Last updated: September 2026
Key Takeaways
- On the API, yes: Claude Opus 5.5 costs 20% less per token than Opus 5 and cache reads cost 60% less. Anthropic's own tests put typical workloads about 40% cheaper at default settings, because the model also uses fewer tokens per task.
- On Pro, Max, and Team, the bill did not change. Five-hour usage limits went up, and subscribers get a rate-limit reset they can save and use when they choose. For a subscriber, "cheaper" means more work per window, not a lower invoice.
- The catches come from Anthropic's own page: most cybersecurity requests are answered by Opus 4.8 and biology requests by Opus 5, thinking cannot be switched off, API accounts opened since August 31 cannot edit prior context, and output carries Anthropic's text watermark.
Yes, if you pay per token. Claude Opus 5.5, released September 22, 2026, costs 20% less per token than Opus 5 and, by Anthropic's own estimate, about 40% less on typical workloads. If you pay a subscription, your bill did not change; your limits did. On either path, the fine print decides which model answers, and that is the part this article covers that the launch summaries skip.
Every figure below comes from Anthropic's announcement, its Help Center, or a named independent source, checked on September 22, 2026.
Opus 5.5 Pricing: What Changed on the API
The per-token cut is 20%, and the workload cut is an estimate. Per Anthropic's September 22 announcement, Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, against $5 and $25 for Opus 5. Cache reads drop from $0.50 to $0.20 per million, a 60% cut, and cache writes from $6.25 to $5. A fast mode in Claude Code and the Claude Platform runs at up to 2.5x speed for $8 and $40. Current rates live on Anthropic's pricing page.
The 40% figure is a different kind of number. Anthropic says its tests show Opus 5.5 "at default settings" costs 40% less than Opus 5 "on typical workloads," because it needs fewer tokens and fewer steps to finish a task and generates output more than 30% faster. That is a vendor estimate at medium effort, not a price. At maximum effort the picture changes: Artificial Analysis measured Opus 5.5 using about 1.6 times the output tokens of Opus 5 per task, which left cost per task level with the older model despite the lower rate.
The cache-read cut is the one that moves real bills. Anthropic states that cache reads make up the majority of agentic and coding costs, because a coding agent re-reads the same repository context and instructions on every turn. A 60% cut on the line item that dominates the invoice matters more than a 20% cut on the headline rate.
What Changed on Pro, Max, and Team
Subscription prices did not change; capacity did. Anthropic's announcement says it is "increasing five-hour usage limits on Pro, Max, and Team plans" and giving subscribers "a rate limit reset, which you can now save and use whenever you choose." It does not put a percentage on the increase. The New Stack reports the five-hour increase at 20% and quotes Anthropic saying the model's lower token use makes five-hour and weekly limits go 25% further; AlphaSignal describes the reset as one-time. Treat those two numbers as reported until the Help Center article states them.
| Where You Pay | What Changed | What Did Not |
|---|---|---|
| Claude API | Token rates down 20%, cache reads down 60%, cache writes down, fast mode added at a higher rate | 1M context; zero data retention still offered |
| Pro | Higher five-hour limit; a saveable rate-limit reset | Monthly price |
| Max | Higher five-hour limit; a saveable rate-limit reset; Opus 5.5 available | Monthly price; the weekly-limit structure |
| Team | Higher five-hour limit; a saveable rate-limit reset | Seat price |
| Claude Code | Opus 5.5 available; fast mode available at a higher rate | Usage shared with the Claude apps under the same limits |
Limit changes per Anthropic's September 22 announcement; the size of the increase and the reset mechanics are reported, not yet documented in the Help Center as of publication.
The practical reading for subscribers: if Opus 5.5 finishes a task in fewer tokens and the window got larger, the same plan does more work per five hours. That is a real gain. It is not the same as a cheaper plan, and it depends on the reroute rules in the next two sections, since a request answered by an older model still spends your window.
Is Opus 5.5 Better Than Opus 5 and Fable 5.1?
