How much does Claude Opus 5.5 cost?

Input is $4 per million tokens and output $20 per million, each 20% below Opus 5 (which was $5/$25). Cache reads fell 60%, from $0.50 to $0.20 per million — the line item that dominates agentic and coding workloads. Fast mode, which runs up to 2.5x standard output speed, is priced at $8 input and $40 output. Anthropic claims typical workload cost drops about 40% versus Opus 5, but only part of that comes from the 20% price cut; the rest comes from needing fewer steps and fewer tokens to finish the same task.

Where do the 40% cost savings actually come from?

Two separate sources, and conflating them is the most common mistake when reading this launch. The first is the list price cut of 20% on input and output, plus a 60% cut on cache reads. The second is token efficiency — the model completing the same work in fewer steps and fewer tokens. In one Anthropic example, auditing and fixing a 200,000-line codebase took under three hours, where Opus 5 took over 20 hours and consumed 2.5x the tokens. A 20% discount cannot produce a 40% saving on its own; the remainder is genuine efficiency, which is the more interesting part of the release.

What are the notable limitations or behaviour changes in Opus 5.5?

Several are worth flagging. Thinking can no longer be disabled, which removes a latency and cost lever that users previously had. Preserved thinking blocks prevent API users from editing prior thinking context on any change to the conversation for accounts created on or after 2026-08-31, and Anthropic has added a statistical watermark to output. There is no separate Fast Mode discount — fast mode costs double. Anthropic also discloses that Opus 5.5 frequently suspects it is being tested, which raises questions about evaluation validity, and that there is no reliable pre-release method for finding every problem. Finally, Anthropic concedes that leaderboard gaps understate how close the model really is to Fable 5.1.

Claude Opus 5.5 Review 2026: 20% Cheaper Per Token, 40% Cheaper Per Task

Q: What is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's new flagship, released September 22, 2026, and the first model in the 5.5 line. Anthropic says it matches the more expensive Fable 5.1 on most tasks while running 30%+ faster than Opus 5. It tops the Artificial Analysis Intelligence Index at 58 points on max effort — the highest score that evaluator has recorded — putting it several points above Claude Fable 5.1 and GPT-6 Astra (both 53) and well above its predecessor Opus 5 (51). It is available through the Claude API alongside AWS, Google Cloud and Microsoft Azure, and to Pro, Max, Team and Enterprise subscribers.

Q: Is Claude Opus 5.5 better than Claude Fable 5.1?

By the headline numbers, yes: Opus 5.5 scores 58 on the Artificial Analysis Intelligence Index against Fable 5.1's 53, and it is priced lower at $4/$20 versus Fable 5.1's $10/$50. But Anthropic itself concedes that benchmark gaps poorly reflect real-world experience and that the internal difference between Opus 5.5 and Fable 5.1 is smaller than the leaderboards suggest. Fable 5.1 also retains the higher ceiling for the most demanding work. The practical read is that Opus 5.5 is the better default, while Fable 5.1 remains the specialist choice.

Claude Opus 5.5 Review 2026: 20% Cheaper Per Token, 40% Cheaper Per Task — and #1 on AA's Index at 58

September 23, 2026 · Updated 2026 · AI Coding · Agentic AI · LLM Pricing · Read time: ~11 min

Verdict first: Claude Opus 5.5 (released September 22, 2026) is the rare model launch where the interesting number is not the benchmark but the arithmetic. It tops the Artificial Analysis Intelligence Index at 58 on max effort — the highest that evaluator has recorded — placing it several points above Claude Fable 5.1 and GPT-6 Astra (both 53) and well clear of Opus 5 (51). It runs 30%+ faster than Opus 5 and, per Anthropic, matches the far pricier Fable 5.1 on most tasks. The pricing is where it gets genuinely worth understanding: input $4 / output $20 per million tokens, each 20% below Opus 5, with cache reads cut 60% from $0.50 to $0.20 — yet Anthropic claims typical task cost drops 40%. A 20% discount cannot produce a 40% saving by itself. The other half comes from token efficiency — fewer steps, fewer tokens per completed task — and that is the real story of this release. The trade-offs are real too: thinking can no longer be disabled, preserved thinking blocks stop API users editing prior reasoning on new accounts, output now carries a statistical watermark, and Anthropic itself concedes the leaderboard gap understates how close Fable 5.1 is.

Claude Opus 5.5 review 2026 — 20% cheaper per token, 40% cheaper per task, and number one on the Artificial Analysis Intelligence Index at 58

Claude Opus 5.5 — $4/$20 per million tokens with cache reads at $0.20, #1 on the AA Intelligence Index at 58, and 30%+ faster than Opus 5

What Is Claude Opus 5.5?

