artificialanalysis.ai No paid content found

What is Artificial Analysis?

Artificial Analysis is an independent benchmarking and comparison site for AI models and API providers. It scores models across quality, price, output speed and latency, and also publishes leaderboards, capability indexes and trend articles aimed at helping you pick a model or provider for a specific use case.

What it covers

• Intelligence and capability indexes — composite scores for general reasoning plus professional-domain evaluations, updated as benchmarks change.
• Coding agents — a separate index for agentic coding performance.
• Image, video and speech — leaderboards beyond text models.
• Cost and speed — output tokens per second and a weighted cost-per-task figure, so you can see the trade-off between quality and spend rather than quality alone.
• Model Recommender — a tool that produces recommendations based on your own priorities across intelligence, speed and cost.

Who it is for

How to use it well

Treat the indexes as a starting filter, not a final verdict. Composite scores bury the details that matter for your workload: a model that leads on general intelligence may be slower or pricier per task than one that is good enough for your prompts. Decide which axis you actually need to optimise for, use the leaderboards to shortlist two or three candidates, then run your own evaluation on representative inputs before committing.

A concrete next step: open the Intelligence and cost-per-task views side by side, note the models that fall within your acceptable quality band, and then check their output speed against your latency budget. The site's own changelog is also worth skimming, since index versions and benchmark components change over time and can shift rankings.

How do I use the Model Recommender to choose a model for my use case?

Use the Model Recommender as a shortlist tool, not a final decision. It turns your priorities — intelligence, speed, and cost — into a ranked set of candidates from Artificial Analysis' independent benchmark data, then you verify the top few against your own workload. Artificial Analysis

How to work through it

1. Define the use case in plain terms. Is this a coding agent, a chat assistant, a document summarizer, or something with images or speech? The site separates capability indexes by domain (including coding agents, image and video, and speech), so start in the area that matches your task.
2. Rank your three trade-offs. You cannot maximize intelligence, speed, and cost at once. Decide which one is a hard constraint and which two you can flex.
3. Read the cost figure correctly. The site reports cost per task — a weighted average cost in USD for completing a task on the Intelligence Index — rather than a raw per-token price. That better reflects real spending, because a cheaper model that needs more steps can cost more overall.
4. Check speed separately. Output tokens per second and latency are distinct from intelligence; a fast model that fails your task is not cheap at any price.
5. Confirm the version and configuration. Models appear with effort or reasoning settings (for example adaptive reasoning at low, medium, or high effort). Pick the configuration you would actually deploy, not the best-scoring variant.
6. Test on your own examples. Run your ten hardest real prompts through two or three finalists and compare quality, latency, and cost.

A practical decision rule

For a concrete scenario: a support team handling thousands of short queries daily would shortlist on cost per task and speed, then check that the cheapest candidates still pass a sample of tricky tickets. A research group doing occasional deep analysis would instead shortlist on intelligence and ignore speed.

If the recommender's output feels too broad, use the custom benchmark builder to weight the metrics that match your workload, then re-rank.

Which AI model currently ranks highest on the Artificial Analysis Intelligence Index?

Claude Opus 5.5 takes the top spot on the Artificial Analysis Intelligence Index, according to the site's own changelog. Note that the leaderboard moves quickly — the page shows Intelligence Index v4.3, which swapped in AutomationBench-AA and upgraded Terminal-Bench to 4.0, so rankings can shift with each index revision.

How to read that ranking

A top intelligence score is a starting point, not a buying decision. Artificial Analysis presents several separate axes, and the leader on one is often not the leader on another:

A model at the top of the Intelligence Index can still be the wrong pick if your workload is latency-sensitive or high-volume, where a slightly weaker but cheaper, faster model often wins on total cost.

Practical next step

If you're choosing for a real project, don't stop at the headline ranking. Use the site's Model Recommender to weight intelligence, speed and cost against your own priorities, and check the Coding Agent Index and Capability Indexes if your use case is coding or a specific professional domain — those rankings can differ from the general index. Re-check before committing, since evaluations are added frequently and the index version itself changes.

For the current standing, see Artificial Analysis.

How can I compare AI models and API providers by cost, speed, and intelligence?

Use Artificial Analysis when you want a side‑by‑side read on models and API hosts rather than a single vendor's marketing page. Its core framing is three axes: intelligence (higher is better), output speed in tokens per second (higher is better), and cost per task as a weighted average in USD for Intelligence Index tasks (lower is better). It also publishes a Model Recommender for personalized recommendations across intelligence, speed, and cost, and lets you build custom benchmarks.

