What is the smartest AI model? November 2026 rankings

As of November 4, 2026, Gemini 4 Carbon at high effort leads the Artificial Analysis Intelligence Index v5.0 with 59 points. Claude Opus 5.5 at max effort with default fallback follows at 58, and Carbon's initial rollout is restricted. GPT-6.1 Sol max scores 55 at $1.27 per index task.

That $1.27 is a weighted average across the benchmark suite. A single prompt's cost depends on its token use and the model's pricing.

A leaderboard is a useful place to start. The tradeoff here is Sol's much lower benchmark cost against its four-point gap to Carbon. Whether that gap is worth paying for depends on your work and effort setting. Access is the other constraint.

How do the leading models compare?

The table shows selected configurations from the November 4 snapshot. Sol is included as a lower-cost comparator, and the selection leaves out configurations between it and the leaders.

Sources: Artificial Analysis leaderboard, Carbon comparison, Sol analysis and Google's Carbon announcement. These benchmark costs use standard prices and exclude Requesty charges.

Effort settings move the scores. Opus 5.5 at xhigh effort with default fallback also scores 57, while Astra at xhigh scores 55. Artificial Analysis rounds the scores, so close results have more precision behind them than the table shows.

What does the Intelligence Index measure?

The Intelligence Index v5.0 combines nine evaluations covering areas including coding, knowledge work, factuality and scientific reasoning. The v5 update introduced two new benchmarks and three upgrades. Private datasets cover most of the index.

Two details are worth keeping in mind:

• The index score is an aggregate across evaluations. A model can lead the aggregate and still trail on a particular task.
• The cost per index task is a weighted average using each evaluation's token costs and index weight. An individual prompt's cost depends on its token use and pricing.

Take coding agents: in a separate Artificial Analysis comparison, Carbon high scores 59 with Antigravity CLI, while Opus max scores 62 with Claude Code. Those results use a different index and different agent setups, so they cannot isolate the model as the cause of the difference.

If coding is your main use case, compare the coding benchmarks and costs before choosing an overall leader.

Which model is worth paying for?

API providers charge for input tokens (text sent to the model) and output tokens (text the model generates). The benchmark cost also depends on how many tokens a model uses across the index tasks.

Token rates come from Google, Anthropic, OpenAI's Astra announcement and OpenAI's pricing page. Index costs come from the Carbon comparison and Sol analysis.

Google announced introductory Carbon rates of $2 per million input tokens and $10 per million output tokens, rising to the standard rates shown above after the introductory period. The $7.76 benchmark figure uses standard pricing. For Astra, OpenAI's model documentation says prompts with more than 272,000 input tokens incur higher rates for the full request.

Token rates tell only part of the cost story. Artificial Analysis reports an average of 123,000 output tokens per index task for Carbon high and 173,000 for Opus max.

The effort setting can move the bill too. In the v5 announcement, Opus 5.5 scores 35 at low effort and 58 at max, with costs of $0.98 and $10.04 per index task respectively. For a broader provider comparison, see our LLM API pricing guide.

What would 100 benchmark-like tasks cost?

Our assumption: a hypothetical batch of 100 tasks has the same weighted task mix, token use and pricing basis as the Artificial Analysis index. This calculation illustrates the benchmark economics; it does not forecast the cost of ordinary prompts.

At those averages, Sol saves $649, or about 84%, against Carbon. Against Opus, it saves $877, or about 87%. Sol scores four points below Carbon and three below Opus on the index, and those aggregate differences do not tell us whether it would pass the same tasks in this batch.

Sensitivity check: suppose all 100 tasks run on Sol first, and 20 of them then need a full additional attempt on Carbon. At the same hypothetical averages, the batch costs $127 + (20 x $7.76) = $282.20. That is $493.80, or about 64%, less than running all 100 tasks on Carbon alone.

The 20% fallback rate is an assumption, and the calculation does not tell us whether either approach succeeds on those tasks. The weakest assumption is that your workload resembles the index mix.

The estimate excludes Requesty charges, subscription fees and taxes, and it does not account for changes in access or prices. See our guide to cost per task for a way to compare models on your own workload.

Is the smartest model available to you?

Google said Carbon would initially be available to a subset of cyber defenders and phased in to trusted partners through its Fairwind Program, before broader access. Its announcement describes that initial rollout; general access on November 4 is unconfirmed.

Consumer subscriptions need a separate check. As of November 4, 2026, general Carbon access is unconfirmed, and Google's initial rollout does not establish access through a consumer plan. Check which model your app subscription provides, then assess API access separately if you intend to build with it.

Before choosing a model, check the conditions for your account and application:

How should you choose a model for your own work?

1. 1
   Build a representative test set

   Collect tasks from the work you want the model to do, including cases where an incorrect answer would be costly. Define what a successful answer looks like before comparing models.

2. 2
   Compare quality and cost

   Run the same test set through the models and effort settings available to you. Record successful answers, token use and latency, then divide total spend by the number of successful tasks.

3. 3
   Configure routing and fallback

   Use the results to choose which tasks need a higher-scoring model and which can run on a cheaper one. Configure a fallback for provider failures, then monitor whether production results match the test.

Our guide to what an LLM router does covers routing policies for different workloads and provider failures.

So what?

Teams with high-volume workloads and a tight budget should test Sol first. Paying for Carbon or Opus can be justified when a representative test shows that their higher success rate is worth the extra cost, and when the model is available to you.

Run that test set and compare cost per successful task. Requesty's OpenAI-compatible gateway, performance monitoring and fallback policies can help teams test available models and manage provider failures.

Scores, costs and access can change. Check the current leaderboard, Sol page and Grok page, as well as Carbon's access status, before making a decision. On Requesty, Claude Opus 5.5, GPT-6 Astra, GPT-6.1 Sol and Grok 4.7 are available through one API key today; Gemini 4 Carbon is not in our model catalog yet.