Grok 4.5

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.”