H3 A practical comparison method

1. Define the job. If you're building a coding agent, weight tool use and reliability over raw chat quality; if you're doing bulk classification, cost per task dominates.
2. Set a floor, not a target. Pick the minimum intelligence score your task tolerates, then optimize speed and cost within that band. Chasing the top intelligence score usually costs more than the task needs.
3. Compare like with like. Check whether reasoning effort settings, context length, or non‑reasoning variants are being compared. A high‑effort configuration and a low‑effort one are different products for budgeting purposes.
4. Validate on your own data. Public indices compress many tasks into one number; run a small sample of your real prompts before committing.

H3 Trade‑offs you'll actually feel

H3 Where providers fit in The same model served by different API hosts can differ in latency, throughput, and price, so treat model choice and provider choice as two separate decisions. Benchmark the model first, then compare hosts on the metrics that matter to your workload.

Next step: open the leaderboards, filter to your task type, and shortlist two or three configurations that clear your intelligence floor. Then run a small A/B on your own prompts before locking in a provider.

What does the Coding Agent Index measure and which agents rank highest?

The Coding Agent Index on Artificial Analysis measures how well AI coding agents perform on software-engineering tasks, scored on a per-agent basis rather than as a single headline model number. It sits alongside the site's other leaderboards (Intelligence, Image & Video, Speech, and the professional-domain Capability Indexes), so you can compare a coding agent's task performance against its speed and cost rather than treating benchmark rank as the whole picture.

What it captures

• Task-level success on coding work, aggregated into an index where higher is better.
• Agent behaviour as configured, not just the underlying model — the page's changelog shows evaluations logged for specific reasoning-effort and fallback settings (for example, adaptive reasoning at low, medium, high, xhigh and max effort).
• A moving target: the index is versioned and updated as benchmarks change. The page notes Intelligence Index v4.3 replaced one banking benchmark with AutomationBench-AA and upgraded Terminal-Bench to 4.0, so scores are not comparable across index versions.

Which agents rank highest

The page evidence does not state the current Coding Agent Index ranking or name its top agents — it only confirms the index exists and was recently updated. Treating any specific agent as the leader without checking the live leaderboard would be guesswork, especially given how often the index is revised. The most recent named results on the page concern Claude Opus 5.5 taking the top spot on the Intelligence Index, which is a different leaderboard and should not be read as a coding-agent ranking.

How to use it

If you are choosing an agent for a real workflow, filter by your constraints first — budget per task, acceptable latency, and whether your work is greenfield code or maintenance in a large existing repository — then look at where candidates land on the Coding Agent Index. A useful decision criterion: prefer an agent that stays competitive across two or three index versions rather than one that spiked in a single release, since benchmark swaps can reshuffle rankings.

Next step: open the Coding Agent Index leaderboard directly at Artificial Analysis, check the index version date, and cross-reference with the Cost per Task and output-speed columns before committing.

Can I build a custom benchmark to evaluate AI models for my specific needs?

Yes. Artificial Analysis offers a feature called Optima that lets you build a custom benchmark, alongside a Model Recommender that produces personalized recommendations weighted toward your priorities across intelligence, speed and cost. Both are described on the site as tools for matching models to a specific use case rather than relying only on a general leaderboard.

What you can customize

Based on the page evidence, the adjustable levers are:

The site also publishes capability indexes v1.1 covering six professional domains, plus separate tracks for coding agents, image, video and speech. If your work sits in one of those domains, start from the relevant index before building something bespoke.

A practical scenario

Suppose you run a customer-support assistant that answers short questions at high volume. Raw intelligence matters less than latency and cost per task, so you would weight speed and cost heavily and treat intelligence as a floor. A coding team would invert that: intelligence and coding-agent performance first, cost second.

Decision criteria

• Build a custom benchmark when your workload differs from the general mix — unusual input lengths, domain jargon, strict latency budgets or heavy volume.
• Skip it when a published capability index already matches your domain closely.
• Keep the benchmark small and re-run it when new model versions appear; the changelog shows evaluations being added and indexes revised frequently, so a fixed snapshot ages quickly.

Next step

Open Artificial Analysis and try Optima or the Model Recommender with your own weightings, then compare the result against the standard Intelligence Index to see whether your priorities genuinely change the ranking.