Cheaper per token, and better on Anthropic’s own tests. Claude Opus 5.5 was released on 22 September 2026, two months after Opus 5, and Anthropic sums it up in two claims: it “performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” The price list confirms a cut of a fifth per token. The rest of that 40% is a claim about using fewer tokens per task, which only your own workload can confirm.Anthropic, Introducing Claude Opus 5.5, 22 Sep 2026, read at source 23 Sep 2026: “We’re introducing Claude Opus 5.5, the first model in our new Claude 5.5 family. It performs at the level of Claude Fable 5.1 on most work and costs 40% less to run than Opus 5.” Anthropic, Claude Opus 5, read at source 23 Sep 2026: “Released July 24, 2026.”
- The list price fell by a fifth, and cache reads fell further. The context window and output limit are unchanged.
- Anthropic’s 40% saving is per task, not per token, and it compares the two models at default settings that are not the same.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads.”
- Anthropic now tells developers to start with Opus 5.5. Opus 5 is still available as a legacy model, with no retirement notice.
- Every benchmark is vendor-reported. Opus 5.5 is ahead of Opus 5 on all of them, and Anthropic itself says its lead over Fable 5.1 looks bigger on paper than in use.
- Switching is more than a name change. Thinking can no longer be turned off, and forced tool use now returns an error.
- Subscribers get higher five-hour usage limits on Pro, Max, Team and seat-based Enterprise. Anthropic did not say by how much.
Every price here was correct when this site last checked it, on 23 Sep 2026. Prices change often and this site no longer updates them, so check the vendor’s own page before you rely on one: Anthropic pricing.
What does it cost compared with Opus 5?
Less on every line of the price list. Input is $4 per million tokens against $5, and output $20 against $25, which Anthropic describes as “20% less than Opus 5”. Cache reads, the price of re-sending text the model has already seen, drop from $0.50 to $0.20.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Input and output tokens are $4 and $20 per million, 20% less than Opus 5. Cache reads (which make up the majority of agentic and coding work costs) are $0.20 per million tokens, 60% less than Opus 5.”
| Opus 5.5 | Opus 5 | Fable 5.1 | |
|---|---|---|---|
| Input | $4 | $5 | $10 |
| Output | $20 | $25 | $50 |
| Cache read | $0.20 | $0.50 | $0.25 |
| Cache write, 5 minutes | $5 | $6.25 | $12.50 |
| Batch, input / output | $2 / $10 | $2.50 / $12.50 | $5 / $25 |
| Fast mode, input / output | $8 / $40 | $10 / $50 | not listed |
Anthropic, Migrating to Claude Opus 5.5, read at source 23 Sep 2026: “Claude Opus 5.5 costs less than Claude Opus 5 ($4 / $20 USD per million input / output tokens, compared with $5 / $25”. Fable 5.1 from the Models overview, same day: “$10 / input MTok, $50 / output MTok”. Cache, batch and fast-mode rows from Anthropic’s Pricing tables, same day. A snapshot: re-check before budgeting.
The cache discount is the unusual part. Most Claude models charge a tenth of the input price for a cache hit; Opus 5.5 charges a twentieth. That puts its cache read below Fable 5.1’s in dollars, even though Fable’s multiplier is lower.Anthropic, Pricing, read at source 23 Sep 2026: “On Claude Fable 5.1 and Claude Mythos 5.1, a cache hit costs 2.5% of the standard input price ($0.25 USD per million tokens). On Claude Opus 5.5, a cache hit costs 5% of the standard input price ($0.20 USD per million tokens).” How caching pays off is on prompt caching.
Fast mode, a quicker and dearer way to run the same model, costs $8 in and $40 out, against $10 and $50 for Opus 5. Anthropic’s price list sells it on the Claude API only, not on the cloud platforms.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Fast mode for Opus 5.5 is also available in Claude Code and the Claude Platform with up to 2.5x speed. It costs $8 per million input tokens and $40 per million output tokens.” Anthropic, Pricing, same day: “Fast mode is available on the Claude API (first-party) only”.
For a workload that re-reads a large, stable prompt, the cache price matters more than the headline rate, and that is where Opus 5.5 moved most.
Where the 40% saving comes from
Only half of it is the price list. Anthropic says the model “costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs.” Its test was run “at default settings”.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Our tests show that at default settings it will cost 40% less than Opus 5 on typical workloads.” and “It costs less per token than Opus 5 and uses fewer tokens per task, which nets out to a 40% drop in costs.”
