docs(specs): record the GPT-5.6 caching verification — it works, 10.6x on warm turns

[philmerrell]

The end-to-end proof the plan was waiting on

Four turns through the SPA on us.openai.gpt-5.6-sol, read back from GET /admin/costs/sessions/{id}/calls (session f76ab27a…):

turn	cacheStatus	input	cacheRead	cacheWrite	output	cost
1	first_write	2	0	3,679	5	$0.02038
2	hit	2	3,679	21	6	$0.00190
3	hit	2	3,700	22	6	$0.00192
4	hit	2	3,722	22	6	$0.00193

Warm turns cost 10.6× less than the cold turn. Every prediction the plan made holds, and each PR is confirmed by a specific column rather than by assertion.

What each column proves

PR-1 (#945) — usage normalization. inputTokens is 2 on every turn, not ~3,681. OpenAI reports input_tokens inclusive of both cache buckets; here they're disjoint and sum exactly to the prefix (2+0+3,679 = 3,681; 2+3,679+21 = 3,702), matching the SPA's own context-window readout to the token.

cacheWriteInputTokens is populated at all only because we recover cache_write_tokens — Strands drops it. Turn 1 alone would otherwise have double-billed 3,679 tokens at the input rate plus the 1.25× premium. That is precisely the correction #947 made to this spec, now demonstrated on real data rather than argued from a doc.

PR-2 (#949) — transport. Reached a live model: base URL, per-request bearer token, inference-profile id all correct.

PR-5 (#951) — model-derived TTL. Gaps of 30s / 78s / 18s classified hit, wastedUsd = $0.00 throughout, no avoidable re-writes.

#956 — explicit caching OFF. This is the implicit shape the probe predicted: cacheRead grows with the conversation (3,679 → 3,700 → 3,722) while cacheWrite is just the appended delta (~21). Under explicit mode cacheRead would be flat and uncached input would climb every turn. The 57%-worse finding is now confirmed in the agent loop, not just at the transport — so shipping it off was right.

#959 — IAM. bedrock:CallWithBearerToken, without which none of the above runs.

Two extra confirmations

Prefix fingerprints behaved exactly as the cache contract requires — toolConfigHash 8eafb0765ed810a2 and systemPromptHash 5a71749b3bee37ab identical across all four calls, historyHash changing each turn. Nothing nondeterministic leaked into the cacheable prefix.

The cost math reproduces to the cent. Deriving each row from the disjoint buckets and the catalog rates matches the reported cost exactly on all four calls, so CostCalculator is verified against live usage too.

⚠️ The dollars are provisional

They use the model-card rates seeded on the dev row (4.40 in / 26.40 out / 0.44 cache read / 5.50 cache write), and 26.40 is derived from the GovCloud 1:6 input:output ratio, not published. The ratios above are real and independent of that; the absolute dollars wait on PR-3, which is still blocked on AWS not publishing commercial rates.

Docs only — no code.

🤖 Generated with Claude Code