Grok 4.5

Original Article ↗ Hacker News Discussion ↗

Grok 4.5, xAI’s new large language model trained with vast amounts of real-world Cursor coding data, is being positioned as an Opus-tier software engineering model that’s significantly cheaper and faster, with strong benchmark results on harder coding suites like DeepSWE and TerminalBench. Early users report impressive performance for code generation, debugging, and test refactoring, though some still find Anthropic and OpenAI models more reliable in certain domains, and note pricing quirks around large-context and cache-hit usage. A major fault line is non-technical: many are wary of adopting Grok due to Elon Musk’s politics, past safety controversies (including CSAM claims and “MechaHitler” incidents), and the risk of covert political steering in model alignment, while others argue all major LLMs are politically biased and prioritize cost–performance over ethics.

Popularity & Adoption

  • Several commenters say Grok is far less used than GPT/Claude/Gemini, citing OpenRouter token throughput as an order of magnitude lower.
  • Reputation is hurt by association with its owner, prior “MechaHitler”/CSAM controversies, and perceived ideological agenda; some refuse to use it on principle.
  • Others don’t care about the politics and evaluate it purely on price/performance, or note that many people happily use Chinese models despite their own issues.

Coding Performance & Benchmarks

  • Broad agreement that earlier Grok versions were behind for SWE; 4.5 is the first seen as competitive.
  • On coding benchmarks (DeepSWE, TerminalBench, SWE-Bench Pro) it appears roughly “Opus 4.7 tier” / comparable to GLM‑5.2, though some still find GPT‑5.5/Opus 4.8 more reliable.
  • Many praise its speed and token efficiency (high tokens/sec, fewer reasoning tokens than peers). Some describe it acing complex refactors, debugging, test-suite overhauls, and Kubernetes fixes.
  • Others report it “borderline unusable” on simple refactors, suggesting harness issues, domain dependence, or just inconsistencies.
  • Tool-calling and agentic behavior are reported as much improved vs Grok 4, but still debated.

Pricing & Token Economics

  • API pricing: $2/$6 per 1M in/out tokens under 200k context, doubling above 200k; cache hits cost 25% of input price (higher than many US and Chinese competitors).
  • Some see it as “Opus-class at Haiku prices” once efficiency is factored in; others say it’s still expensive vs DeepSeek/GLM.
  • Subscription complaints: the $40 plan only yields ~8 hours/month of Grok Build; SuperGrok and Cursor bundles give more, but value vs Codex/Claude Code is disputed.

Training Data, Benchmarks & Privacy

  • Training used “trillions of tokens” of Cursor IDE interaction data plus large RL environments. Many see this as the real moat for coding quality.
  • Concern: Cursor trains on user work by default, even for paying customers, raising IP‑leakage worries.
  • Synthetic-data “model collapse” concerns are discussed; consensus in-thread is that curated, mixed real/synthetic data is fine.
  • CursorBench was partly contaminated by training data; they excluded it from the public benchmark chart, but some still see this as benchmark juicing risk.

Politics, Bias & Trust

  • A large subthread debates bias: multiple external analyses (cited in-thread) are interpreted as showing major labs lean left, with Grok slightly right but closest to center; others question these studies.
  • Many are uneasy that a model can be tuned so directly to one billionaire’s politics and fear silent backend “nudging” in non-political domains, making it a supply‑chain risk.
  • Counterpoint: all frontier labs shape outputs via RLHF/RLAIF; none are politically neutral, so Grok is just more explicit and, to some, more “balanced.”
  • Earlier image behavior around minors and the “MechaHitler” persona are repeatedly cited as red lines by critics; defenders say these issues have been patched and focus on censorship differences vs other labs.

Access, Region Locking & Release Timing

  • Grok 4.5 is initially unavailable in the EU (VPNs work); xAI says EU support is coming soon. Region‑locking is seen as unusual but not unprecedented.
  • Commenters note frequent clustering of big model announcements (Grok vs GPT releases) and speculate about competitive timing, corporate intelligence, and PR strategy.

Market Strategy & Competition

  • Some see Grok stuck in the middle: not best-in-class and not cheapest; hard to justify over Anthropic/OpenAI/DeepSeek/GLM.
  • Others argue xAI’s huge cluster, excess capacity, and integration with Cursor/X/Tesla give it room to undercut on cost and build a niche, even at low market share.
  • There’s broad agreement that Anthropic still leads on serious coding/research, with OpenAI and Google strong in other areas; Grok 4.5 is viewed as moving xAI from “non‑contender” to “second tier but relevant.”