's July 28 post, which described Grok 4.6 as "the 1.5T model," the same size he had previously attached to Grok 4.5; his framing, and the launch post's, is that the gains over Grok 4.5 come from training rather than scale.^[11][1] No model card accompanied the launch-day materials; for Grok 4.5, the model card followed six days after release.^[1][8]

The launch post describes three stages of work on top of Grok 4.5's recipe. First, a "longer supplemental training run" used curated model-generated data for reasoning and advanced technical concepts, high-quality engineering data, and what the company calls an improved optimizer and training recipe. Second, SpaceXAI used Grok 4.5 itself to regenerate the supervised fine-tuning trajectories across reasoning efforts, agent harnesses, and domains such as STEM, software engineering, and knowledge work, filtering out problematic traces with model-based checks. Third, the model went through reinforcement learning on agentic tasks, including domain-specific environments for kernel optimization, web development, and computer-aided design.^[1][8]

The documented interface matches Grok 4.5 in most respects: text and JPEG or PNG image input, text output, a 500,000-token context window, and a February 1, 2026 knowledge cutoff, with no access to real-time information unless a search tool is enabled.^[4][5] Two things changed. The API imposes no text output limit, and the reasoning-effort control gained a fourth setting, xhigh, above the default high; SpaceXAI's documentation states that xhigh is available on grok-4.6, that requests sending it to grok-4.5 are treated as high, and that reasoning cannot be disabled.^[3][4][7] The model works with the Responses and Chat Completions APIs and supports function calling plus SpaceXAI's server-side web search, X search, and code-execution tools.^[4]

Benchmarks

SpaceXAI's launch post says Grok 4.6 "achieves frontier intelligence across several agentic coding and knowledge work benchmarks" and that it matches GPT-5.6 Sol on the Artificial Analysis Intelligence Index.^[1] The announcement's own comparison table, reproduced below, is more mixed than that summary: Grok 4.6 leads on three of the ten rows, Claude Fable 5 on four, and GPT-5.6 Sol on two, with the Intelligence Index row going to Fable 5. The table compares Grok 4.6 at high reasoning effort against competitors at their maximum settings, and its footnote states that third-party scores are "the best of self-reported or publicly available results," so these are assembled figures rather than head-to-head runs.^[1][2]

Scroll sideways for more →

Source: SpaceXAI launch page as archived on August 12, 2026.^[1][2] The benchmark graphic attached to the launch tweet differs from the web table in one cell, giving Claude Fable 5 64.9% rather than 63.6% on FrontierCode v1.1 (Extended); both appeared on launch day and the discrepancy was unexplained as of that evening.^[9][2]

The gains over Grok 4.5 are largest on the knowledge-work rows: 227 points on GDPval-AA v2, 264 on AA-Briefcase, and a 10-point gain on Terminal-Bench v3.0, though that last benchmark remains the model's weakest result against competitors. The Harvey LAB row refers to Harvey's Legal Agent Benchmark, an open-source evaluation of agents on legal work whose live leaderboard is hosted by Vals AI; it is the table's biggest lead and also its least symmetrical comparison, with GPT-5.6 Sol listed at 2.5%. Vals AI's hosted leaderboard added Grok 4.6 on August 13 at 15.8 percent accuracy, matching the vendor's figure.^[1][19]

Independent measurement

Artificial Analysis evaluated Grok 4.6 at high reasoning effort on launch day and scored it 61 on Intelligence Index v4.1.1, a composite of nine evaluations including GDPval-AA v2, Terminal-Bench v2.1, Humanity's Last Exam, and GPQA Diamond. That placed it sixth of 183 tracked models, tied with GPT-5.6 Sol at maximum effort and Claude Opus 5 at high effort, and behind Claude Opus 5 at maximum and xhigh effort (63) and Claude Fable 5 (62). Grok 4.5 (high) sat at 56 on the same board.^[14][15]

The cost profile shifted against the predecessor even though the token prices did not. Artificial Analysis measured $0.84 per Intelligence Index task for Grok 4.6 against $0.36 for Grok 4.5, with the model generating 72 million output tokens across the index, around the median for comparable models; the evaluator's full run cost $1,068.^[14][15] Grok 4.5's launch story had leaned heavily on below-median token use, so on this evidence the intelligence gain was partly bought with verbosity. Artificial Analysis had not yet published an output-speed measurement for Grok 4.6 as of August 12.^[14]

Arena AI (formerly LMArena) listed grok-4.6 in both high and xhigh configurations for head-to-head voting on launch day but had not yet ranked it in any modality.^[16]

Pricing and availability

Grok 4.6 costs $2 per million input tokens, $0.50 per million cached input tokens, and $6 per million output tokens for requests under 200,000 input tokens; at or above that threshold the rates double to $4, $1, and $12.^[6] The headline input and output prices are identical to Grok 4.5's, supporting the launch tweet's "same price" claim, but the cached-input rate is higher: SpaceXAI's pricing page lists Grok 4.5 cached input at $0.30 ($0.60 long-context), so prompt-cache-heavy workloads pay more on the new model.^[6][9] Both launch posts also mention "a fast variant" at "twice the price," without further specification.^[1][8]

Server-side tools are billed separately, at $5 per 1,000 calls for web search, X search, and code execution, $2.50 per 1,000 for collections search, and $10 per 1,000 for file-attachment search.^[6] SpaceXAI recommends setting a prompt_cache_key so that a conversation's requests route to the same server, warning that cache hits are otherwise unreliable.^[4]

