Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index

Original Article ↗ Hacker News Discussion ↗

Grok 4.6, xAI’s latest large language model, is drawing attention for offering near-frontier coding and reasoning performance at lower token costs and higher usage limits than many Anthropic and OpenAI plans, especially when bundled through tools like Cursor. Commenters weigh its strengths—speed, concise communication style, and generous subsidized access—against reports that it still trails top-tier models like Anthropic Fable 5 and GPT‑5.6 on complex tasks, while also probing broader concerns about Musk’s politics, aggressive data center build-out, carbon impact, and whether a third or fourth closed frontier model can sustain a viable niche in a market increasingly pressured by powerful open-weight alternatives.

Model quality and benchmarks

  • Many see Grok 4.5 as “decent frontier” but below top models like Fable/Opus 5/Sol/GPT‑5.6; some speculate 4.6 is around Opus 4.8 in practice but still under the newest Anthropic/OpenAI tiers.
  • Several daily users report Grok 4.5/4.6 is fast, concise, and “good enough” for engineering; they value shorter answers and simpler, human‑like code.
  • Others say its speed is partly illusory: it often tackles only a small slice of a task and prematurely claims it’s done, with weaker outputs than Anthropic/OpenAI.
  • Compared with Kimi K3, Grok is generally described as better and cheaper; better than Gemini in some coding tasks, but not best‑in‑class overall.
  • 4.6 is noted as more tool‑use/verification‑oriented (e.g., screenshotting UIs) than 4.5, but no longer quite as “blazing fast.”

Pricing, token efficiency, and subscriptions

  • Grok is repeatedly described as very cost‑effective, especially via Cursor and xAI subscriptions. Some feel like a $30 sub replaced hundreds to thousands of dollars of per‑token spend elsewhere.
  • Benchmarks cited in the thread claim Grok’s token use is on par with recent OpenAI models and per‑token cost is lower; others counter that GPT remains the most reasoning‑efficient.
  • Grok 4.5 listed at $2 input / $6 output vs Opus 4.8 at $5 / $25. Cache read price reportedly rose from $0.30 to $0.50 in 4.6, which users say hits agentic coding costs hard.
  • Cursor: generous Grok/Composer usage plus ~$20 of non‑first‑party credits; some call it the cheapest way to use Grok for coding.
  • OpenAI’s subscription “reset” behavior frustrates people trying to budget usage; some respond by chasing whichever frontier model is currently most subsidized.

Usage patterns and workflows

  • Contrary to “no one uses Grok” claims, multiple commenters use it as a primary coding model or as part of multi‑model workflows (e.g., Fable for planning, Grok for rapid iteration/debugging).
  • It’s used in Cursor, Grok CLI/Build, Copilot (historically), and custom harnesses (OpenCode, Pi coding agent) that route tasks to different models.
  • Some teams report effectively no cap on overall AI spend and still choose Grok for style and speed, not just price.
  • For security work, users say Grok and some Chinese models will do vulnerability analysis and PoCs where Anthropic/OpenAI block or heavily sanitize.

Ethics, politics, and boycotts

  • A substantial contingent refuses to use Grok/xAI at all, citing:
    • The owner’s alleged Nazi salutes, election interference, far‑right support in multiple countries, behavior around foreign aid and Ukraine, and general governance style.
    • Concerns that Grok’s system prompts and political slant are shaped directly by those views.
  • Others argue almost all rich actors “interfere” with elections, that boycotts are inconsistent if one also uses Chinese or other controversial providers, or that they separate product quality from owner politics.
  • Some prefer Grok because they believe OpenAI/Anthropic skew progressive and claim Grok is closer to “center,” while others contest what “center” even means.

Business strategy, compute, and competition

  • Reasons proposed for xAI/SpaceX investing heavily in Grok despite strong incumbents:
    • Monetizing surplus GPU capacity by renting to other labs while ramping demand.
    • Vertical integration for internal use (e.g., Tesla, robots, multi‑planetary ambitions) and avoiding dependency on OpenAI/Anthropic.
    • A bet that they can reach or surpass the frontier given large compute, capital, and training data (including Cursor’s dataset).
    • Philosophical motives and desire for a politically different model family.
  • Skeptics question the business case: consumer and enterprise markets already dominated, open‑weight models squeezing the low end, and likely eventual consolidation. Some label it FOMO or investor‑driven bubble behavior.
  • Debate over infrastructure:
    • xAI/SpaceX praised for owning datacenters and planning their own chips, potentially enabling cheaper tokens.
    • Others doubt chip‑fab timelines (especially EUV/FEL approaches), note regulatory shortcuts (e.g., gas‑turbine datacenters), and question the reliability of self‑reported metrics.

Environmental and ancillary notes

  • A cited chart (from Stanford HAI via media report) portrays Grok training as especially CO₂‑intensive due to portable gas generators, claimed to be less efficient than other large setups; some see AI carbon debates as overblown in general but accurate for Grok.
  • Grok’s Imagine 2.0 image/video generation is described as low quality and “plastic‑looking.”
  • Some users praise Grok’s communication style versus ChatGPT’s verbosity and Claude’s rhetorical tics, and value having a distinct “model family” for creative tasks alongside Anthropic, OpenAI, and Gemini.