xAI Grok 4.7 Release: Capabilities, Benchmarks, and Community Reaction

Grok 4.7 delivers stronger long‑context coding and safety at the same price as 4.6

xAI announced Grok 4.7 on 2026‑09‑21, positioning it as the most capable model for coding and knowledge work in the Grok line. The model runs at the same $2 / M input and $6 / M output token rates as Grok 4.6, yet it claims longer reasoning windows, better self‑verification, and the strongest safeguard stack the company has released.

Core technical improvements

  • Larger base model: Grok 4.7 uses a bigger architecture than 4.6, with roughly 40 % more parameters according to community speculation.
  • Extended RL training: The reinforcement‑learning phase was lengthened and weighted toward multi‑hour tasks, improving performance on benchmarks that stress long‑running code generation (e.g., CursorBench 4.0).
  • Native Grok Bot integration: The model now understands the Grok Bot harness directly, which boosts conversational and general‑knowledge tasks.
  • Enhanced self‑checking: New verification loops let the model catch its own errors more reliably, a key factor in the reported safety gains.

Benchmark performance at a glance

Benchmark Grok 4.7 Grok 4.6 GPT‑5.6 Sol Fable 5.1
CursorBench 4.0 (software engineering) 46.3 % 40.4 % 41.7 % 51.8 %
DeepSWE v1.1 (software engineering, high‑effort) 71.0 %* 65.2 % 72.7 % 70.0 %
EEBench (electrical engineering) 64.0 % 53.0 % 39.4 % 56.4 %
AA Briefcase v1.1 (multi‑hour office work) 1,657 pts 1,546 pts 1,487 pts 1,678 pts
Terminal‑Bench 4.0 (multi‑hour terminal work) 38.0 % 20.3 % 37.3 % 57.9 %
Harvey Legal Agent 19.6 % 15.8 % 2.5 % 6.7 %
HealthBench Professional 56.7 % 48.5 % 60.5 % 62.1 %

*high‑effort score.

Takeaway: Grok 4.7 leads its own lineage on most engineering and professional‑work benchmarks, especially in long‑context tasks, while remaining price‑competitive with its predecessor.

Safety and dual‑use controls

  • New safeguard stack: The release notes claim the strongest refusal and jailbreak resistance to date.
  • HackerBench v0.3: Grok 4.7 lets only 3.3 % of risky dual‑use prompts through, while still permitting legitimate security research.
  • LatchBio biosafety benchmark: The model scores 62.4 %, the highest among evaluated frontier models.
  • Red‑team access: Select cybersecurity partners receive invite‑only access to Grok 4.7’s defensive capabilities.

Pricing and availability

  • Token pricing: $2 / M input tokens, $6 / M output tokens (identical to Grok 4.6). A “fast” variant costs double for twice the output speed.
  • Access points: Available through Cursor, Grok Build, the Grok API, third‑party coding harnesses, and cloud model routers.
  • Free trial: Users can try Grok Build at x.ai/build.

Community reaction on Hacker News

Positive impressions

  • Long‑running coding: Users reported Grok 4.6 already handling complex build‑root setups; many expect 4.7 to improve on that.
  • Frontend development: Several commenters praised Grok’s ability to generate HTML/CSS/JS snippets quickly.
  • Safety perception: The new safeguard stack is viewed as a step forward for dual‑use domains.

Criticisms and concerns

  • Token efficiency regression: Multiple commenters noted a sharp rise in output token consumption (e.g., 240 M tokens for a 46 % CursorBench score versus 97 M for 4.6), suggesting the model is less economical despite unchanged pricing.

    "Significant regression in token efficiency compared to Grok 4.6 suggested by artificialanalysis.ai Intelligence Index Comparisons." – @notduckrabbit

  • Benchmark relevance: Some users question the choice of comparators (e.g., GPT‑5.6 Sol, Fable 5.1) and the omission of Chinese models such as Kimi or DeepSeek.

    "Why Chinese models from Kimi, Deepseek are not added in comparison benchmarks?" – @shdtabasum

  • Cache pricing opacity: The public price sheet lists only input/output rates, while cache read costs remain hidden, leading to accusations of deceptive pricing.

    "Every Grok release obscures their cache pricing while highlighting their input/output pricing… Long‑running agentic workflows are dominated by cache reads." – @GodelNumbering

  • Performance vs. cost trade‑off: Users report that the fast variant doubles cost without proportionate speed gains, and that caching penalties make Grok more expensive than alternatives like DeepSeek 4.1 Flash.

    "Grok it's really expensive. I'm getting really amazing results using DeepSeek 4.1 Flash for fraction of the price." – @meerita

  • Mixed benchmark outcomes: While Grok 4.7 outperforms 4.6 on many tasks, it falls behind competitors such as Astra and Gemini 3.8 Flash on certain specialized benchmarks (e.g., image‑to‑HTML generation, omniscience tests).

    "Astra is something else. It is expensive, but uses way fewer tokens for my tasks." – @saejox "Grok 4.7 is near the top of the Redactle LLM leaderboard but still not as good as Gemini 3.8 Flash." – @pampas

  • User‑experience regressions: Some early adopters observed slower response times and higher token usage, leading to quicker quota exhaustion in subscription plans.

    "4.7 is definitely slower & more expensive… I haven't been making fast progress today with benchmarking it." – @mchusma

What the mixed feedback means for the frontier

  • Benchmark gains are modest: Grok 4.7’s headline improvements (e.g., +6 % on CursorBench) are real but not transformative; the model still trails the very top frontier models on several tasks.
  • Pricing strategy matters: Keeping input/output rates flat while increasing token consumption may erode cost‑effectiveness, especially for heavy‑use developers and agents.
  • Safety advances are a differentiator: The strongest jailbreak resistance to date could make Grok 4.7 attractive for regulated industries despite its higher token usage.
  • Community scrutiny is intensifying: Users are demanding transparency on cache pricing, benchmark methodology, and real‑world token efficiency, signaling that future releases will need clearer cost disclosures.

Bottom line

Grok 4.7 pushes xAI’s coding‑focused model forward with longer context windows, better self‑verification, and the most robust safety stack the company has released. However, the modest benchmark improvements, higher token consumption, and opaque cache pricing have generated a split response among developers. The model’s value proposition now hinges on whether its safety guarantees and long‑running task performance outweigh the increased operational costs for real‑world workloads.

Sources

Related