At launch the model was available through the xAI API under the identifier grok-4.6, as the default model of the Grok Build coding agent, in Cursor on all plans, and through the OpenRouter, Vercel, and Cloudflare gateways; OpenRouter's listing records a 500,000-token context and a dated internal snapshot name of grok-4.6-20260810.^[4][17] The launch materials do not mention availability in the consumer Grok apps or on X, the same staging Grok 4.5 followed.^[1][3]

Reception

Launch-day coverage was thin compared with Grok 4.5's, which had arrived alongside a $60 billion acquisition story. Reuters carried a brief wire item, "SpaceXAI Introduces Grok 4.6," and early trade coverage led with the Artificial Analysis tie with GPT-5.6 Sol.^[18][20] The most substantive independent datapoint available on day one was Artificial Analysis' scoring described above, which supported the vendor's frontier claim on aggregate intelligence while showing the model's per-task cost more than doubling against its predecessor.^[14][15]

The comparison targets Musk chose are worth noting: his July 24 chart measured Grok 4.5 against Claude Opus 5, and the launch table benchmarks Grok 4.6 against Claude Fable 5 and GPT-5.6 Sol at their maximum effort settings while Grok runs at high, one setting below its own xhigh.^[10][1][7] With three frontier releases planned in roughly six weeks (4.5 in July, 4.6 in August, 4.7 promised weeks later), the schedule itself became part of the product story, and it did not hold exactly: the release landed five days after Musk's August 7 estimate.^[11][12][3]

Roadmap: Grok 4.7 and Grok 5

Musk's July 28 post described Grok 4.7 as "the 2.1T model released a few weeks later," saying it "will be better than 4.6 in every way, except slightly slower to serve"; on the August 4 earnings call he placed it three to four weeks out.^[11][12] For Grok 5, the call guidance was a release around the end of 2026, trained with SpaceX's internal engineering and flight data, which Musk suggested could make it the strongest available model for engineering work.^[12][13] All of these are founder statements about unreleased products rather than documented specifications, and the Grok 4.6 schedule itself slipped five days between Musk's estimate and the release.^[11][3]

See also

• Grok
• Grok 4.5
• Grok Build
• xAI
• Cursor
• SpaceX-xAI merger
• Artificial Analysis
• SpaceX Starmind

References

1. ^SpaceXAI. "Introducing Grok 4.6." August 12, 2026. x.ai/...grok-4-6
2. ^SpaceXAI. "Introducing Grok 4.6." Archived copy of the launch-day page, captured August 12, 2026, 15:44 UTC. web.archive.org/...grok-4-6
3. ^SpaceXAI. "Release Notes." SpaceXAI Docs. Updated August 12, 2026. docs.x.ai/...release-notes
4. ^SpaceXAI. "Grok 4.6." SpaceXAI Docs. Updated August 12, 2026. docs.x.ai/...grok-4-6
5. ^SpaceXAI. "Models." SpaceXAI Docs. Updated August 12, 2026. docs.x.ai/...models
6. ^SpaceXAI. "Pricing." SpaceXAI Docs. Accessed August 12, 2026. docs.x.ai/...pricing
7. ^SpaceXAI. "Reasoning." SpaceXAI Docs. Accessed August 12, 2026. docs.x.ai/...reasoning
8. ^Cursor. "Introducing Grok 4.6." August 12, 2026. cursor.com/...grok-4-6
9. ^SpaceXAI (@SpaceXAI). "Introducing Grok 4.6. It delivers frontier intelligence and is a significant improvement over Grok 4.5 at the same price." X, August 12, 2026. x.com/...2087562800982077492
10. ^Elon Musk (@elonmusk). "Grok 4.5 and Opus 5 are alone on Pareto frontier" and reply "Grok 4.6 in 2 weeks and Grok 4.7 in 4 weeks." X, July 24, 2026. x.com/...2080724087593226311
11. ^Elon Musk (@elonmusk). "Grok 4.6 releases around August 7. This will be the 1.5T model with significantly improved SFT & RL. Grok 4.7 will be the 2.1T model released a few weeks later..." X, July 28, 2026. x.com/...2082123925283041545
12. ^Not a Tesla App. "Highlights From SpaceX's First-Ever Earnings Call: Starship, Starlink, Grok and More." August 2026. notateslaapp.com/...tarship-starlink-grok-and-more
13. ^Tesla Oracle. "Space, Connectivity, AI: Key takeaways from Elon Musk's SpaceX Q2 2026 Earnings Call." August 10, 2026. teslaoracle.com/...ks-spacex-q2-2026-earnings-call
14. ^Artificial Analysis. "Grok 4.6 (high): Intelligence, Performance and Price Analysis." Accessed August 12, 2026. artificialanalysis.ai/...grok-4-6
15. ^Artificial Analysis. "LLM Leaderboard." Accessed August 12, 2026. artificialanalysis.ai/...models
16. ^Arena AI. "Leaderboard." Accessed August 12, 2026. arena.ai/leaderboard
17. ^OpenRouter. "SpaceXAI: Grok 4.6." Model listing, accessed August 12, 2026. openrouter.ai/...grok-4.6
18. ^Reuters (via TradingView News). "SpaceXAI Introduces Grok 4.6." August 12, 2026. tradingview.com/...:0-spacexai-introduces-grok-4-6
19. ^Vals AI. "Harvey's Legal Agent Benchmark." Accessed August 12, 2026. vals.ai/...hlab
20. ^TradingKey. "SpaceXAI Officially Launches Grok 4.6: Performance on AI Intelligence Analysis Index Matches GPT-5.6 Sol." August 12, 2026. tradingkey.com/...-gpt-5-6-sol-ai-index-tradingkey