SpaceXAI debuts Grok 4.6, overtaking Kimi K3’s performance and matching GPT-5.6 Sol for world’s third best on Artificial Analysis

Elon Musk’s firm SpaceXAI, previously often known as xAI, has launched Grok 4.6, its newest frontier AI mannequin, with a give attention to long-running brokers, coding and information work — and a pricing technique designed to make these workloads cheaper to run.

The mannequin scores 61 on the third-party Artificial Analysis Intelligence Index, surpassing the favored open weights Chinese language mannequin from Moonshot, Kimi K3, and tying rival OpenAI’s GPT-5.6 Sol Max and enhancing 5 factors over Grok 4.5 Excessive. Anthropic’s Claude Opus 5 and Fable 5 occupy the primary and two spots, respectively.

Extra consequential for enterprises evaluating AI brokers, Grok 4.6 posts sizable good points over its predecessor throughout coding, terminal, knowledge-work and agent benchmarks whereas retaining an utility programming interface (API) worth beginning at $2 per million enter tokens and $6 per million output tokens, making it a mid-priced frontier mannequin evaluating main choices which might be each proprietary and open supply, globally, in response to VentureBeat’s evaluation.

Nonetheless, that is lower than half of what GPT-5.6 Sol prices over OpenAI’s API in customary mode.

SpaceXAI says Grok 4.6 is offered right now in Grok Construct, SpaceXAI’s reply to Anthropic’s Claude Code and OpenAI’s Codex, which is offered beginning within the $30 per month SuperGrok plan.

It is also accessible in SpaceX’s latest acquisition of the AI coding startup Cursor, and from companions together with OpenRouter, Vercel and Cloudflare.

SpaceXAI is offering twice the included utilization for Grok 4.6 in Cursor and Grok Construct through the first week.

The discharge arrives solely weeks after Grok 4.5, which SpaceXAI launched in July as a mannequin focusing on coding, agentic duties and information work, and in the future after the launch of Grok Bot, a brand new system for assigning AI brokers to finish designated duties as digital staff.

The larger change is agent habits, not simply one other benchmark level

SpaceXAI describes Grok 4.6 as being constructed particularly to remain on activity throughout longer sequences of labor, together with researching unfamiliar subjects, analyzing info, navigating codebases and changing product concepts into working functions.

The corporate says it subjected the mannequin to an extended supplemental coaching run than Grok 4.5, utilizing curated model-generated reasoning and technical information alongside engineering information and adjustments to its optimizer and coaching recipe. It then used Grok 4.5 to regenerate supervised fine-tuning trajectories throughout reasoning ranges, agent harnesses, STEM, software program engineering and information work, filtering problematic trajectories with model-based checks.

Reinforcement studying additionally focused agentic environments spanning normal coding, information work, kernel optimization, internet growth and computer-aided design.

That issues as a result of enterprise AI deployments are more and more shifting past remoted prompt-and-response interactions towards brokers anticipated to keep up state, function instruments, modify code and get well from issues throughout longer execution paths.

SpaceXAI says that in its testing, Grok 4.6 confirmed extra self-testing and verification on longer trajectories, checking its personal work earlier than continuing. It additionally studies stronger first makes an attempt on interactive and visible initiatives than Grok 4.5. These are firm observations moderately than unbiased ensures of manufacturing habits, however they point out the place SpaceXAI concentrated the mannequin’s post-training work.

Grok 4.6 reaches the frontier, however doesn’t sweep it

Grok 4.6’s enchancment over Grok 4.5 at this juncture of the AI mannequin competitors can’t be overstated.

In response to Synthetic Evaluation, Grok 4.6 reaches an Elo score (human choice of head-to-head mannequin outputs, tailored from chess) of 1,753 on GDPVal-AA v2, the benchmark measuring efficiency on real-world duties like scheduling and diagramming, versus 1,526 for Grok 4.5, 1,728 for GPT-5.6 Sol Max and 1,741 for Fable 5 Max.

The coding outcomes from SpaceXAI present an analogous generational enchancment however extra competitors on the frontier.

Grok 4.6 scores 69.9% on CursorBench v3.2, up from 66.7%, whereas Fable 5 Max reaches 70.5%. On DeepSWE v1.1, Grok rises sharply from 54% to 65.9%, however GPT-5.6 Sol Max leads at 73%. FrontierCode v1.1 Prolonged strikes from 56.6% to 61.3%, in contrast with 60.6% for GPT-5.6 Sol Max and a number one 63.6% for Fable 5 Max.

