Grok 4.7
Grok 4.7, xAI’s latest large language model, is being judged as a modest step up in raw capability over Grok 4.6 but a regression in token efficiency and value, especially for long-running agentic coding workflows. Commenters report it can be strong for certain tasks like web frontends, legal research, and image‑to‑HTML work, yet still lags Anthropic’s and OpenAI’s top models on many benchmarks and often burns far more tokens to reach similar outcomes. Many are also put off by opaque cache pricing, nerfed subscription limits, and the political baggage around its parent company, while others welcome it as a fast, plain‑English alternative that broadens the field of serious proprietary models.
Overall Perceived Quality vs Competitors
- Opinions are split: some say Grok (4.6/4.7) is “solid” and often better than cheaper Chinese models for hard tasks; others find it clearly behind OpenAI/Anthropic on capability and reliability.
- Several users say 4.6 felt roughly like older Opus/Sonnet-era models; 4.7 is seen by some as only a small step up, or even a regression.
- Benchmarks shared in the thread generally place Grok below Astra/Fable and often below Sol/Opus on cost‑per‑task and token efficiency.
Coding & Agentic Behavior
- Many positive reports for:
- Frontend web work, image→HTML, PCBs/tscircuit, niche utilities, and code review (often catching different bugs than Fable/Astra).
- Agentic coding in Cursor and Grok Build, with strong parallel tool calls and concise step logging.
- Negative reports include:
- Frequent premature “Done!” responses, lazy fixes, brittle or insecure code (e.g., SQL injection, bad auth), infinite planning loops in some harnesses, and struggles with complex 3D or WebGL tasks.
- 4.7 in particular is said to be more token-hungry and sometimes worse behaved than 4.6 in certain agent setups.
Language Style and UX
- Big attraction for some: Grok’s “plain English,” brevity, and lack of “Claudish” verbosity.
- Others find it too terse, inventing internal terms then using them without explanation, forcing follow‑up questions.
- Compared to Claude, several users prefer Grok’s tone for casual chat and legal research; for deep technical explanations, Claude is still considered superior by many.
Benchmarks, Token Efficiency, and Model Size
- 4.7 reportedly has ~40% more weights than 4.6 but:
- Multiple references (ArtificialAnalysis, Vals, etc.) suggest worse token efficiency and sometimes slightly lower composite scores.
- Output tokens per benchmark task appear to have increased dramatically (e.g., ~97M → ~240M in one cited index for a small score gain).
- Some suspect heavy RL and benchmark‑chasing have hurt real‑world behavior and prose.
Pricing, Caching, and Subscriptions
- Base API price ($2/M input, $6/M output) is mid‑range, but:
- Cache reads are relatively expensive ($0.50–$1.00/M) vs competitors; several call this deceptive since long agent runs are cache-dominated.
- Subscription experience is mixed:
- Some report excellent value, especially via Cursor Ultra (effectively “infinite” tokens).
- Others say SuperGrok quotas were heavily nerfed; current plans are described as “terrible value” or burning through weekly usage in a day under coding workloads.
Use Cases and Adoption
- Reported uses:
- Coding (especially frontend, infra, Go), code review alongside other models, legal research, everyday Q&A, and voice assistant in cars and mobile apps.
- Some users say they rarely see Grok recommended in their circles and question its relevance; others claim growing enterprise traction because of price/performance and availability in popular harnesses.
Political and Ethical Concerns
- A significant subset refuses to use Grok/xAI for political/ethical reasons tied to its owner and prior controversial outputs (e.g., offensive content, CSAM concerns).
- Others argue these boycotts are superficial or inconsistent, while some counter that all technology choices are inherently political. The tension is unresolved.