On Anthropic's own table, yes, with a caveat Anthropic wrote itself. Opus 5.5 scores 66.4% on Terminal-Bench 4.0 against 55.8% for Fable 5.1 and 52.3% for Opus 5; 54.4% on FrontierCode against 50.3% and 48.0%; and 1846 Elo on GDPval-AA v2.1 against 1735 and 1708. Then the announcement adds that "benchmark margins have become a less reliable guide to real-world differences" and that in Anthropic's own use the gap to Fable 5.1 "is narrower than these scores suggest." Artificial Analysis, which runs its own index, puts Opus 5.5 at 58 at maximum effort, the highest score it has recorded.
Two footnotes in the table matter for anyone reading the numbers as a product spec. Anthropic ran its evaluations with production safeguards enabled; when they intervened, cybersecurity tasks were completed by Opus 4.8 and biology and frontier-model tasks by Opus 5. And the AutomationBench figure was run without fallback models, so every safeguard intervention counted as a failure. The scores describe the deployed system, routing included, not an isolated model answering every prompt. That is the right way to test a hosted product, and it is also the reason the next section exists.
The Catch: Which Model Actually Answers
You choose Opus 5.5; for some requests Anthropic's classifiers choose a different model. The announcement says the new safeguards "all fall back to another model transparently," that "most cybersecurity tasks will be re-routed to Opus 4.8," and that biology and frontier-LLM-development requests fall back to Opus 5. Routine software work, including finding and fixing bugs, stays on Opus 5.5. The sanctioned way around the reroutes is verification: a Cyber Verification Program that Anthropic says will expand to Opus 5.5 with three tiers of trusted access, and a Life Sciences Verification Program for vetted organizations.
| Request Type | Model That Answers | Route to Full Access |
|---|---|---|
| Routine coding, including bug finding and fixing | Opus 5.5 | None needed |
| Most cybersecurity tasks | Opus 4.8 | Cyber Verification Program (expansion to Opus 5.5 announced, not yet open) |
| Biology research and development | Opus 5 | Life Sciences Verification Program (applications open) |
| Frontier LLM development | Opus 5 | Not stated |
| Everything else | Opus 5.5 | None needed |
Routing per Anthropic's September 22 announcement. Whether a rerouted request is billed at the Opus 5.5 rate or counted against a subscription window at the Opus 5.5 rate is not stated on the launch page.
The precedent is three months old. When Fable 5 returned in July with its cyber classifier, blocked requests notified the user, were answered by Opus 4.8, and, by early user reports, still consumed plan usage; our Fable 5 analysis tracked the coding false positives Anthropic said it was tuning out. For Opus 5.5, "transparently" is Anthropic's word, and the announcement does not say what the user sees when a reroute happens or what it costs. As our Opus 5 versus Fable 5 guide put it, no hosted model can promise it will not hand your request to a different model mid-task. The consumer version is shorter: you pay for Opus 5.5, and on a defined set of requests you receive an older model, on a decision made on Anthropic's side.
Four Other Terms That Changed
None of these is hidden; all of them sit below the benchmark table.
- Thinking cannot be turned off. Opus 5.5 is no longer available with thinking disabled. Effort levels are the cost control, and every response carries reasoning tokens you pay for on the API and spend from your window on a subscription.
- Preserved thinking for newer API accounts. The anti-distillation safeguard introduced with Fable 5.1 applies to Opus 5.5 for API accounts created on or after August 31, 2026. It stops the caller from editing Claude's prior context, which is a real change for integrations that rewrite conversation history; Anthropic's Help Center article explains the mechanics.
- Watermarked output. Opus 5.5 ships with the text watermarking Anthropic introduced for EU AI Act compliance. Zero data retention remains available, as with prior Opus models.
- The family is not finished. Sonnet 5.5 and Haiku 5.5 follow "in the coming weeks" with many of the same efficiency changes. If your workload runs on Sonnet today, the price question for you is still a few weeks out; our Sonnet 5 breakdown is the baseline the next release will move.
What the Price Cut Does Not Change for Local AI
A rented model getting cheaper does not change what a local model costs, and the gap between them is smaller than the price gap suggests. On Artificial Analysis's current index, Opus 5.5 scores 58 at maximum effort and Qwen3.8-27B, an Apache 2.0 model that fits a 24 GB card at 4-bit, scores 52. Six points separates the most capable hosted model from the best open model most people can run at home. The 27B is slow and verbose to get there, and a 4-bit local build sits somewhere below its hosted score, but the number is what it is.