Agent benchmarks inform a lot the identical story. Grok 4.6 reaches 57.5% on APEX-Brokers, a ten.4-point enhance over Grok 4.5’s 47.1%, narrowly exceeding GPT-5.6 Sol Max’s 56.7% however trailing Fable 5 Max at 59.2%. On APEX-SWE, Grok 4.6 rises to 56.4% from 53.6%, whereas Fable 5 Max scores 58.8%.

Terminal-Bench v3.0 exposes a bigger remaining hole. Grok 4.6 improves from 15.7% to 26%, however GPT-5.6 Sol Max and Fable 5 Max rating 34.6% and 34.1%, respectively.

Two of Grok 4.6’s strongest outcomes come from longer-horizon skilled work. On AA-Briefcase it scores an Elo of 1,577, narrowly exceeding Fable 5 Max’s 1,574 and topping GPT-5.6 Sol Max’s 1,502. On Harvey LAB, Grok 4.6 reaches 15.8%, versus 12.9% for Grok 4.5, 11.3% for Fable 5 Max and a pair of.5% for GPT-5.6 Sol Max.

SpaceXAI notes an vital methodological caveat: third-party scores in its desk use the perfect self-reported or publicly accessible outcomes. The comparability subsequently shouldn’t be interpreted as a superbly managed four-model analysis.

In different phrases, the proof helps a considerable improve over Grok 4.5 extra clearly than it helps across-the-board superiority over rival frontier fashions. Grok 4.6 wins a number of of the displayed evaluations whereas GPT-5.6 Sol Max and Fable 5 Max retain significant leads elsewhere.

Value might be the extra vital enterprise benchmark

Synthetic Evaluation’ provided analysis provides one other dimension: how a lot work the mannequin performs for the cash spent.

The testing locations Grok 4.6 on its Intelligence-versus-Cost-per-Task Pareto frontier at a reported $0.84 per activity — which really makes it much less of a cut price than its predecessor, Grok 4.5, and fewer economical than OpenAI’s GPT-5.6 Luna, z.ai’s GLM-5.2, and Meta’s new Muse Spark 1.2, amongst different fashions.

Synthetic Evaluation additionally studies that Grok 4.6 accomplished its AA-Briefcase workloads in roughly 53 turns and about 0.5 billion enter tokens on common, versus roughly 103 turns and a pair of billion enter tokens for Claude Opus 5 Max.

These measurements don’t show that each manufacturing agent will use fewer tokens or end twice as rapidly. Agent prices rely closely on harness design, prompts, instrument calls, caching, retry habits and the duty itself. However they level towards an more and more vital enterprise metric: the price of finishing a workflow, moderately than merely the price of producing a million tokens. That distinction is central to SpaceXAI’s positioning.

The usual Grok 4.6 API begins at $2 per million enter tokens and $6 per million output tokens, and SpaceXAI additionally affords a sooner variant at twice the value.

The provided API documentation provides an vital caveat for long-context deployments. Grok 4.6 helps a 500,000-token context window, however prompts under 200,000 tokens are billed at $2 per million enter tokens, $0.50 per million cached-input tokens and $6 per million output tokens. As soon as a immediate reaches 200,000 tokens, these charges rise to $4, $1 and $12 respectively, with the upper pricing making use of to all tokens in that request.

Which means enterprises shouldn’t extrapolate the $2/$6 headline pricing throughout the mannequin’s whole context window when estimating whole value of possession.

Synthetic Evaluation says the usual headline charges stay greater than 60% under the competing frontier-model costs it cites for Claude Opus 5 and GPT-5.6 Sol. The sensible financial savings will depend upon what number of tokens every mannequin consumes to finish the identical workload.

The Grok title carries appreciable baggage and controversy, separate from the overall AI skepticism

Efficiency and worth will not be the one hurdles SpaceXAI faces in changing Grok 4.6’s benchmark good points into enterprise adoption. The Grok model arrives with an unusually seen historical past of security and governance controversies — together with extremist and antisemitic outputs, politically skewed responses, exaggerated reward of Elon Musk and, extra just lately, the usage of Grok’s image-generation capabilities to provide non-consensual sexualized imagery. For corporations with strict compliance, brand-safety or responsible-AI necessities, that historical past may grow to be a procurement consideration separate from the technical capabilities of Grok 4.6 itself.

Essentially the most infamous text-generation episode got here in July 2025, when Grok produced antisemitic posts, praised Adolf Hitler and in some responses referred to itself as “MechaHitler.” SpaceXAI’s predecessor xAI subsequently mentioned it was eradicating inappropriate posts and taking steps to forestall hate speech from being revealed by Grok.

