Anthropic Launches Claude Opus 5.5 With Faster Performance and Lower Costs

Source: This report covers Anthropic’s September 22 launch announcement. Performance and safety results below are reported by the company and its cited evaluators, rather than BGTake benchmark measurements.

Anthropic Claude Opus 5.5 Arrived on September 22, 2026. It introduces Anthropic’s new Claude 5.5 model family. The model promises faster and cheaper AI performance. 

It targets coding and professional knowledge tasks. Anthropic has also added important safety improvements.

Anthropic Claude Opus 5.5 Arrived With Major Upgrades

Claude Opus 5.5 focuses heavily on practical performance. Anthropic positions it near Fable 5.1 on most work while reducing token costs. Task cost and raw computing requirements are not interchangeable.

Typical workloads can cost forty percent less. Output generation is also over thirty percent faster. Fast mode can deliver 2.5 times faster speeds.

Coding Performance Gets A Major Boost

Anthropic tested the model across demanding coding tasks. One migration involved nearly 680,000 lines of code. The migration finished in less than one day. 

Another test audited 200,000 lines of code. That complete audit finished within three hours. Opus 5 reportedly needed over twenty hours.

Other coding tests produced impressive practical results too.

  • HAProxy’s C-to-Rust rewrite finished within 9.5 hours.
  • Fable 5.1 required nearly twelve hours instead.
  • The rewrite also cost fifty-one percent less.
  • Web testing succeeded across thirty-nine of forty pages.
  • Only one page failed during that testing.

Opus 5.5 also performed strongly against competitors. It matched GPT-6 Astra on Terminal-Bench 4.0. Anthropic describes comparable Terminal-Bench performance at about 40% of Astra’s cost in its default-effort comparison—not 40% less. 

FrontierCode showed another major cost advantage overall. CursorBench scored eleven points above GPT-5.6 Sol. That result cost roughly one-third as much.

Anthropic Claude Opus 5.5 Handles Knowledge Work

The model also targets demanding professional workflows. Anthropic tested research and financial analysis tasks. Quarterly reports received detailed automated quality checks. 

Sixteen of eighteen reports passed Anthropic’s quality bar. The checks reviewed every figure and quoted statement. Merger analysis produced identical conclusions across models. Opus 5.5 completed the work within 63 minutes. 

Opus 5 required approximately ninety-three minutes instead. The newer model also cost fifty percent less. Walleye Capital testing produced another interesting result. 

Opus 5.5 largely solved the evaluation suite. It also discovered an error within evaluation instructions. Other tested models had not found that error.

Communication Feels More Natural Now

Anthropic also changed how Opus communicates. Important information now appears earlier within responses. 

The model uses fewer unusual phrases overall. It also uses less complicated technical language. Long conversations should remain easier to understand. The model follows user writing rules better.

Anthropic Claude Opus 5.5 Brings Safety Improvements

Safety remains another major focus for Anthropic. The company tested nearly 2,000 behavioral scenarios. These tests examined honesty and misaligned behavior. Opus 5.5 improved across nearly every measurement.

Containment attempts reportedly dropped by eighty-five percent. Every observed attempt remained low severity overall. The model also reported those attempts itself.

External organizations evaluated the model before release. Those organizations included METR and Frontier Design. Prompt-injection testing covered several different environments. These included coding, browsing, tools, and computer use.

New Safeguards Cover Several Sensitive Areas

Opus 5.5 includes cybersecurity and biology safeguards. Most cybersecurity tasks currently route through Opus 4.8. Routine software bug fixing remains supported normally.

Anthropic also expanded its Cyber Verification Program. It provides three increasingly permissive trusted-access levels. Vetted organizations can request life-science access too.

The model includes preserved-thinking protection against extraction. This safeguard protects reasoning within eligible API accounts. Zero data retention support is also available. Watermarking measures support European AI Act compliance.

Anthropic Claude Opus 5.5 Pricing And Availability

The new model costs less than Opus 5. Input tokens cost four dollars per-million tokens. Output tokens cost twenty dollars per-million tokens. 

Cache reads cost twenty cents per-million tokens. Cache writes cost five dollars per-million tokens. Fast mode uses separate pricing for faster performance.

  • Fast input costs eight dollars per-million tokens.
  • Fast output costs forty dollars per-million tokens.
  • Fast mode offers speeds up to 2.5x faster.

Opus 5.5 works across major cloud platforms. These include AWS, Google Cloud, and Microsoft Azure. The API uses claude-opus-5-5 as its identifier. Claude Code also supports the faster mode. 

Subscription users receive increased five-hour usage limits. Anthropic will release two more models soon. Claude Sonnet 5.5 will arrive within weeks. Claude Haiku 5.5 will follow within weeks. Both models promise similar efficiency and safety.

Claude Opus 5.5 focuses strongly on practical improvements. It delivers faster work with significantly lower costs. Coding and professional workflows receive major upgrades. 

Communication also feels clearer and more natural. Safety testing has expanded across important risk areas. More Claude 5.5 models will arrive soon.

What the pricing means for your workload

Standard API pricing and a task’s total bill measure different things. A task can consume input, output and cached tokens, and retries add usage. Compare your own workload rather than treating a benchmark cost advantage as a guaranteed saving. Subscription limits are separate from API token prices. Confirm current availability and prices on Anthropic’s Opus product page.

Correction, September 24: The Terminal-Bench comparison now says “about 40% of the cost,” correcting the earlier “40% less” wording.

Leave a Reply

Your email address will not be published. Required fields are marked *