Freeze Perplexity Fast for Reproducible Tests

What changed in August

Perplexity’s changelog labels the change “August 2026” but does not provide an exact day. It says dynamic fast requests automatically pick up GPT-5.6 Luna, minimal reasoning effort, and priority processing. Priority processing uses twice the model’s standard token prices.

The same monthly changelog says presets now receive stable prompt-cache keys automatically. Perplexity estimates that caching can reduce costs by about 5% for applications that frequently use presets, depending on cache utilization. That is an estimate conditional on cache use—not a guaranteed discount for a single request.

The important measurement lesson is simple: a friendly preset name is an alias for a provider-managed bundle. It is not a durable version identifier.

Current fast snapshot

We checked Perplexity’s current preset documentation on August 24, 2026. Its frozen examples expose the following operational values:

Perplexity’s frozen examples also include the complete system instructions. Preserve those instructions or a cryptographic hash of the exact text. Recording only the model leaves a large part of the request contract undocumented.

Choose dynamic or frozen deliberately

For visibility research, “both” is usually the strongest design. Run a dynamic arm to observe the product users receive and a frozen arm to distinguish alias drift from query or web drift. Do not average the arms into one score.

A minimum reproducibility record

{
  "observed_at_utc": "2026-08-24T00:00:00Z",
  "endpoint": "/v1/agent",
  "mode": "frozen",
  "requested_preset": "fast",
  "model": "openai/gpt-5.6-luna",
  "service_tier": "priority",
  "reasoning_effort": "minimal",
  "max_steps": 1,
  "max_output_tokens": 8192,
  "prompt_cache_key": "fast",
  "tools": [{"type": "web_search", "max_results": 10}],
  "prompt_sha256": "record-your-own-hash",
  "system_instructions_sha256": "record-your-own-hash",
  "run_id": "provider-response-id",
  "latency_ms": "measured-client-side",
  "usage": "copy-provider-usage-object",
  "citations": "preserve-ordered-resolved-urls"
}

Replace the placeholders with observed values; do not publish them as results. Save the raw response before normalizing URLs. A normalization step can merge tracking variants, but the original URL is necessary when debugging redirects, freshness, and source identity.

A three-prompt control test

1. Stable fact: choose a question whose answer and primary source are unlikely to change during the test window.
2. Fresh fact: choose a time-sensitive question with an authoritative dated source.
3. Source-sensitive comparison: ask for a comparison that requires multiple official documents.

Run each prompt three times in the dynamic arm and three times in the frozen arm within a bounded window. Record answer completion, exact cited URLs, citation order, claim-source fit, elapsed time, token usage, cache usage if reported, and calculated cost. With only three prompts, the output is a diagnostic artifact—not a platform benchmark.

Our guide to latest-model aliases explains the same versioning risk across providers. The small-experiment method shows why interface and client state also belong in the observation record.

What freezing cannot control

• the live web, changed pages, deleted pages, robots rules, or index freshness;
• search ranking and candidate retrieval behind the API;
• undocumented provider infrastructure changes;
• regional, account, rate-limit, or availability differences;
• the behavior of Perplexity’s consumer web or app products.

A frozen request is therefore a stronger control, not a time machine. Archive the prompt, configuration, response, citations, timing, usage, and documentation snapshot together.

When a provider changes an alias, open a new observation period instead of rewriting old records. Keep the previous configuration and documentation capture so a later reviewer can reconstruct which contract each run used.

Sources, method, and limits

Primary sources: Perplexity API changelog and Perplexity Agent API preset documentation, checked August 24, 2026.

Method: We copied explicit fields from the current frozen fast examples and separated provider-documented behavior from the test fields a practitioner must collect.

Limits: Perplexity gives no exact day for the August preset change. The displayed values are a dated documentation snapshot and may change. We did not send a paid API request for this article, so it reports no latency, citation, or cost result.