AISource: OpenAI

OpenAI adds optional 12-hour prompt caching to the Responses API

OpenAI added 12-hour prompt caching for requests that reuse context hours apart. Cache writes cost 2.5 times the ordinary input-token rate, making repeated reuse key to savings.

By News DeskPublished here Thu, Oct 22, 2026, 3:25 PM UTC (UTC)1 min read

OpenAI introduced optional 12-hour prompt caching for the Responses API on October 20, 2026. The feature keeps shared instructions, tools and context eligible for discounted cached-input billing across requests separated by hours, according to OpenAI. Sources: OpenAI

For GPT-5.6 and later, developers enable the option by setting prompt_cache_options.ttl to "12h"; the default remains "30m". Eligible prefixes require at least 1,024 visible input tokens. Sources: OpenAI

OpenAI lists total write charges at 2.5 times the ordinary input-token rate for 12-hour caching and 1.25 times for 30-minute caching. At GPT-6.1 Sol’s standard, short-context rates, one million tokens written once and fully read twice calculates to $5.20, versus $6.00 for three uncached passes. A 30-minute cache serving both reads totals $2.70; these calculations exclude output, changing prompt content and additional writes. Sources: OpenAI

OpenAI defines 12 hours as a minimum period of billing eligibility, without guaranteeing physical cache reuse or lower latency. It is not a deletion deadline and does not change the maximum data-retention policy. Sources: OpenAI

Why it matters

Developers can keep shared context eligible for discounted input billing across hours-long gaps. The higher write price means 12-hour caching needs enough reuse to justify its premium over the 30-minute option.

Attribution

Sources

1OpenAIChangelog | OpenAI APISource published Tue, Oct 20, 2026, 12:00 AM UTC (UTC)
2explainx.aiOpenAI 12-Hour Prompt Caching: 2.5x Write, VerifiedSource published Wed, Oct 21, 2026, 12:00 AM UTC (UTC)
3OpenAIPrompt caching | OpenAI API
4OpenAIPricing | OpenAI API

Read the original source for full details and context.

All AI stories

Review · AI

Exa 3.0 adds lower-cost Lite search and generally available Dynamic Highlights

Exa 3.0 adds Lite, a search tier priced below Auto, to Exa’s web search and content retrieval API for AI applications. Exa announced the update on October 19, 2026, and said it was immediately available through its API. The company says Dynamic Highlights is now generally available. It selects text across results to reduce duplication and allocate more space to useful sources, according to Exa. Exa also says it added a larger embedding model, expanded coverage and built Lite to use Auto’s index with less retrieval compute.

ExaOct 22, 2026

Article · AI

vLLM 0.32.0 adds structured decisions: how to compare them with constrained JSON

vLLM 0.32.0 enables an endpoint that reads model scores for predefined answers. A proposed document-triage test compares that approach with constrained JSON across accuracy, calibration, latency, and fallback handling.

vLLM on GitHubOct 22, 2026

Article · AI

Build a Voxtral 3 booking prototype that checks corrections before committing

Mistral’s Voxtral 3 booking demonstration shows a corrected reference before a lookup. This guide proposes a prototype to test whether corrections also prevent obsolete bookings, with controls for delayed tools, confirmation and database writes.

Mistral AIOct 21, 2026