For an agent that runs on a schedule, use Claude Opus 5.

Why Claude Opus 5

The reasons it wins for this job.

Anthropic built Opus 5 for complex agentic work and recommends starting there. It scores 51 on the Artificial Analysis Intelligence Index, against 53 for Fable 5.1, at half the per-token price.
Task budgets tell the model how many tokens a loop may spend, so it paces itself and wraps up instead of being cut off mid-task.
Managed Agents runs the loop and a per-session sandbox on Anthropic's side, with scheduled deployments, so a weekly job needs no server of your own.
You can delegate reading-heavy sub-steps to Sonnet 5 workers and keep Opus 5 for the decisions, which cuts cost without lowering the quality of the final call.

How to use it

4 steps to a first result.

1Start from a workflow, not an agentIf the steps are always the same, write them as code with a model call at each step. Use an agent only where the path genuinely varies.
2Define done and a budgetGive each run an outcome to check and a token budget. Log both.
3Keep permissions narrowTools that write or send need approval, or a separate confirmation step outside the model.
4Schedule it and review a sampleRun it on a schedule and read five transcripts a week. Regressions show up there before they show up in metrics.

Claude API, TypeScript (tool loop)

Start from this.

Edit the parts in capitals, then run it.

import Anthropic from "@anthropic-ai/sdk";

const client = new Anthropic();
const tools: Anthropic.Tool[] = [/* YOUR TOOLS: name, description, input_schema */];
const messages: Anthropic.MessageParam[] = [{ role: "user", content: "GOAL AND DEFINITION OF DONE" }];

while (true) {
  const res = await client.messages.create({
    model: "claude-opus-5",
    max_tokens: 16000,
    output_config: { effort: "high" },
    tools,
    messages,
  });
  messages.push({ role: "assistant", content: res.content });
  if (res.stop_reason !== "tool_use") break;

  const results: Anthropic.ToolResultBlockParam[] = [];
  for (const block of res.content) {
    if (block.type !== "tool_use") continue;
    const output = await runTool(block.name, block.input); // YOUR CODE
    results.push({ type: "tool_result", tool_use_id: block.id, content: JSON.stringify(output) });
  }
  messages.push({ role: "user", content: results }); // all results in one message
}

Nothing is sent anywhere. It copies to your clipboard.

Alternatives that also work

If Claude Opus 5 is not an option.

GPT-5.6 SolOfficial page →

OpenAI · closed · cost: medium

Pick it when your stack is on OpenAI and cost matters more than the top score. Sol is OpenAI's flagship for professional work at a medium cost tier.

GLM-5.3Hugging Face →

Z.ai · open weights · cost: free weights; you pay for the hardware · 838K HF downloads/mo

Pick it when the agent must run on your own GPUs. It posts the best open-weight Terminal-Bench 3.0 score on its model card.

Claude Sonnet 5Official page →

Anthropic · closed · cost: medium

Pick it when the steps are simple lookups and edits. It also works well as the worker under an Opus 5 lead.

Watch out for

Most agent failures in production are tool failures: a timeout, an expired token, a changed API. Return errors to the model as tool results instead of crashing the run.
Append to the conversation, never edit earlier turns. Newer Claude models reject edited histories that carry thinking blocks.

Doing this just once? For a long, one-off task an agent should run end to end, use Claude Fable 5.1. A one-off agent task, such as auditing a site, migrating data between tools or assembling a report from ten systems, fails on the long tail of steps. Fable 5.1 is Anthropic's model for long-horizon agentic work, and for one run its higher price barely registers.

Sources

Checked 17 September 2026. Models change monthly; we re-check this page when they do.

Picking the model is the easy part.

Wiring it into a workflow that runs every week, with evals, fallbacks and a cost you can predict, is the work. Fifteen minutes, no deck.

Book fifteen minutes →