Skip to main content

​
What’s New

OpenAI’s GPT-5.6 family ships in three tiers, each with a pro variant:
ModelTier
openai/gpt-5.6-solFlagship capability
openai/gpt-5.6-sol-proSol with pro reasoning
openai/gpt-5.6-terraBalanced performance and cost
openai/gpt-5.6-terra-proTerra with pro reasoning
openai/gpt-5.6-lunaFastest, most affordable
openai/gpt-5.6-luna-proLuna with pro reasoning
There are no breaking changes. Existing requests keep working as-is — you can swap the model slug and stop there. The rest of this guide covers the additive pieces worth adopting: These additions are supported by OpenAI GPT-5.6 and newer. Explicit prompt caching is only available through the Responses API; reasoning.mode and reasoning.context work on both the Chat Completions and Responses APIs.
You don’t need to filter these fields per model. OpenRouter strips reasoning.mode, prompt_cache_breakpoint, and prompt_cache_options before forwarding requests to providers that don’t support them, so requests that fall back to other models keep working.

​
Pro Mode with reasoning.mode

GPT-5.6 pro variants apply deeper multi-pass reasoning before returning a single final answer. Use them for hard, quality-first tasks where latency and token usage matter less. There are two equivalent ways to request pro mode:
  1. Send reasoning.mode: "pro" with the standard model — OpenRouter reroutes the request to the matching *-pro model listing.
  2. Call the *-pro model listing directly.
{
  "model": "openai/gpt-5.6-sol",
  "input": "Prove that there are infinitely many primes.",
  "reasoning": {
    "mode": "pro"
  }
}
mode is independent of effort, so you can combine mode: "pro" with any supported effort level. Pro mode bills at the same per-token rates as standard mode but typically consumes more tokens. See Reasoning Mode for details.

​
Persisted Reasoning with reasoning.context

When you echo reasoning items back in conversation history, reasoning.context controls which reasoning the model can use:
ValueBehavior
omitted / autoThe model’s default context mode
all_turnsThe model can reference reasoning from all turns in the input
current_turnOnly reasoning from the current turn is used; prior reasoning items are ignored
Use all_turns for multi-turn work where goals stay stable and you want the model to build on its earlier chain of thought. Use current_turn for a fresh reasoning pass. 
{
  "model": "openai/gpt-5.6-sol",
  "input": [
    { "role": "user", "content": "Solve this math problem step by step: ..." },
    {
      "type": "reasoning",
      "id": "rs_abc123",
      "encrypted_content": "...",
      "summary": [
        { "type": "summary_text", "text": "Analyzed the equation..." }
      ]
    },
    {
      "type": "message",
      "role": "assistant",
      "content": [{ "type": "output_text", "text": "The answer is 42." }]
    },
    { "role": "user", "content": "Now explain it differently." }
  ],
  "reasoning": {
    "effort": "high",
    "context": "all_turns"
  },
  "include": ["reasoning.encrypted_content"]
}
The default value of context may vary by model, so set it explicitly if your use case needs a specific behavior. See Reasoning Context Mode for details.

​
Explicit Prompt Caching

Automatic prompt caching still works with no code changes. Explicit caching adds direct control over cache boundaries instead of relying on OpenAI’s automatic breakpoint placement. Cached prefixes have a minimum 30-minute TTL.
Explicit caching is not available in the Chat Completions API yet. The prompt_cache_breakpoint and prompt_cache_options fields only exist in the Responses API request format. Chat Completions requests still get automatic prompt caching — you just can’t control the breakpoints.
Two controls:
  • prompt_cache_breakpoint: placed on an individual input_text content block to mark the end of a reusable prefix. Everything through that block becomes the candidate cached prefix. Automatic caching stays enabled.
  • prompt_cache_options: placed at the request root. Setting mode to "explicit" disables OpenAI-managed breakpoints so only blocks marked with prompt_cache_breakpoint participate in caching. Use ttl to request a cache duration (e.g. "30m").
{
  "model": "openai/gpt-5.6-sol",
  "prompt_cache_key": "my-session-key",
  "prompt_cache_options": {
    "mode": "explicit",
    "ttl": "30m"
  },
  "input": [
    {
      "role": "user",
      "content": [
        {
          "type": "input_text",
          "text": "<REUSABLE_PREFIX>",
          "prompt_cache_breakpoint": {
            "mode": "explicit"
          }
        },
        {
          "type": "input_text",
          "text": "<TASK_SPECIFIC_SUFFIX>"
        }
      ]
    }
  ]
}
See Prompt Caching for details.

​
Cache-Write Billing

For GPT-5.6 and later, OpenAI bills cache writes at 1.25x the model’s uncached input rate. Cache reads keep the discounted cache-read rate. Cache activity is reported in the Responses API’s usage.input_tokens_details:
  • cache_write_tokens: prompt tokens written to the cache
  • cached_tokens: prompt tokens read from the cache
Track both to understand net caching cost. If your traffic rarely re-reads cached prefixes, explicit mode lets you avoid unnecessary cache writes.

​
Migration Checklist

  1. Pick your tier: openai/gpt-5.6-sol for frontier capability, openai/gpt-5.6-terra for balanced cost, openai/gpt-5.6-luna for high-volume workloads.
  2. Keep your existing reasoning effort as the baseline, then compare one level lower — GPT-5.6 is more token-efficient than earlier generations, so lower settings often hold quality.
  3. Enable reasoning.mode: "pro" only where evaluations show a quality gain worth the extra latency and tokens.
  4. Set reasoning.context explicitly in multi-turn workflows that echo reasoning items back.
  5. Watch cache_write_tokens and cached_tokens in usage. Switch to explicit caching if automatic cache writes are costing more than the reads save.

​
Breaking Changes

None. All GPT-5.6 API additions are optional:
  • Requests without reasoning.mode, reasoning.context, or prompt_cache_options behave the same as before
  • Automatic prompt caching continues to work without configuration
  • Standard-mode billing rates are unchanged; cache-write billing only applies when prompts are written to the cache

​
Resources