Anthropic released Claude Opus 5.5 on September 22, a little over two months after Opus 5 shipped in July. No keynote, no big launch post — just an updated pricing page and a model card. The number that actually matters if you're running Claude inside an automation: $4 per million input tokens and $20 per million output, down from Opus 5's $5 and $25.
What actually got cheaper?
The base rate dropped a flat 20% on both input and output. Cache writes dropped by roughly the same amount — $5 for a five-minute cache versus $6.25 on Opus 5, $8 for a one-hour cache versus $10. Cache reads dropped the most: $0.20 per million tokens against $0.50, a 60% cut.
That's the number behind Anthropic's claim that Opus 5.5 costs "40% less than Opus 5 on typical workloads," and it's worth being precise about what it means. It isn't a blanket 40% off every token. It's a 60% cut to the one line item that dominates the bill in a multi-step agent loop — the kind that keeps re-reading the same system prompt or reference document at every step. Read a document once and move on, and you'll see the flat 20%. Run an agent that loops back to the same context five or six times per task, and you land much closer to the advertised 40.
What does this look like on a workflow we've already priced?
We ran this exact math over the summer for a contract-review agent — ten contracts a day, 30,000 tokens in per contract, 1,500 tokens out for the risk summary — at $5/$25 per million tokens, the rate Opus 5 launched at in July and Opus 4.8 already used. That worked out to $1.88 a day, about $56 a month, in our post on what an AI agent actually costs to run.
Same workflow on Opus 5.5's $4/$20 rate: (10 × 30,000 ÷ 1,000,000 × $4) + (10 × 1,500 ÷ 1,000,000 × $20) = $1.20 + $0.30 = $1.50 a day, about $45 a month.
Eleven dollars a month won't move anyone's budget on its own. The direction is the point — every Opus-tier workflow running today just got cheaper without anyone touching a line of code.
Now compare that to a workflow that actually leans on cache reads, since that's where Opus 5.5 pulls ahead of a flat 20%. Say a support-escalation agent checks a 20,000-token policy manual at each of four steps per ticket — read policy, draft a response, check the edge case, finalize — and handles 30 tickets a day. The first read of the manual each ticket is a cache write; the next three are cache reads. Add roughly 400 output tokens per step.
On Opus 5: 600,000 daily write tokens at $6.25/MTok ($3.75), 1.8 million read tokens at $0.50/MTok ($0.90), and 48,000 output tokens at $25/MTok ($1.20). That's $5.85 a day, about $176 a month.
On Opus 5.5, same volume: writes at $5/MTok ($3.00), reads at $0.20/MTok ($0.36), output at $20/MTok ($0.96). That's $4.32 a day, about $130 a month — a 26% drop, noticeably more than the contract agent's flat 20%, because three-quarters of this workflow's token volume sits on the line item that got cut 60%.
Did capability drop to hit this price?
Usually "cheaper model" means "worse model" too. That's not what happened here. Anthropic's own benchmark numbers went up, not down: 66.4% on Terminal-Bench 4.0, its agentic coding benchmark, against Opus 5's 52.3%. 81.8% on OSWorld 2.0, a computer-use benchmark, against 74.0%. Output generates roughly 30% faster as well.
If you moved a workflow off Opus for cost or speed reasons in the last two months, that's the one worth re-testing first.
What do you actually have to do about it?
Nothing structural. Opus 5.5 is a new model ID — claude-opus-5-5 — on the Claude API, AWS, Google Cloud, and Azure, so switching is a one-line change wherever your code calls the model, not a rebuild. One thing worth checking: API accounts created after August 31 can no longer turn thinking mode off on Opus 5.5, a "preserved thinking" safeguard Anthropic added at the same time. Test that against any integration that assumed reasoning could be disabled.
The businesses who get the most out of this release aren't the ones already running Opus at scale. They're the ones who priced out an Opus-tier build eight or nine months ago, decided it didn't clear the bar, and shelved it. That bar just moved 20 to 40%. If you'd rather have someone run the updated numbers against your specific workflow than guess, that's what an AI automation scoping call is for.
— Cole
Sources
- Anthropic — Claude Opus 5.5 announcement
- Anthropic — Claude Platform pricing docs (model pricing table, cache pricing, fast mode)