Grok 4.7

TL;DR Highlight

Grok 4.7 raises coding benchmark scores at the same $2/$6 price as Grok 4.6

Who Should Read

A backend developer comparing cost and performance to pick a model for coding agents

Core Mechanics

Bigger base model, longer RL
- It uses a new, larger base model than Grok 4.6.
- Reinforcement learning ran longer, weighted toward tasks that take many hours.
- xAI says it verifies its own work and handles long context better.
Same price as 4.6, below rivals
- Input is $2 and output is $6 per million tokens.
- GPT-5.6 Sol Max is $4/$20 and Fable 5.1 Max is $10/$50.
- A fast variant gives 2x output speed at 2x price.
CursorBench 4.0 beats 4.6
- Grok 4.7 xHigh scores 46.3% versus 40.4% for Grok 4.6 High.
- GPT-5.6 Sol Max scores 41.7% and Fable 5.1 Max scores 51.8%.
- xAI claims frontier price-performance.
Results vary by benchmark
- Terminal-Bench 4.0 is 37.6% versus 20.3% for 4.6, but Fable 5.1 Max gets 57.9%.
- Harvey Legal Agent Benchmark is 19.6%, the highest among the compared models.

Community Reactions

Don't judge by input price alone
- One commenter says cache reads cost $0.50/M for Grok versus $0.25/M for Fable.
- Long agentic runs are dominated by cache reads, so the headline price misleads.
Slower and burns more tokens
- One commenter found it slower and pricier than 4.6, as if tokens were burned to climb benchmarks.
- Another cites artificialanalysis.ai as showing a token-efficiency regression versus 4.6.
Simon Willison's effort-level test
- Low and medium used similar token counts, and xhigh sometimes used fewer than high.
- Calling the xAI API directly still showed little difference between levels.
Comparison fairness questioned
- A commenter questions comparing 4.7 xhigh with 4.6 high.
- Another notes the charts show score and cost but not time or code quality.
Mixed results by task

How to Apply

Run three of your agent tasks on 4.7
- Run three typical coding tasks on your current model and on 4.7 with identical prompts.
- Record success, time, total tokens, and cost.
Include cache-read price in cost math
- Don't compare on input and output rates alone.
- Recompute monthly cost using the cache-read share in your agent logs.
Measure tokens per effort level
- Run the same task at low, high, and xhigh and compare reasoning tokens.
- If tokens rise without better success, pin the lower level.
Route models by task type
- Split frontend, backend, and terminal tasks and compare success per model.
- Set the winner as the default for each task type.

Try It

copy & run

# Paste into Claude Code or ChatGPT: model comparison experiment helper

I want to decide whether to switch my coding agent model.
Candidates: current model {{current model}}, new model Grok 4.7
Tasks I run often: {{3 tasks}}

Create the following.
1. A table to record when running both models on the same prompts (success, time, input/output/cache-read tokens, cost, code quality notes)
2. A cost formula that includes cache-read pricing
3. An experiment order for varying reasoning effort across low/high/xhigh

Next step: after filling in the table, paste the results and ask which model to use for which task.

Terminology

RL (reinforcement learning)Training that scores outcomes and reinforces good behavior instead of memorizing answers. Like a game AI improving through repeated play.

reasoning effortA setting for how long the model thinks before answering. Higher can improve accuracy but costs more tokens and time.

harnessThe execution framework around a model: tool calls, file access, loops. Like the chassis around an engine.

cache readA lower per-token fee for reusing an already-sent long prompt. Agents reread the same context, so it dominates cost.

CursorBenchA coding benchmark made by Cursor. Version 4.0 stresses long-running coding tasks.