Keep your weekend agent alive with a one-evening model-string audit

The deadline makes a one-evening audit realistic

The first sign that a weekend agent is about to break is not a crash. It is a model string in a config file that used to work and now points at a retired model.

Perplexity's Sonar model tiers are set to retire on September 27, 2026, when the Agent API becomes the primary interface for that work. Perplexity kept its legacy Sonar endpoints operational through September 27, 2026.

Perplexity shifted the base sonar model to its Agent API on September 25, 2026, with unchanged request requirements and reduced prices. It removed routing for sonar-pro and sonar-reasoning-pro on September 27, 2026. Perplexity's Agent API does not accept the sonar-pro or sonar-reasoning-pro model identifiers, and it offers no drop-in successor for those two models. Flag those two strings for a new plan.

You have a few hours after dinner, not a migration team. Most side projects stay side projects, and that is fine. Use a one-evening model-string audit: find every place the code names a model, decide what should happen next, and test the path before the next scheduled run.

Weekend agents are usually a small set of parts: a prompt, a model string, a search call, and a timer. When one of those pieces changes, the whole thing can look broken even though the rest is fine.

Run the audit before the model string stops working

Start with the boring part: search the repo. A model string can hide in an environment file, a notebook, a scheduled job, or a search-plus-answer helper that someone copied from an example.

On LLM Gateway, the perplexity/sonar model identifier remains available from September 25, 2026. Treat the audit as a mapping exercise. Perplexity's migration table pairs Sonar with the fast preset, Sonar Pro with low, Sonar Reasoning Pro with medium, and Sonar Deep Research with high. The Agent API has been available since February 2026.

A search-plus-answer feature is the usual place this breaks. The search call may still work, but the model that turns chunks into an answer can be pinned to a retired name. An overnight job makes the first error easy to miss.

Old strings are the quiet problem. A model name in a comment is not a bug, but it becomes one when the next builder copies it.

For a scheduled agent, the test is not a command-line call. It is the same path the scheduled job takes, with the same environment variables and the same timeout.

The Perplexity MCP Server is a possible shortcut. Version 1.0.0 moved its model-backed tools from legacy Sonar models to Agent API presets, kept tool names and response shapes unchanged, and made all three tools, on average, faster and less expensive than their Sonar predecessors. The audit may be a version check.

Keep this list on the screen while you work. Each step is small enough to finish in a minute.

• Grep the repo for perplexity/sonar, sonar-pro, and sonar-reasoning-pro.
• Map each model tier to the preset named in the migration table, and flag the two higher tiers for a new plan.
• Update the code so the request shape still matches the new preset.
• Check whether the job uses the fetch_url tool, then note the current price.
• Re-run the scheduled job once, then watch the first response for an error.

Run the list once, then leave the repo in a state you would trust for a weekend.

The second check is the bill

Model sunsets can change the bill in two ways. One is the model price. The other is the tooling around it.

In the 2026-08-14 pricing-page change, Perplexity did not alter Sonar token rates, the per-1,000-request search fee, the $5.00-per-1,000 Search API, or Embeddings rates. That makes the audit easier: you are mostly changing the request path, not the bill.

Perplexity's Agent API fetch_url tool cost was $0.0005 per invocation after a reversal of the 2026-07-29 reduction to $0.00025. An agent that fetches pages on every run can let that line item outgrow the model call.

Side projects can survive a small price bump. They cannot survive a surprise invoice from a fetch loop that runs every hour.

Answering a few questions a day may keep the fetch line invisible. A queue can make it the main cost.

The Agent API can route to model families from OpenAI, Anthropic, Google, xAI, Z.AI, Moonshot AI, and NVIDIA, as well as Perplexity's Sonar model. For a weekend project, a fallback is a single config line, not a platform decision.

Use the fallback for the edge case, not the default. Keep the preset that matches the old tier, and use another family only when the preset is too slow or too expensive.