Which AI models are best at getting work done?

An agent needs to do more than answer a question. It needs to read documents, use tools, and finish the task. I built the Agent Index to compare models across those skills. Here’s the methodology, the results, and how I’d use them to choose a model.

Methodology #

This report uses the Agent Index v1.2 preview, a November 2, 2026 snapshot of published evaluations across seven benchmarks:

• AA-Briefcase: producing professional deliverables.
• LiveBench Agentic: completing multi-step tasks.
• LiveBench Browser Use: interacting with web apps visually.
• Terminal-Bench 4.0: working in a terminal.
• HLE-Diamond: expert reasoning and knowledge.
• GDP.pdf: understanding documents.
• LiveBench Coding: coding.

These test individual skills an agent needs to get work done, rather than a complete workflow from start to finish.

• The overall ranking requires all seven results. 26 of 86 models qualify; missing results aren’t estimated or scored as zero.
• Cost is separate from capability. Pricing, speed, and latency don’t contribute to the index score. We compare them alongside the Agent Index to find models that offer the strongest capability for the cost.

The benchmark charts below show the top five selected reports available for each component, including models that don’t have enough coverage to receive an overall index score. I use the same report selection as the index, rather than choosing a model’s highest score from any setup.

Results: The top 5 LLM models for knowledge work #

To no one’s surprise, Claude Opus 5.5 leads the composite ranking at 90.4 out of 100, followed by GPT-6 Astra and Claude Sonnet 5.5.

Opus would be my starting point for broad knowledge work. But the strongest model isn’t always the one that makes sense for your budget. How much capability can you keep while bringing the cost down?

Small ranking gaps also deserve some caution. Sonnet and Fable are only 0.1 points apart, and this preview doesn’t estimate uncertainty.

Results: The best models for cost and performance #

But what if you want your agent to get more done for less? Here’s the Agent Index against blended token pricing, with the higher-capability, lower-cost quadrant highlighted.

Agent Index vs Cost Higher capability, lower token cost

View all 22 model prices and scores #

The quadrant uses the median score and price of the 22 models with eligible pricing. It’s a way to narrow the field, not a claim that every model inside it offers the best value. Three options I’d look at:

• GPT-6.1 Terra for a middle ground. It scores67.4 at $3 per million tokens , putting it in the highlighted quadrant. That’s roughly the same index score as Kimi K3.1 Pro at half the blended token price.
• Claude Haiku 5.5 for a tighter budget. At45.4 and $0.30 per million tokens , it sits above the median capability score while costing a tenth as much as Terra. It’s worth testing when that level of capability is enough for the task.
• Claude Sonnet 5.5 for stronger capability at a lower premium. It sits outside the low-cost quadrant, but scores85.8 at $6 per million tokens : half Opus’s blended price for 4.6 fewer index points. I’d shortlist it when I want capability closer to the top of the ranking without paying the top model’s price.

Prices use a 50/50 blend of input and output rates from the lowest eligible sourced standard offering. They aren’t total workflow costs. I’d still compare cost per accepted task, including retries, tool use, and human cleanup, before choosing a model.

Cost narrows the shortlist, but the next question is whether a model is good at the work you need it to do. An agent producing presentations needs different strengths from one working in a terminal. Here’s how the models compare on each benchmark.

Individual benchmark results #

AA-Briefcase: producing professional deliverables

AA-Briefcase evaluates agents producing spreadsheets, presentations, memos, and other professional deliverables. The selected track uses its v1.1 combined Elo score, which brings together rubric performance, analytical quality, and presentation.

Sonnet 5.5 leads this component at 1,814 Elo. If your product needs to produce something a person can use or share, this is a result I’d look at closely. A higher Elo score describes relative performance within this evaluation; it isn’t a percentage of tasks completed.

LiveBench Agentic: completing multi-step tasks

This component looks at multi-step work. It matters when an agent needs to do more than return a single response and has to work through a task.

Claude Fable 5.1 leads the selected reports at 77.9%. I’d use this component when shortlisting models for workflows with several steps, then check whether that advantage holds with my own tools and instructions. Where the chart labels an older release, the model was evaluated on a different question set.

