25 September 2026 · AI & LLMs · 5 min read
Flagship models almost never get cheaper at launch — they get better and cost the same, or both go up together. Anthropic broke that pattern on September 22 when Opus 5.5 shipped at $4 per million input tokens and $20 per million output, down from Opus 5's $5/$25. That's the headline number every pricing tracker ran with. The number I actually care about is buried a line further down the rate card: cache reads dropped 60%, from $0.50/MTok to $0.20/MTok. For anything that leans on prompt caching — which is most serious agent and coding workloads — that second cut moves the bill more than the first one does.
Checked against Anthropic's published pricing (platform.claude.com/docs/en/about-claude/pricing) and corroborated across independent pricing trackers, 2026-09-25. All figures per million tokens.
| Input | 5m cache write | Cache read | Output | |
|---|---|---|---|---|
| Claude Opus 5 | $5.00 | $6.25 | $0.50 | $25.00 |
| Claude Opus 5.5 | $4.00 | $5.00 | $0.20 | $20.00 |
| Change | -20% | -20% | -60% | -20% |
Input, cache write, and output all fell by the same 20% — a clean, proportional cut across the base rate card, same shape as when Sonnet 5 undercut its own predecessor at launch. Cache reads are the outlier: instead of holding at the usual 0.1x-of-input ratio (which would put a $4 input model's cache read at $0.40), Anthropic cut it to $0.20 — a 0.05x ratio, twice as cheap relative to input as the standard discount. Anthropic also says Opus 5.5 uses fewer tokens to finish the same task and generates output more than 30% faster than Opus 5, which is a real cost lever but isn't in the rate card itself — the numbers below are the rate card only, not that efficiency gain.
Take a coding-agent workload that's exactly the shape prompt caching is built for: a 6,000-token repo context resent on every turn, 20 turns inside a single 5-minute session, 50,000 sessions a month — the same scenario I used to price Sonnet 5's caching. On Opus 5, a cache write costs 1.25x input ($6.25/MTok) and a hit costs 0.1x input ($0.50/MTok). On Opus 5.5, write is $5.00/MTok and a hit is $0.20/MTok.
| Per 20-turn session | At 50,000 sessions/mo | |
|---|---|---|
| Opus 5 (1 write + 19 hits) | $0.0945 | $4,725.00 |
| Opus 5.5 (1 write + 19 hits) | $0.0668 | $3,340.00 |
That's $1,385/mo saved, a 29.3% cut — bigger than the 20% headline rate drop, because 19 of every 20 calls in this session shape are cache reads, and reads got cut three times as deep as everything else. The more a workload leans on caching, the more the 60% cut dominates over the 20% one; a workload with little or no caching only sees the smaller number.
A document-analysis pipeline processing 2 million input tokens and 200,000 output tokens a day, no caching involved — a fair comparison of the base rate alone.
| Input (2M/day) | Output (200K/day) | Monthly (30 days) | |
|---|---|---|---|
| Opus 5 | $10.00 | $5.00 | $450.00 |
| Opus 5.5 | $8.00 | $4.00 | $360.00 |
$90/mo, exactly the 20% the headline number promises — no caching in the mix means no extra lift from the read-price cut. That's the honest baseline: if your workload doesn't cache, budget for 20% off, not 30%.
If you're already on Opus 5 and not using prompt caching on repeated context (a system prompt, a tool schema, a repo snapshot, retrieved documents that don't change turn to turn), that's the bigger miss here — not the model version. Turn caching on first, upgrade to 5.5 second, and the two compound: in the session example above, an uncached Opus 5 session costs $0.24/session in base rates alone, versus $0.0668 for a fully cached Opus 5.5 session — a 72% drop from combining both, not either alone. If your workload is pure one-shot completions with no repeated context, the 20% base-rate cut is the whole story and it's still worth the swap, since Anthropic's own efficiency claims (fewer tokens, 30% faster) come free with the version bump.
New to usage-based LLM pricing? Start with the free API-cost guides.
Deploy it yourself: DigitalOcean — $200 free credit ↗ · Hostinger VPS ↗
Pricing checked 2026-09-25 against Anthropic's official Claude API pricing docs (platform.claude.com/docs/en/about-claude/pricing) and cross-checked against independent pricing coverage of the September 22, 2026 Opus 5.5 launch. Reference estimates using Anthropic's published self-serve rates — Bedrock, Google Cloud and Microsoft Foundry pricing, enterprise volume discounts and future rate changes vary; confirm current pricing at the source before budgeting. Workload figures (request volumes, token counts) are illustrative scenarios computed against real published per-token rates, not vendor-supplied numbers. The 30%-faster/fewer-tokens efficiency claim is Anthropic's own and is not reflected in the per-token dollar figures above, which use rate-card prices only.