Claude Opus 5.5 is Anthropic's new flagship, released September 22, 2026, and the first model in the 5.5 line. The positioning is narrower than "the biggest model" — it is "Fable 5.1 capability at Opus prices." Anthropic says it matches the more expensive Fable 5.1 on most tasks while running over 30% faster than Opus 5.

Two things make it worth a full review rather than a changelog entry:

It is now the top of the Artificial Analysis Intelligence Index at 58 — a score that, as we will see, comes with an important caveat about whose harness produced it.
The cost story is compound, not simple. The headline 40% saving is not a 40% price cut, and understanding where the rest comes from tells you more about where the industry is heading than the benchmark table does.

It is available through the Claude API alongside AWS, Google Cloud and Microsoft Azure, and to Pro, Max, Team and Enterprise subscribers. Anthropic says Sonnet 5.5 and Haiku 5.5 arrive in the coming weeks.

The Pricing Math: Why 20% Becomes 40%

Item	Opus 5.5	Opus 5	Change
Input (per 1M)	$4	$5	−20%
Output (per 1M)	$20	$25	−20%
Cache read (per 1M)	$0.20	$0.50	−60%
Fast mode (up to 2.5× speed)	$8 in / $40 out	—	—

The cache-read cut is the one that matters most in practice. For agentic and coding workloads, cache reads are the dominant cost line — an agent re-reading the same large context every turn pays the cached-input rate far more often than the miss rate. Cutting it 60% changes the economics of long loops more than the headline 20% does.

But here is the part worth slowing down for. Anthropic claims typical workload cost falls about 40%, and the 20% price cut cannot account for that. The remainder is token efficiency — the model finishing the same work in fewer steps and fewer tokens. Anthropic's own example: auditing and fixing a 200,000-line codebase took under three hours, where Opus 5 took over 20 hours and consumed 2.5× the tokens.

Why this distinction matters for your budget: a pure price cut is predictable — you multiply your old spend by 0.8. An efficiency gain is workload-dependent — it shows up on long agentic tasks and may show up much less on short, single-shot calls where there were never extra steps to remove. Model your own task mix before assuming the 40% figure applies to you.

Benchmarks: #1 at 58

Artificial Analysis Intelligence Index:

Model	Score
Claude Opus 5.5 (max)	58 🥇
Claude Opus 5.5 (xhigh)	56
Claude Opus 5.5 (high)	54
Claude Fable 5.1	53
GPT-6 Astra	53
Claude Opus 5	51
Xiaomi MiMo-V2.6-Pro (best open)	46

Artificial Analysis describes 58 as the highest score it has recorded, several points above the previous best. Two things stand out beyond the top line:

Four of Opus 5.5's five effort settings (max, xhigh, high, medium) sit on the intelligence-versus-cost Pareto frontier — meaning there is no "wrong" effort level to pick, only different points on a good curve. That is an unusual and genuinely useful property.
The gap between the best closed and best open model is now 12 points (58 versus MiMo-V2.6-Pro's 46). Open weights have been closing fast this year; this release widens the gap again.

Where Opus 5.5 leads: it comes first in six of the ten evaluations, including Humanity's Last Exam (61.4%), SciCode (66.9%), GDPval-AA, AA-Briefcase, AA-Omniscience and AutomationBench-AA.

Other headline scores (Anthropic-reported):

Benchmark	Opus 5.5	Comparison
Terminal-Bench 4.0	66.4%	Opus 5 52.3% · GPT-6 Astra 57.9% · Fable 5.1 55.8%
FrontierCode v1.1	54.4%	GPT-6 Astra 53.3%
CursorBench 4.0 (max)	57.8%	Fable 5.1 51.8% · GPT-5.6 Sol 41.7%
GDPval-AA v2.1 (44 occupations)	1846 Elo	Fable 5.1 1735 · Opus 5 1708

Two details in that table are more interesting than the headline wins. First, FrontierCode at default (medium) effort scored 54.6%, above its own max-effort 54.4% — a 0.2-point gap that the source itself calls within noise. Second, on GDPval-AA the default effort setting beats GPT-6 Astra at max effort, at roughly one-fifth the per-task cost, and on CursorBench the default medium setting still leads GPT-5.6 Sol's best by about 11 points at around a third of the cost.

The pattern: the wins are largest on real multi-step work — terminal tasks, multi-file Cursor sessions, 44-occupation knowledge work — and thinnest on synthetic single-shot benchmarks. That is consistent with the token-efficiency story: this model was tuned for finishing jobs, not for answering questions.

Vendor Numbers vs AA's Own Harness

This is the detail most coverage has skipped, and it is worth stating plainly.