LiveBench Browser Use: interacting with web apps

Browser Use tests screenshot-based interaction with web apps. That makes it relevant to agents that need to interpret a visual interface and act on it.

GPT-6.1 Sol leads at 93.3%. This is specifically a visual browser result. It doesn’t measure every agent that browses through text or the DOM. A published zero for a model that can’t accept screenshots counts as a zero for this evaluation, rather than missing data.

Terminal-Bench 4.0: working in a terminal

The index uses the Terminal-Bench 4.0 track with 66 tasks, a shared mini-swe-agent harness, and three repeats. Keeping the harness shared makes the setup more consistent than comparing each provider’s own agent stack.

Kimi K3.1 Pro leads this selected track at 67.2%. That’s a useful example of why I wouldn’t just pick the overall winner for every job. If terminal work is central to your agent, this component deserves more attention. Native-agent terminal scores elsewhere on the leaderboard are separate results.

HLE-Diamond: expert reasoning and knowledge

HLE-Diamond uses a revised 1,000-question exam. The index selects the track without tools, so this component measures reasoning and knowledge rather than an agent’s ability to complete a workflow with external tools.

I include this because reasoning can be important to the work an agent does, but it only contributes 10% of the composite score. A strong result here needs to be read alongside the model’s performance on documents, deliverables, and tool-based tasks.

GDP.pdf: understanding documents

The document component uses GDP.pdf’s All-pass result across five attempts. It’s a different measure from the GDPval evaluations you may see elsewhere on the leaderboard.

If your workflow starts with a pile of PDFs, this is one of the components I’d inspect before choosing a model. But document understanding is still only part of the job: your own evaluation should check whether the agent uses that information correctly in its final output.

LiveBench Coding: coding capability

Coding contributes 5% of the index. I include it because an agent may need code to transform data or complete a task, while keeping the focus on broader knowledge work.

Claude Fable 5.1 leads the selected reports at 97.7%. Several models score highly here. A strong coding result on its own doesn’t establish that a model will also handle documents, browser interaction, or professional deliverables well.

The composite Agent Index #

The individual results answer different questions. The composite gives us one starting point for comparing models across them.

I gave the most weight to producing professional work, completing multi-step tasks, and using a browser. Coding matters, but this isn’t meant to be another coding leaderboard.

Source and methodology ↗

These are editorial weights. If your agent mostly writes code or processes documents, your priorities will be different. The individual results let you make that judgment rather than accepting the overall ranking as the answer.

How the score is calculated

You can’t directly average an Elo score with a percentage. First, each benchmark result gets mapped onto a common 0–100 scale using fixed low and high reference scores. Then I apply the weights.

Normalized score = 100 × (reported score − low reference) / (high reference − low reference)
Agent Index = sum of each normalized score × its weight

Normalized scores are capped at 0 and 100. The reference points stay fixed for this version, so adding a model doesn’t rescale everyone else’s scores. Changing the benchmarks, weights, or reference points requires a new index version.

A score of 90 doesn’t mean the model completes 90% of knowledge work tasks. It describes performance relative to the index’s reference points and selected benchmarks.

What the composite doesn’t tell us

Seven benchmarks don’t cover everything an agent needs to do. Some measure overlapping skills, and three components come from LiveBench. Tool budgets and reasoning effort can differ. Some suites have few tasks, and the preview doesn’t provide confidence intervals.

Personal assistance and broader computer use don’t yet have dedicated components. Browser Use measures screenshot-based interaction, and HLE-Diamond measures reasoning without tools. Those boundaries matter when you interpret the overall score.

I’d use this index to choose a few candidates, then run the same real tasks through each one with the same tools and instructions. Define what an accepted result looks like before testing. Check completion, human cleanup, time, and cost. That’s where you’ll learn whether a benchmark advantage shows up in your product.

Final comparison #

Here are all 26 models with complete coverage in this snapshot. Prices are blended token prices, not estimates of total workflow cost. “Not available” means there wasn’t an eligible sourced price for the chart.

Source and methodology ↗

Opus 5.5 leads overall in this snapshot. The best model for your agent will depend on the work you need it to finish and what it costs to get an accepted result.