Sonnet 5.5 cache reads now $0.10/M: long agent session cost

Claude Sonnet 5.5 cache hits are listed at $0.10 per million tokens on Anthropic's pricing page read on 2026-10-13, which is 0.05 times its $2 base input price. A turn that reuses 150,000 cached tokens, adds 10,000 fresh input tokens and writes 4,000 output tokens costs $0.075 at that rate, down from $0.090 at $0.20.

Sume's rate table does not show the new number yet. The repository's Sonnet 5.5 tier on origin/main holds a cache-read rate of $0.20 per million, stamped anthropic_list_2026-09-28. This page works out what the difference means for a long agent session, and which receipt field to trust.

What changed on the vendor page

Anthropic's table lists Sonnet 5.5 at $2 per million input tokens, $2.50 for 5-minute cache writes, $4 for 1-hour cache writes, $0.10 for cache hits and refreshes, and $10 per million output tokens. A footnote says cache hits on Opus 5.5 and Sonnet 5.5 are priced at 0.05 times the base input price. Opus 5.5 at $4 base therefore stays at $0.20, and only Sonnet 5.5 moves from the figure the Sume table carries.

Writes and base input are unchanged against that table. Only the cache-hit line differs, so the saving appears only on turns that read a large cached prefix.

Cost of one turn at three cache sizes

Each row assumes 10,000 fresh input tokens at $2 per million ($0.020) and 4,000 output tokens at $10 per million ($0.040) on top of the cache read. Long agent threads carry a big prefix of system prompt, tool schemas and earlier turns, so the cached part usually dominates the input.

What it does to a 40-turn session

Run the 150,000-token row for 40 turns and the session costs $3.00 at $0.075 against $3.60 at $0.090, a $0.60 difference. At 100 sessions a day that is $60 a day. The saving grows with the cache size and shrinks once output tokens dominate, because output is still $10 per million.

The check is a few lines. Edit the token counts to match a receipt from your own workload:

def turn(cached: int, fresh: int, out: int, cache_rate: float) -> float:
    return cached * cache_rate / 1e6 + fresh * 2.0 / 1e6 + out * 10.0 / 1e6

for rate in (0.20, 0.10):
    one = turn(150_000, 10_000, 4_000, rate)
    print(f"cache ${rate:.2f}/M: ${one:.3f} per turn, ${one * 40:.2f} for 40 turns")

Which number to trust on a Sume receipt

Do not plan a Sume run from any list price, new or old. Sume's Errors and spend page says usage.cap.counted_credits covers metered generation and does not include the agent's own LLM turn. The record of what the run cost in total is usage.debited_credits, and the billing records are GET /v1/usage and GET /v1/balance.

So the practical steps are short. Set generation_spend_cap_credits for the generation you expect, as described under Spend caps on Create a run, and it has no default on Agent Completions. Then read usage.debited_credits on a finished run of the model you use. Compare that to your token estimate rather than to this page. If the Sonnet 5.5 rate in Sume's table is updated to $0.10, a rerun of the same brief is the cleanest way to see the move.

• Cache hits at $0.10 versus $0.20 matter most past about 100,000 cached tokens per turn.
• Opus 5.5 is unchanged at $0.20 per million cache hits on Anthropic's page.
• The cap bounds generation, not tokens, so tokens are checked on usage.debited_credits.