The electricity side is small. A 24 GB RTX 3090 is rated at 350 watts of board power; run flat out for eight hours a day, it draws 2.8 kilowatt-hours, which at the EIA's 2026 average residential rate of 18.2 cents per kilowatt-hour in its Short-Term Energy Outlook costs about 51 cents a day, or roughly $15 a month. At Opus 5.5 rates, $15 buys 750,000 output tokens or 3.75 million input tokens. The card is the real cost, and it is a one-time cost that does not scale with tokens; our local AI hardware guide covers what that card should be and our Qwen3.8-27B guide has the speeds it delivers.
Which job goes where has not moved. Hosted wins multi-hour agentic coding, computer use, and 1M-token context by default, and Opus 5.5 made all three cheaper. Local wins the jobs where the terms are the product: an always-on home agent with no five-hour window, documents that never leave the machine, no classifier deciding which model answers, no watermark, and no terms revision on someone else's schedule. Our models-by-VRAM guide has the right pick for every memory tier. A price cut on a rented model is not a reason to buy a card, and it is not a reason to skip one.
The same terms question sits at the edge of your network. A rented ISP gateway is a monthly fee for a box you never own, running remote-management firmware you do not control. Replacing it with a modem and router you own ends both; our guide to owning your modem covers what changes.
Who Should Switch Tonight
If you run coding agents on the API, switch: the cache-read cut lands on the line that dominates your invoice, and the model finishes in fewer steps. If you subscribe, Opus 5.5 is available now; use the saved reset for the one long session that used to hit the wall, and expect any security-flavored request to come back from Opus 4.8. If your work is cybersecurity or biology, the model you want is behind a verification program, and the launch page says so. If you run local, nothing here changes your hardware or your model pick; it changes the price of the alternative you already declined.
Frequently Asked Questions
Is Claude Opus 5.5 cheaper than Opus 5?
Yes, on the API. Opus 5.5 costs $4 per million input tokens and $20 per million output tokens, 20% below Opus 5's $5 and $25, and cache reads fell from $0.50 to $0.20 per million, per Anthropic's September 22, 2026 announcement. Anthropic estimates typical workloads cost about 40% less at default settings because the model uses fewer tokens per task.
Did Claude Pro or Max get cheaper with Opus 5.5?
No. Subscription prices did not change. Anthropic increased five-hour usage limits on Pro, Max, and Team and gave subscribers a rate-limit reset that can be saved and used later. Because Opus 5.5 uses fewer tokens per task, the same limit covers more work, which is the form "cheaper" takes on a subscription.
What is the Opus 5.5 rate-limit reset?
A reset of your usage limit that Anthropic is giving subscription users with the Opus 5.5 launch, which you can save and apply whenever you choose instead of waiting for the window to roll over. Anthropic's announcement describes it in one sentence; reports describe it as one-time. The Help Center should carry the mechanics.
Is Opus 5.5 better than Fable 5.1?
On Anthropic's benchmarks it leads Fable 5.1 on agentic coding, knowledge work, and computer use, and Artificial Analysis scores it 58 at maximum effort, its highest recorded result. Anthropic adds that in its own use the gap to Fable 5.1 is narrower than the scores suggest. Opus 5.5 is also cheaper per token than Fable 5.1's published rates.
Which requests does Opus 5.5 hand to an older model?
Per Anthropic, most cybersecurity tasks are re-routed to Opus 4.8, and biology and frontier-LLM-development tasks fall back to Opus 5. Routine coding, including bug fixes, stays on Opus 5.5. Verified organizations can apply to the Cyber Verification Program and the Life Sciences Verification Program for fuller access.
Can you turn off thinking in Opus 5.5?
No. Opus 5.5 is no longer available with thinking disabled, per Anthropic's model documentation. The controls are the effort levels, and Anthropic's own charts show default effort delivering most of the score at a fraction of the cost; every response includes reasoning tokens.
Can you run Claude Opus 5.5 locally?
No. Opus 5.5 is a closed, hosted model with no downloadable weights, and any file offered under that name is not it. The closest local alternative by independent score is Qwen3.8-27B at 52 on the Artificial Analysis index against 58 for Opus 5.5, running on a single 24 GB GPU under Apache 2.0.