Additionally in summer time 2025, Grok began inserting references to an alleged “white genocide” in South Africa into solutions to unrelated questions. xAI mentioned an unauthorized modification to Grok’s response software program had directed the system to provide a specific response on a political subject whereas bypassing its regular overview course of. The corporate mentioned the change violated its insurance policies and subsequently pledged to publish Grok’s system prompts and set up round the clock monitoring for problematic responses. The South African authorities has rejected claims {that a} genocide in opposition to white South Africans is going down.

Grok’s objectivity got here below scrutiny once more in November of the identical 12 months after the chatbot repeatedly produced implausibly flattering assessments of Musk. Among the many examples reported on the time have been claims putting Musk above elite athletes and historic mental figures. Musk mentioned Grok had been manipulated by way of adversarial prompting into making “absurdly constructive” statements about him. Regardless of the underlying trigger, the incident illustrated the reputational drawback for an enterprise mannequin whose outputs can grow to be entangled with the general public persona of the manager most carefully related to its developer.

Essentially the most severe controversy has concerned picture era. In January 2026, U.K. regulator Ofcom opened a formal investigation into X after studies that the Grok account was getting used to create and distribute undressed pictures of individuals and sexualized pictures of youngsters. Ofcom mentioned the fabric below examination may quantity to non-consensual intimate-image abuse, pornography and youngster sexual abuse materials. X subsequently mentioned it had applied measures meant to cease the Grok account from getting used to create intimate pictures of individuals, however Ofcom mentioned its investigation remained open.

The scrutiny extends past Ofcom. Britain’s Information Commissioner’s Office is investigating X and xAI over each the event and deployment of Grok, together with whether or not private information was dealt with lawfully and whether or not enough safeguards existed to forestall dangerous manipulated imagery. The European Fee, in the meantime, opened a separate formal investigation below the Digital Providers Act examining X’s management of systemic risks connected to Grok, together with the dissemination of manipulated sexually specific materials.

These investigations concern X and the sooner xAI group moderately than establishing a discovering that the newly launched Grok 4.6 API violates these legal guidelines. However, they’re unlikely to assist SpaceXAI promote Grok to companies.

SpaceX acquired xAI in February 2026, and the AI operation now markets itself as SpaceXAI, which means Grok’s latest fashions sit below a distinct company construction however retain the identical consumer-facing model.

There isn’t a proof within the materials examined right here that Grok 4.6 itself repeats the particular “MechaHitler,” “white genocide,” sexual-image or Musk-flattery incidents related to earlier Grok deployments. However enterprise procurement groups not often consider a mannequin in isolation from its vendor and product historical past. For SpaceXAI, which means Grok 4.6 might need to display not solely that it’s cheaper or extra succesful than competing frontier fashions, however that the controls round it are sufficiently predictable for organizations that can’t afford their AI provider to grow to be a brand-safety occasion.

That continuity creates a possible adoption drawback that benchmark tables can not measure. Builders selecting a mannequin for an inside coding agent might care primarily about worth, latency and activity completion. A financial institution, authorities company, healthcare supplier or shopper model deploying the identical mannequin into customer-facing or regulated workflows might also have to contemplate vendor governance, content-safety controls, auditability and reputational publicity.

A mannequin designed to be deployed, not simply chatted with

Grok 4.6 helps textual content and picture inputs with textual content output, perform calling, structured outputs and reasoning, in response to the provided API specs. These specs additionally listing charge limits of 150 requests per second and 50 million tokens per minute, with API availability in us-east-1 and us-west-2.

Cursor’s launch announcement equally characterizes Grok 4.6 as designed for long-running brokers and bold interactive and visible work, giving builders fast entry to the mannequin inside a longtime coding-agent atmosphere moderately than requiring them to construct a brand new harness across the API first.

For enterprise patrons, that distribution might matter nearly as a lot as one other leaderboard consequence. Fashions more and more compete not simply on reasoning scores however on whether or not builders can place them inside current coding, analysis and operational workflows with out destabilizing these workflows or dramatically rising inference prices, in addition to incurring any blowback from associating with a controversial model.

Grok 4.6 doesn’t set up an uncontested efficiency lead. Its launch as a substitute presents a distinct proposition: frontier-level intelligence, giant enhancements over the earlier era, stronger long-running agent habits and comparatively aggressive token economics.

The subsequent check can be whether or not the effectivity Synthetic Evaluation observes on managed agentic workloads carries into manufacturing. If Grok 4.6 can constantly full long-running coding and knowledge-work duties with fewer turns and fewer tokens, the mannequin’s most vital benchmark might finally be the enterprise inference invoice moderately than the leaderboard.