Gpt 56 Explicit Prompt Cache Controls Cost Proof 2026

• Date: 2026-07-20 (UTC)
• Script: scripts/gpt56-explicit-cache-probe.py
• Artifact: data/lab-runs/gpt-56-explicit-prompt-cache-controls-cost-proof-2026.openai.json
• Endpoint: https://api.openai.com/v1/responses
• Models: gpt-5.6-luna, gpt-5.5-2026-04-23
• Purpose: measure what GPT-5.6's explicit prompt-cache controls actually change versus implicit caching, and versus GPT-5.5's 24h retention
• Prefix: 9,208 characters of synthetic, invented "operations handbook" text, SHA-256 80761fcc454ee750... (full digest in the JSON artifact)
• Budget guard: every call passed through scripts/proof_budget.py (check_budget before, record_usage after)
• Safety boundary: no confidential, customer, credential, or private data. The handbook text is entirely fictional.

Command

python3 scripts/gpt56-explicit-cache-probe.py

Configurations

Identical prefix and identical question in every configuration. 4 calls per config (B2: 1 call), ~2 seconds apart.

Raw output

[A_gpt56_implicit] call 1: input=2032 cached=0 cache_write=2029
[A_gpt56_implicit] call 2: input=2032 cached=2029 cache_write=None
[A_gpt56_implicit] call 3: input=2032 cached=2029 cache_write=None
[A_gpt56_implicit] call 4: input=2032 cached=2029 cache_write=None
[B_gpt56_explicit] call 1: input=2032 cached=0 cache_write=2001
[B_gpt56_explicit] call 2: input=2032 cached=2001 cache_write=None
[B_gpt56_explicit] call 3: input=2032 cached=2001 cache_write=None
[B_gpt56_explicit] call 4: input=2032 cached=2001 cache_write=None
[B2_gpt56_explicit_ttl_24h] call 1: ERROR HTTP 400: {
  "error": {
    "message": "Invalid value: '24h'. Supported values are: '30m'.",
    "type": "invalid_request_error",
    "param": "prompt_cache_options.ttl",
    "code": "invalid_value"
  }
}
[C_gpt55_retention_24h] call 1: input=2032 cached=0 cache_write=None
[C_gpt55_retention_24h] call 2: input=2032 cached=1792 cache_write=None
[C_gpt55_retention_24h] call 3: input=2032 cached=1792 cache_write=None
[C_gpt55_retention_24h] call 4: input=2032 cached=1792 cache_write=None

Measured summary

What worked

• Both GPT-5.6 modes produced a cache hit on the very next call, with no warm-up delay beyond the ~2s gap between requests.
• The explicit breakpoint was accepted on the developer-role content block and produced a stable, repeatable cached prefix of exactly 2,001 tokens across all three repeat calls.
• GPT-5.5 with prompt_cache_retention: "24h" also hit on the second call, at a smaller cached prefix (1,792 tokens).

What failed

• prompt_cache_options.ttl: "24h" was rejected with HTTP 400: Invalid value: '24h'. Supported values are: '30m'. The 24-hour cache lifetime that GPT-5.5 offers has no equivalent on GPT-5.6 in this run.

What is new on 5.6

• cache_write_tokens appeared as a populated field in usage.input_tokens_details on the first GPT-5.6 call of each config, and was absent on GPT-5.5. Per OpenAI's pricing page, cache writes bill at 1.25x the uncached input rate.

Limitations

• Single prefix, single account, single region, one session on 2026-07-20. Cache behaviour is infrastructure-dependent and may vary by tenant, load, and time.
• 4 calls per config is enough to show hit/miss and the write field, not enough to characterise eviction, cross-machine routing, or hit rate under production concurrency.
• Only gpt-5.6-luna was tested. Sol and Terra were not exercised.
• Repeat calls were ~2 seconds apart, so nothing here tests the 30-minute TTL boundary itself.
• All dollar figures in the published article are arithmetic over OpenAI's published price table and these measured token counts. No billing statement was read; no dollar amount was directly measured.