The defaults are not the same. Effort, the setting that decides how much the model thinks before it answers, defaults to medium on Opus 5.5 and to high on Opus 5.Anthropic, Migrating to Claude Opus 5.5, read at source 23 Sep 2026: “Effort is the only thinking control on Claude Opus 5.5, and its default is medium where Claude Opus 5’s is high, so a request that omits effort now runs at medium.” Some of the saving at default settings will come from that lower default, not from the model itself. Anthropic’s own launch post reports that the model beats Opus 5 at a lower effort on several of its coding tests, so the lower default may be justified.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026, on Terminal-Bench 4.0: “Opus 5.5 at default effort beats Opus 5 at max effort for about a fifth of the cost.” But a team that pins effort at high on both models should not expect the same 40%.Reasoning, September 2026 — the per-token cut is 20%, so the rest of a 40% saving has to come from fewer tokens per task, and a lower default effort produces fewer tokens by design. Anthropic also says Opus 5.5 generates output more than 30% faster than Opus 5.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Opus 5.5 also generates output more than 30% faster than Opus 5.”
The per-token cut is certain; the per-task saving is a measurement you have to repeat on your own work, at the effort level you actually use.
Is Opus 5 going away?
Not yet. Anthropic’s documentation now says to start with Opus 5.5, where a week earlier it named Opus 5.Anthropic, Models overview, read at source 23 Sep 2026: “If you’re unsure which model to use, start with Claude Opus 5.5 for most workloads.” The same sentence named Claude Opus 5 when this site read it on 16 Sep 2026 (see which Claude model to use). Opus 5 has moved to a legacy list. Its page says it “is still available” and suggests moving on, and the deprecation table still lists it as Active, not deprecated, with a retirement floor of 24 July 2027.Anthropic, Claude Opus 5, read at source 23 Sep 2026: “Although Claude Opus 5 is still available, you should consider migrating to Claude Opus 5.5 for improved performance.” Anthropic, Model deprecations, same day: claude-opus-5 “Active”, deprecated “N/A”, “Not sooner than July 24, 2027”; and “Legacy: The model will no longer receive updates and may be deprecated in the future.” Anthropic promises at least 60 days’ notice before a retirement.Anthropic, Model deprecations, read at source 23 Sep 2026: “providing at least 60 days’ notice before model retirement for publicly released models.” How that works is on when AI models retire.
What stays the same: a 1M-token context window and 128k maximum output. The knowledge cutoff moves from May 2026 to June 2026.Anthropic, Migrating to Claude Opus 5.5, read at source 23 Sep 2026: “keeps Claude Opus 5’s 1M token context window and 128k max output tokens.” Reliable knowledge cutoff “Jun 2026” on the Claude Opus 5.5 page and “May 2026” on the Claude Opus 5 page, both read 23 Sep 2026.
The launch post calls Opus 5.5 “the new leading model”. The documentation is more careful: it still sends the hardest work to Fable 5.1, which costs two and a half times as much per token.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “It’s the new leading model, and early testers saw large jumps in performance on their most complex work.” Anthropic, Models overview, same day: “Use Claude Fable 5.1 for demanding reasoning and long-horizon agentic work, or when your evals on Claude Opus 5.5 at higher effort still fall short.”
Nothing forces a move today. Code pinned to Opus 5 keeps working, but it is now on a model that gets no further updates.
What do the benchmarks say?
| Benchmark (what it tests) | Opus 5.5 | Opus 5 | Fable 5.1 |
|---|---|---|---|
| Terminal-Bench 4.0 (agentic coding) | 66.4% | 52.3% | 55.8% |
| FrontierCode v1.1 (agentic coding) | 54.4% | 48.0% | 50.3% |
| CursorBench 4.0 (agentic coding) | 57.8% | 46.6% | 51.8% |
| GDPval-AA v2.1 (knowledge work, Elo) | 1846 | 1708 | 1735 |
| AutomationBench (business workflows) | 40.0% | 26.9% | 31.4% |
| Humanity’s Last Exam, with tools (reasoning) | 67.7% | 63.6% | 65.6% |
| Terminal-Bench-Science 0.1 (scientific research) | 58.7% | 29.0% | 52.6% |
| OSWorld 2.0 (computer use, partial) | 81.8% | 74.0% | 80.7% |
Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026. Anthropic’s table also lists GPT-6 Astra and GPT-5.6 Sol, left out here. “Unless otherwise noted, all Claude Opus 5.5 results use adaptive thinking at max effort.” The page reports Terminal-Bench 4.0 for Opus 5.5 “at xhigh effort”.
On Anthropic’s numbers, Opus 5.5 is ahead of Opus 5 on every row, most sharply on the science test, where the score roughly doubles. It is also ahead of the dearer Fable 5.1 on every row.
