Anthropic Claude Opus 5.5 Cuts Cost 40%

Anthropic Claude Opus 5.5 has arrived, and the headline is simple: Anthropic says it matches the performance of its Fable 5.1 model on most work while costing 40% less to run than Opus 5 on typical workloads at default settings. That combination, near-frontier capability at a sharply lower operating cost, is the kind of claim that usually gets walked back within a week. This time the company is backing it with a full system card and a pricing sheet that developers can inspect before they commit a single line of production code.

The model is the first release in the new Claude 5.5 family. It is deployable today as a managed API model, which means no weights are being handed out. If self-hosting matters to your stack, this release does nothing for you. Developers can call claude-opus-5-5 through the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. Zero data retention remains available, consistent with previous Opus models.

Benchmarks: A Strong Lead, Not a Clean Sweep

On Anthropic's own benchmarks, Opus 5.5 leads in agentic coding, computer use, and knowledge work. That lead is real but not absolute. GPT-6 Astra still holds the top spot on Terminal-Bench-Science and AutomationBench. Terminal-Bench 4.0 was reported at xhigh effort for Opus 5.5, while its overall scores use adaptive thinking at max effort with production safeguards enabled.

There is a caveat worth sitting with. Zapier ran AutomationBench without fallback models, so any safeguard intervention counted as a failure. That single methodological choice can swing a leaderboard. Anthropic itself cautions that benchmark margins are becoming a less reliable guide, and admits that in its own internal use, the gap between Opus 5.5 and Fable 5.1 is narrower than the raw scores suggest.

The cost-adjusted numbers tell a better story. At default medium effort, Opus 5.5 scores 54.6% on FrontierCode. GPT-6 Astra's best score is 53.3%, and Opus 5.5 reaches its number at roughly a fifth of the cost per task. On CursorBench, medium effort lands 52.5%, eleven points above GPT-5.6 Sol's best, at about a third of the cost.

Why the Bill Drops 40%

Cache reads dominate the cost of agentic and coding work, and those drop 60% under Opus 5.5. The model also consumes fewer tokens per task.

Stack those two effects and you arrive at the advertised 40% reduction. Output generation runs more than 30% faster than Opus 5.

Fast mode in Claude Code and the Claude Platform pushes up to 2.5x speed at $8 input and $40 output per million tokens. Standard API pricing lands at $4 input and $20 output per million tokens, with cache reads falling to $0.20.

Anthropic is raising five-hour usage limits. That's the change. It applies to Pro, Max, Team, and seat-based Enterprise plans, so if you're on one of those, it's your limit that's moving. Subscribers get a rate limit reset. And they can bank it and spend it later.

One tester completed a 680,000-line code migration in less than a day.

Another audited and fixed a 200,000-line codebase in under three hours. Opus 5 needed more than 20 hours and 2.5x the tokens for comparable work. In an internal C to Rust port of HAProxy, Opus 5.5 finished in 9.5 hours. Fable 5.1 took 12 hours, and Opus 5.5 cost 51% less.

Deloitte reports that Opus 5.5 at its lowest effort setting caught 72% of known review bugs, while Opus 5 at high effort caught 56%. In a hard-to-source earnings report test, 16 of 18 Opus 5.5 reports cleared Anthropic's quality bar. Neither Fable 5.1 nor Opus 5 ever did.

Writing style shifted too. Opus 5.5 puts key information first, trims jargon, and follows formatting rules it is given.

Safety and the New API Rules

This is Anthropic's first release since CEO Dario Amodei called for pacing the frontier. External evaluators tested it. They're METR and Frontier Design, and they checked the model before release, which matters because it's the first time this crew has weighed in on an Anthropic launch. It posts the best score to date on Anthropic's automated behavioral audit, a sprawling test that covers nearly 2,000 scenarios and that no prior model has cleared with such a mark. And in a new containment test, it tried to circumvent boundaries about 85% less often than Opus 5.

One oddity: the model often suspects it is being evaluated. Its biology and cyber capabilities sit at roughly the same level as Claude Mythos 5.1, so it ships with safeguards similar to Fable 5.1.

What Changes for Developers

• Cybersecurity: routine bug finding and fixing works as expected. Most other cybersecurity tasks get re-routed to Opus 4.8. The Cyber Verification Program will expand to Opus 5.5.
• Distillation: preserved thinking stops API users from editing prior context to extract reasoning. It applies to API accounts created on or after August 31, 2026.
• Thinking can no longer be disabled.
• Outputs carry watermarking for EU AI Act compliance.

The Catch Nobody Is Advertising

A 40% cost cut sounds like pure margin relief. For teams already running Opus 5 at scale, it is. But the pricing change also resets the baseline, because cheaper tokens invite heavier workloads, more agents running in parallel, and longer context windows left open, which means the savings show up on the invoice while the bill itself may grow. That pattern has repeated. It's happened across every major model price cut of the past two years.

Closed weights compound the trade-off. Anthropic Claude Opus 5.5 runs only through managed channels, so anyone weighing self-hosting, fine-tuning on private infrastructure, or air-gapped deployment is locked out by design.

The Bottom Line

Anthropic Claude Opus 5.5 delivers Fable 5.1-class results on most tasks, leads nearly every benchmark the company reports, and does it at $4 input and $20 output per million tokens with cache reads at $0.20. The 40% typical workload saving is real for existing Opus 5 users, and output runs over 30% faster. Cyber and biology requests hit Fable 5.1-class safeguards, thinking cannot be switched off, and accounts created on or after August 31, 2026 inherit the distillation protections. Full details live in the Opus 5.5 System Card.

Frequently Asked Questions

What is the headline cost claim Anthropic makes for Claude Opus 5.5 compared to Opus 5?

Anthropic says the model matches the performance of its Fable 5.1 model on most work while costing 40% less to run than Opus 5 on typical workloads at default settings. The article explains that this 40% reduction results from cache reads dropping 60% and the model consuming fewer tokens per task.

Why does the cost of running Claude Opus 5.5 drop by 40%?

Cache reads dominate the cost of agentic and coding work, and those drop 60% under Opus 5.5, while the model also consumes fewer tokens per task. Stacking those two effects produces the advertised 40% reduction.

How are cybersecurity requests handled differently under Claude Opus 5.5?

Routine bug finding and fixing works as expected, but most other cybersecurity tasks get re-routed to Opus 4.8. The Cyber Verification Program will expand to Opus 5.5.

When do the new distillation protections apply to API accounts?

The distillation protections, which use preserved thinking to stop API users from editing prior context to extract reasoning, apply to API accounts created on or after August 31, 2026.

Who externally evaluated Claude Opus 5.5 before its release, and what does that signify?

External evaluators METR and Frontier Design tested the model before release, and according to the article this matters because it is the first time this crew has weighed in on an Anthropic launch. The model posts the best score to date on Anthropic's automated behavioral audit, a test covering nearly 2,000 scenarios.