Grok 4.6 scores 61 on the Artificial Analysis Intelligence Index
TL;DR Highlight
Grok 4.6 scores 61 on the AA Index, matching GPT-5.6 Sol at 60%+ lower token prices
Who Should Read
A backend developer choosing an LLM for agents or coding tools while comparing performance against token cost
Core Mechanics
- Index score of 61 joins the top tier
- It gained 5 points over Grok 4.5 and 23 over Grok 4.3.
- It matches GPT-5.6 Sol (max) at 61, behind Claude Opus 5 (max) at 63 and Claude Fable 5 at 62.
- It is slightly ahead of Kimi K3.
- Strength is agentic work, not static reasoning
- GDPval-AA v2 Elo is 1753, behind only Claude Opus 5.
- 𝜏³-Banking is 50.7%, top two alongside Qwen3.8 Max at 51.3%.
- Terminal-Bench v2.1 is 88.4%, in line with the leaders.
- Pricing unchanged at $2/$6
- Per 1M input/output tokens, it is 60%+ below Claude Opus 5 ($5/$25) and GPT-5.6 Sol ($5/$30).
- Cost per task is $0.84, same as Kimi K3 with slightly higher intelligence.
- It sits on the cost vs. intelligence Pareto frontier.
- Fewer turns and tokens on long tasks
- On AA-Briefcase (a private long-horizon knowledge-work benchmark) its Elo is 1577, Fable 5-tier.
Community Reactions
- Switching from Claude to Grok 4.5 felt faster and more interactive, per one user
- They use Grok as a daily driver and cross-check with Codex and Sol.
- Visual verification with tools improved in 4.6, per one user
- Previously only Opus and Fable consistently screenshotted results; Grok is now right behind them.
- Grok 4.5 rarely did this, they said.
- Cache read price hike matters for real bills, per one commenter
- They said about 80% of their coding-session bill is cache reads and writes.
- Cursor's plans are generous with first-party models like Grok, per commenters
- One said the top plan felt close to unlimited compared to Claude.
How to Apply
- Run one agent task on 4.6 and compare cost
- Within 30 minutes, run one real task on your current model and on Grok 4.6.
- Record turns, input tokens and cost to see whether the 53 vs. 103 turn gap shows up.
- Include cache-hit pricing in cost estimates
- Pull the cache read share from your monthly token logs.
- Recompute total cost with the $0.5 cache rate.
- Use output token price as a selection criterion
- For reasoning-heavy workloads, re-estimate budgets using the $6 vs. $25-30 output price gap.
Try It
# Prompt to paste into Claude or ChatGPT (to design a model-swap experiment)
I run agents on LLM APIs. Use the info below to design a model-swap experiment.
- Current model: {{current model}}, input/output price: {{prices}}
- Candidate: Grok 4.6 ($2 input, $6 output per 1M tokens, cache hit $0.5)
- Three representative tasks: {{task list}}
- Monthly input tokens: {{input tokens}}, output tokens: {{output tokens}}, cache hit ratio: {{ratio}}
Instruction: Make a table template to log turns, input tokens, cost and success per task, plus a monthly cost formula.
Next step: Fill in real logs and compare cost per task between the two models.Terminology
Related Papers
The Judge Can Debias Itself: Awareness and Self-Debiasing for Reliable LLM-as-a-Judge
Warning an LLM judge about bias or letting it self-critique improved 20 of 21 model–bias combinations.
90% of US working hours are spent on tasks not covered well by AI benchmarks
Mapping 29 benchmarks to job tasks shows 90% of US working hours lack good coverage
Postgres 19 delayed after AI reviewers find more bugs than humans can fix
LLM code review surfaced so many bugs that Postgres 19 slipped to November and 8-10 features were reverted
Show HN: Gemini 4 Argon plays StarCraft
LLMs wrote StarCraft bots in one hour; GPT-6 Astra scored highest.
A Plausible Explanation Is Enough to Mislead Multimodal Judges
Adding plausible explanations to fixed answers cut MLLM judges' wrong-answer rejection by 12.7–13.8 points.
Neural networks are clean
Looking at neurons directly shows most have one main job, challenging superposition.