Three things to keep in mind. First, Anthropic itself cautions that the margins overstate the difference from Fable 5.1: “In our own use, the gap between Opus 5.5 and Claude Fable 5.1 is narrower than these scores suggest.”Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “That said, at these levels of capability we’ve found that benchmark margins have become a less reliable guide to real-world differences.” Second, when the new model’s safeguards stepped in during testing, older models finished the task, and Anthropic says that probably lowered the scores.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “When they intervened, cybersecurity tasks were completed by Claude Opus 4.8, and biology and frontier LLM development tasks were completed by Claude Opus 5. This likely reduces Claude Opus 5.5’s performance on these benchmarks.” Third, some rows were run by others but are still reported by Anthropic: AutomationBench by Zapier, and Terminal-Bench’s Opus 5 figure sits close to the public leaderboard’s.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “AutomationBench results were run and reported by Zapier.” and “The public leaderboard (5 trials/task, Claude Code harness) reports Claude Opus 5 at 51.8%; our setup reproduces it at 52.3%, within noise.”
As on our Grok 4.7 vs 4.6 page: compare models inside one vendor’s table from one date, never across two.
The direction is clear and consistent. The size of the lead is Anthropic’s measurement, at settings Anthropic chose, and it says so itself.
What changes if you switch
On the API the model name becomes claude-opus-5-5, and Anthropic lists four breaking changes for code already running on Opus 5:Anthropic, What’s new in Claude Opus 5.5, read at source 23 Sep 2026: “Four breaking changes affect code already running on Claude Opus 5”; “On Claude Opus 5.5, thinking is always on”; “Claude Opus 5.5 doesn’t support forced tool use.”
- Thinking cannot be turned off. A request that disables it returns an error; lower the effort setting instead.
- Forced tool use is gone. Asking the model to always call a tool, or a named tool, returns an error.
- Thinking blocks are tied to the model and the conversation, which matters if a router moves a conversation between models.
- An older computer-use tool is not accepted on the Claude API and Google Cloud.
Two more changes do not fail a request but alter behaviour: the effort default drops to medium, and text the model writes between tool calls now comes back in thinking blocks, so an app that shows it as progress can go quiet. Anthropic’s migration guide covers both.Anthropic, Migrating to Claude Opus 5.5, read at source 23 Sep 2026: “No request fails, but an application that streams that text to its users as progress updates goes quiet between tool calls.”
Safeguards also change. Anthropic says most cybersecurity tasks will be passed to an older model, and that Opus 5.5 carries the same text watermark as Fable 5.1.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “most cybersecurity tasks will be re-routed to Opus 4.8” and “As with Fable 5.1, Opus 5.5 comes with our watermarking measures to comply with the EU AI Act”.
Budget a test run, not just a find-and-replace. Code that disabled thinking or forced a tool call will break on the first request.
For people on a Claude plan
Anthropic says it is raising the five-hour usage limits on Pro, Max, Team and seat-based Enterprise plans, and giving subscribers a reset they can save for later. It did not publish the new limits.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “In addition to the price drop, we’re increasing five-hour usage limits on Pro, Max, Team, and seat-based Enterprise plans. We’re also providing subscription users a rate limit reset, which you can now save and use whenever you choose.”
The model is on Anthropic’s own apps and API and on Amazon Web Services, Google Cloud and Microsoft Azure. Smaller models in the same family are promised but not dated.Anthropic, Introducing Claude Opus 5.5, read at source 23 Sep 2026: “Claude Opus 5.5 is now available on all platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure.” and “Claude Sonnet 5.5 and Claude Haiku 5.5 will follow in the coming weeks, with many of the same improvements to performance, efficiency, and safety.” Until they arrive, Anthropic’s four current models carry four different version numbers; what AI model names mean explains why that is normal for Anthropic.
If you hit limits on a plan, this release should give you more room, but by an amount you will only learn from use.
Should you switch?
On the API, the price cut makes it cheap to try, and Anthropic’s own advice now points there. Run the prompts you rely on through both models at the same effort level, then again at Opus 5.5’s default, and compare cost and quality. That tells you how much of Anthropic’s claimed saving you actually get. How Opus 5.5 sits against other vendors is on LLM API prices compared.
Lower prices make this an easy trial. Your own prompts, at your own effort level, are the only benchmark that answers your question.
What this page could not verify
- Every benchmark number. All come from Anthropic’s announcement, including the rows it says others ran; none was re-run here.
- The per-task saving. It depends on the workload and the effort setting; Anthropic has not published the workloads.
- How far the five-hour limits rose. Not stated.
- Whether Opus 5.5 is the default in the Claude apps. The pages read here cover the API recommendation, not the app default.
- When Opus 5 will be retired, and when Sonnet 5.5 and Haiku 5.5 arrive. No dates were published as of 23 September 2026.
Anthropic, Introducing Claude Opus 5.5, 22 September 2026. Anthropic developer docs: Models overview, Pricing, Model deprecations, Claude Opus 5.5, Claude Opus 5, Migrating to Claude Opus 5.5 and What’s new in Claude Opus 5.5. All read at source on 23 September 2026.
Prices are a snapshot and vendor benchmarks are marketing until someone else runs them.