Anthropic reports Terminal-Bench 4.0 at 66.4% — a substantial margin over GPT-6 Astra's 57.9%. But on Artificial Analysis's own harness, Opus 5.5 and GPT-6 Astra come out level on the same benchmark.

Both numbers can be true. Vendor benchmark runs and third-party harness runs differ in setup, tooling, retry policy and scoring, and the gap between them is exactly why independent evaluation matters. But when a 8.5-point lead narrows to parity depending on who runs the test, the honest reading is that the real-world difference is smaller than the vendor table implies.

Notably, Anthropic says this itself: the company concedes that benchmark gaps poorly reflect real-world experience and that the internal difference between Opus 5.5 and Fable 5.1 is smaller than the leaderboards suggest. That kind of admission is rare in a launch post, and it is the most credible sentence in the release.

Real-World Results: 680K Lines in a Day

The task-level examples are where the efficiency claim gets tested:

680,000 lines of code migrated in under a day — work Anthropic says would take a human team weeks.
A 200,000-line codebase audited and fixed in under three hours, versus Opus 5's 20+ hours and 2.5× the tokens.
HAProxy rewritten from C to Rust in 9.5 hours, against Fable 5.1's 12 hours, at 51% lower cost, with both passing nearly all regression tests.
Page-load optimisation: 39 successes out of 40 attempts.
Walleye Capital completed its quantitative research evaluation at the lowest effort setting; at higher effort the model spotted an off-by-one error in the instructions and flagged that fixing it could cost it points — arguably more useful behaviour than silently succeeding.

Anthropic also notes a style change: the model now leads with conclusions rather than building to them. In a billing-bug example it opens by stating that $9.92 traces to a bug and $1.50 to free-tier adjustments. One tester's summary — "it writes the way I do" — is the kind of feedback that does not show up in an Elo score but does show up in whether people keep using a model.

Safety, Anti-Distillation and the Behaviour Changes

Opus 5.5 is the first Opus to carry Fable 5.1-level cyber, bio and anti-distillation guardrails. Anthropic audited it across nearly 2,000 simulated scenarios and reports boundary-circumvention attempts about 85% lower than Opus 5 and Mythos 5.1. Agent protections include pre-execution checks, an auditable open-source sandbox, and pre-merge code review.

Two access-control notes with real consequences: bio work routes through the Life Sciences Verification Program, and cyber requests are largely routed to Opus 4.8 for general users, with the Cyber Verification Program reserved for certified professionals.

The anti-distillation changes are the ones most likely to affect normal API users:

Preserved thinking blocks — API users can no longer edit prior thinking context on any change to the conversation, for accounts created on or after 2026-08-31.
Thinking can no longer be disabled at all. This removes a latency and cost lever that users previously had — worth knowing if your pipeline relied on switching it off for cheap calls.
A statistical watermark is now embedded in output.

On subscriptions, Anthropic raised the five-hour limits for Pro, Max, Team and per-seat Enterprise plans and added one bankable rate-limit reset that you can apply at a time of your choosing — a small feature that matters a great deal during a deadline.

The Honest Take

Score: 8.8/10. This is a strong release with a genuinely interesting economic structure and a handful of real trade-offs.

What genuinely impresses:

It is #1 on the AA Intelligence Index at 58, several points above the previous ceiling — and that score comes from a third-party evaluator's harness, not a vendor table.
Four of five effort levels sit on the Pareto frontier. You can pick for speed or for cost without stepping off the efficient curve.
The cache-read cut to $0.20 (−60%) targets the line item that actually dominates agent loops, which is a more sophisticated pricing move than an across-the-board discount.
The 40% task-cost claim is partly earned by efficiency, not just discounting. Fewer steps and fewer tokens per task is a durable improvement; a price cut can be reversed next quarter.
The real-work results are concrete — 680K lines migrated, a 200K-line codebase fixed in under three hours, HAProxy C-to-Rust at half the cost of Fable 5.1.
Fable 5.1-level capability at $4/$20 versus Fable 5.1's $10/$50 is a compelling default-tier offer.
Guardrails are measurably stronger — ~85% fewer boundary-circumvention attempts across ~2,000 simulated scenarios.
The bankable rate-limit reset is a genuinely thoughtful subscription feature.
Anthropic's own candour. Conceding that leaderboard gaps understate the real difference to Fable 5.1 costs them a headline and earns them credibility.

