OpenAI and Codex

39.1 Configuration Surface #

Codex reads AGENTS.md natively, which makes it the reference implementation of the standard.⁷¹ Nesting works as specified—nearest file wins.

39.2 Caching Model #

OpenAI’s model changed materially with GPT-5.6, and the differences matter.⁴⁹

GPT-5.6 and later:

• Both implicit and explicit breakpoints supported. prompt_cache_options.mode selects; prompt_cache_breakpoint marks.
• Up to four cache writes per request. Lookup considers the first 2 and the latest 50 explicit breakpoints and reuses the longest match; implicit mode also considers the boundaries at the ends of messages.
• Cache writes at 1.25×, reads at 0.1×.
• TTL: 30m is the only supported value and the default.
• Minimum cacheable prefix: 1,024 visible input tokens.
• Explicit-only mode avoids writing changing suffixes, which is a genuine cost control. What is distinctive is the mode selector that turns implicit boundaries off; the underlying control—mark the stable prefix, leave the changing suffix unmarked—is the same one §22.3 teaches and the same one Anthropic’s per-block cache_control provides (§38.2).

Earlier models:

• Implicit only, breakpoints at model-dependent intervals.
• No cache write charge; model-dependent read rate.
• prompt_cache_retention of in_memory or 24h. These are maxima rather than working lifetimes: in_memory is typically 5–10 minutes against a one-hour ceiling, 24h typically around 30 minutes against a 24-hour one. On models offering both, the default follows the organization’s data-retention policy—24h without Zero Data Retention, in_memory with it. GPT-5.5 and 5.5 Pro support 24h only.
• Minimum cacheable prefix varies with the request settings; cached_tokens rounds down to a multiple of 128.

Follow the migration checklist from OpenAI’s own documentation literally: keep stable prefixes, keep prompt_cache_key values, replace prompt_cache_retention with prompt_cache_options.ttl, confirm the prefix meets the new minimum, and add an explicit breakpoint after stable content if the default breakpoint would include changing material.⁴⁹

39.3 Prompt Cache Keys #

This is the distinctive OpenAI mechanic, covered in §22.7. Cached states are machine-local and traffic above roughly 15 requests per minute can overflow. On models before GPT-5.6, prompt_cache_key influences routing so related requests reach the same cache; from GPT-5.6 the routing is automatic and the key is optional, marking off separate cache accounting instead.

const response = await client.responses.create({
  model: "gpt-5.6-sol",
  prompt_cache_key: "triage-v3:tenant-acme:shard-7",
  prompt_cache_options: { mode: "explicit", ttl: "30m" },
  input: [
    { role: "developer", content: [
      { type: "input_text", text: STABLE_INSTRUCTIONS,
        prompt_cache_breakpoint: { mode: "explicit" } },
    ]},
    { role: "developer", content: `Request time: ${new Date().toISOString()}` },
    { role: "user", content: userMessage },
  ],
});

The order runs stable content with an explicit breakpoint, then the varying developer content, then the user message. Explicit-only mode means the varying suffix is processed at the uncached rate with no write charge, which is exactly right for content that will not be reused.

39.4 Pricing #

Verified September 24, 2026, standard tier, per million tokens:⁵²

Batch is 50 percent off. Data-residency endpoints add a 10 percent uplift for models released on or after March 5, 2026. The long-context column is the threshold cliff from §21.3. It begins above 272K input tokens, and crossing it applies 2× to input, cached reads and cache writes, and 1.5× to output, for the full request.

39.5 Distinctive Features #