Claude Fable 5.2 is the best AI model for agents. GPT-6.1 Terra scores 51.3 for 5% of the cost, and Kimi K3.1 Pro gets through each step fastest.
Claude Fable 5.2 is the best AI model for agents. GPT-6.1 Terra scores 51.3 for 5% of the cost, and Kimi K3.1 Pro gets through each step fastest.
32 models · Updated November 2026 · Data from Artificial Analysis
Claude Fable 5.2
Anthropic · Max effort
The highest score of any model we compared at this setting.
GPT-6.1 Terra
OpenAI · Max effort
79% of Claude Fable 5.2's score for 5% of the cost.
Kimi K3.1 Pro
Moonshot AI · Standard effort
The quickest full answer from a model scoring 46 or more.
Over 45 s at every tested setting
Claude Fable 5.2
Anthropic · Max effort
The highest score of any model we compared at this setting.
GPT-6.1 Terra
OpenAI · Max effort
79% of Claude Fable 5.2's score for 5% of the cost.
Kimi K3.1 Pro
Moonshot AI · Standard effort
The quickest full answer from a model scoring 46 or more.
Each job favors a different model. Pick the one closest to yours.
Long jobs where one wrong step is expensive
Use
Claude Fable 5.2
Anthropic · Max effort
The highest intelligence score, so it plans better and recovers from bad tool results more often. $1,035 per 1,000 tasks.
Long jobs where one wrong step is expensive
Use Claude Fable 5.2
The highest intelligence score, so it plans better and recovers from bad tool results more often. $1,035 per 1,000 tasks.
High-volume agents that run in the background
Use GPT-6.1 Terra
$50.27 per 1,000 tasks, about 5% of what Claude Fable 5.2 costs.
Agents someone is waiting on
Use Kimi K3.1 Pro
85 seconds per step. A 20-step run takes about 28 minutes, against 82 minutes with Claude Fable 5.2. It scores 46.5, so give it clear, well-scoped jobs.
Sub-agents for lookups and simple steps
Use Muse Ember
$6.36 per 1,000 tasks. Let a stronger model plan and hand the searching, reading and formatting to this one.
Staying with OpenAI
Use GPT-6 Astra
The highest-scoring OpenAI model: 56.9 at max effort, with 5 minutes per step. At medium effort it answers in 22 seconds and scores 51.6.
Coding agents
Agents that write and fix code have their own ranking, with Claude and OpenAI picks.
Setting up the agent itself
Generate a starting config for your agent framework, with the model, tools and limits filled in.
Each dot is a model. Higher scores better, further left is cheaper, so the best deals sit top left. The cost axis is logarithmic: every step left is a big saving. Hover a dot for its numbers.
Sort by the pillar you care about, and set your monthly volume to see the bill. Costs include each model's thinking tokens at the setting shown.
An agent calls the model again and again: plan, call a tool, read the result, decide the next step. The wait for each answer is multiplied by the number of steps, so a model that feels fine in chat can make an agent slow.
For a 20-step run at max effort: Claude Fable 5.2 takes about 82 minutes, GPT-6 Astra takes about 2 hours, Kimi K3.1 Pro takes about 28 minutes, GPT-6.1 Terra takes about 2 hours. At high effort, where each answer comes back within 45 seconds, Claude Fable 5.2 gets through the same run in about 12 minutes and scores 61.8 instead of 64.6. If a person is waiting, speed can matter more than the last few points of intelligence.
Most agent steps are routine: search, read, extract, format. Let Claude Fable 5.2 plan and check the work, and hand those steps to GPT-6.1 Terra. If four in five calls go to GPT-6.1 Terra, 1,000 tasks cost about $247 instead of $1,035, roughly 76% less.
We rank each model at max effort, the setting that gives its best score and the one Artificial Analysis and the labs quote. Cost and answer time are shown at that same setting, so you can see what the top score costs. Switch to Answers under 45 s above to see each model at its strongest setting that gives a full answer within 45 seconds. Effort labels mean different things at different labs, so that view picks each setting by wait time, not by name.
Intelligence is the Artificial Analysis Intelligence Index at that setting, which includes agentic tasks alongside coding, knowledge and reasoning. There's no tool-calling score published for every model, so we don't rank on one.
Cost starts from a typical agent task of 30,000 tokens, 60% of it context the agent reads and 40% what it writes. We scale the writing part by how much each model wrote, thinking included, in Artificial Analysis's tests.
Speed is the median time to a full answer in the same tests. For an agent, read it as time per step.
Best value is the most intelligence per dollar and fastest is the quickest step, both among models scoring 46 or more (within 30% of the leader). Neither pick goes to a model its maker has replaced.
Claude Fable 5.2 at max effort. It scores 64.6 on the Artificial Analysis Intelligence Index, the highest of the models we compared, and costs about $1,035 per 1,000 agent tasks. GPT-6.1 Terra gives the most intelligence per dollar at $50.27. If you need answers in under 45 seconds, it scores 61.8 at high effort for $449 per 1,000 tasks.
Claude Fable 5.2, as far as public data shows. Artificial Analysis doesn't publish a tool-calling score for every model, so we use its Intelligence Index, which includes agentic tasks. Claude Fable 5.2 leads it at 64.6.
Muse Ember, at about $6.36 per 1,000 agent tasks. It scores 36, so use it for simple steps or as a sub-agent under a stronger planner.
Kimi K3.1 Pro, at about 85 seconds per step. A 20-step run takes about 28 minutes, against 82 minutes with Claude Fable 5.2 at max effort. With every model held to answers under 45 seconds, GPT-6 Astra is quickest, at about 22 seconds for a score of 51.6.
Enough to hold the task, the tool results and the steps so far, which grows quickly on long runs. All three of our picks accept at least 1 million tokens, so context is rarely the limit. Cost is the one to watch, because the whole history is sent again on every step.
For unattended runs on hard problems, it can be worth it. Claude Fable 5.2 takes about 4 minutes per answer at max effort and costs $1,035 per 1,000 tasks. At high effort it answers in 36 seconds for $449 and scores 61.8 instead of 64.6. For agents a person is waiting on, use a setting that answers in under 45 seconds.
Answer three questions and get a pick for each part of your workload, with the monthly cost.