What should temper your excitement:

Vendor and third-party Terminal-Bench results disagree. 66.4% under Anthropic's run versus parity with GPT-6 Astra under AA's harness — an 8.5-point gap that evaporates under independent testing.
The 40% saving is workload-dependent. It will be largest on long agentic loops and smallest on short single-shot calls.
Thinking can no longer be disabled. If your pipeline relied on that for cheap or low-latency calls, your cost and latency profile changes.
Preserved thinking blocks restrict editing prior reasoning on accounts created from 2026-08-31 — a real constraint on some agent patterns.
A statistical watermark in output may matter for downstream processing or compliance review.
Fast mode is double the price, not a free toggle.
Anthropic discloses that Opus 5.5 frequently suspects it is being tested — which raises legitimate questions about how much the benchmark numbers reflect normal behaviour.
Anthropic also states there is no reliable pre-release method for finding every problem — and FrontierCode's default effort slightly beating its own max effort suggests the effort ladder is not perfectly monotonic.
Cyber capability is gated for general users, who get routed to Opus 4.8 instead.

Who should use it: teams running long agentic loops and large-codebase work, where the cache-read cut and token efficiency compound into real money; coding-heavy workflows, given Terminal-Bench, CursorBench and the migration case studies; anyone who wanted Fable 5.1 capability but could not justify $10/$50; enterprises already inside AWS, Google Cloud or Azure; and cost-sensitive deployments that benefit from having four of five effort levels on the Pareto frontier.

Who should skip it: pipelines that depend on disabling thinking; agent patterns that edit prior thinking context on newer accounts; use cases where a statistical watermark is a problem; workflows that are short and single-shot, where the 40% saving will not materialise; and anyone needing the absolute ceiling for the hardest research problems, where Fable 5.1 still sets the bar.

FAQ

What is Claude Opus 5.5?

Anthropic's new flagship, released September 22, 2026, and the first model in the 5.5 line. Anthropic says it matches the more expensive Fable 5.1 on most tasks while running 30%+ faster than Opus 5. It tops the Artificial Analysis Intelligence Index at 58 on max effort — the highest recorded — above Fable 5.1 and GPT-6 Astra (both 53) and Opus 5 (51). Available via the Claude API plus AWS, Google Cloud and Microsoft Azure, and to Pro, Max, Team and Enterprise subscribers.

How much does it cost?

$4 per million input tokens and $20 per million output, each 20% below Opus 5. Cache reads dropped 60%, from $0.50 to $0.20 per million — the dominant cost line for agentic and coding work. Fast mode (up to 2.5× output speed) is $8/$40. Anthropic claims roughly 40% lower typical task cost, but only half of that comes from the price cut; the rest is token efficiency.

Where do the 40% savings come from?

Two sources that are easy to conflate. First, the list-price reduction — 20% on input and output, 60% on cache reads. Second, token efficiency: completing the same work in fewer steps and fewer tokens. In Anthropic's example, a 200,000-line codebase was audited and fixed in under three hours where Opus 5 took 20+ hours and 2.5× the tokens. A 20% discount mathematically cannot yield a 40% saving, so the remainder is real efficiency — it just varies with your workload mix.

Is it better than Claude Fable 5.1?

On the headline numbers, yes — 58 versus 53 on the AA index, at $4/$20 versus $10/$50. But Anthropic concedes that benchmark gaps poorly reflect real-world experience, and that the internal difference between the two is smaller than the leaderboards suggest. Fable 5.1 still represents the higher ceiling for the most demanding work. Practically: Opus 5.5 is the better default, Fable 5.1 the specialist option.

What behaviour changes should I know about?

Thinking can no longer be disabled, which removes a latency and cost lever. Preserved thinking blocks prevent API users from editing prior thinking context on any conversation change, for accounts created on or after 2026-08-31. Output now carries a statistical watermark. Fast mode costs double. Cyber requests from general users are largely routed to Opus 4.8, with certified professionals using the Cyber Verification Program.

Does the 66.4% Terminal-Bench score hold up?

With a caveat worth understanding. Anthropic reports 66.4%, well above GPT-6 Astra's 57.9%. But on Artificial Analysis's own harness, the two models come out level on the same benchmark. Both can be true — vendor runs and third-party harnesses differ in tooling, retries and scoring — but when a large lead narrows to parity under independent testing, the real-world gap is smaller than the vendor table suggests. Artificial Analysis does place Opus 5.5 first overall on its Intelligence Index at 58.

What is coming next from Anthropic?

Anthropic says Claude Sonnet 5.5 and Haiku 5.5 will launch in the coming weeks. Opus 5.5 is the first of the 5.5 line.

Sources

Benchmark figures reported by Anthropic are vendor-run; the Artificial Analysis Intelligence Index and its Terminal-Bench parity finding come from third-party evaluation with its own harness. Where the two disagree, both are noted. Prices and specifications change frequently — verify against Anthropic's official documentation before making purchasing decisions.

搜索此博客

ring