Claude Opus 5.5: The Big Upgrade Is Efficiency

Imagine handing an AI coding assistant a 200,000-line repository at the end of the day and finding a tested patch waiting the next morning. That is the practical story behind Claude Opus 5.5, released on September 22, 2026. It is the first member of Anthropic’s Claude 5.5 family, meaning a new group of related models, but the interesting part is not the version number. Opus 5.5 is designed to handle difficult work with fewer retries, fewer tool calls, and fewer tokens.

A large language model, or LLM, is software trained to understand and generate text, code, and other data. Opus 5.5 pushes that software toward agentic work: tasks where the model reads files, uses tools, changes code, runs tests, and checks the result across many steps. That is a different job from writing one function in a chat box.

The headline is efficiency

Anthropic says Opus 5.5 performs at roughly the level of its sibling model Claude Fable 5.1 on most work, while costing about 40% less to run than Opus 5 on typical workloads. The improvement is not only a matter of price per request. The model also tends to finish tasks with fewer actions and less generated text.

A token is a small piece of text processed by a language model. Depending on the language, a token might be part of a word, a whole short word, or a punctuation mark. For a short conversation, saving tokens may barely matter. For an agent that reads a large codebase, calls tools dozens of times, and works overnight, token use becomes part of the operating budget.

The published benchmark results show the direction of the change. On Terminal-Bench 4.0, which measures complex command-line tasks, Opus 5.5 scored 66.4%, compared with 52.3% for Opus 5. On FrontierCode, a test of whether code changes would be accepted into a real project, it scored 54.4%, ahead of Opus 5 at 48.0%. Its GDPval-AA score, an Elo-style rating for professional work across many occupations, reached 1846, compared with 1708 for Opus 5.

Those numbers are useful, but they are not a promise that every task will improve by the same amount. The tests use different effort settings, and real projects contain messy requirements, missing documentation, and surprising dependencies that a benchmark cannot fully reproduce.

What is Claude Opus 5.5 good at?

The strongest fit is long-running technical work. Early tests described with the release include a 680,000-line code migration completed in less than a day and an audit of a 200,000-line codebase finished in under three hours. The comparable Opus 5 run took more than 20 hours and used about two and a half times as many tokens.

Another test asked Opus 5.5 and Fable 5.1 to translate HAProxy, open-source software that routes web traffic between servers, from C into Rust. Both versions passed nearly all of HAProxy’s regression tests, which are automated checks that make sure existing behavior still works. Opus 5.5 finished in 9.5 hours instead of 12 and cost 51% less in that evaluation.

Why does this feel different from autocomplete? Traditional autocomplete suggests the next line while you type. An agentic coding system can inspect the repository, form a plan, edit several files, run the test suite, investigate failures, and revise its patch. The value comes from keeping the whole loop moving without losing track of the original goal.

The price story is about cache reads

Opus 5.5’s listed prices are lower across the main categories:

Prompt caching means storing a reusable part of a request, such as a system instruction or a large project description, so later requests do not need to process that material at the full input price. A cache read is the cheaper operation of reusing that stored context. This matters for coding agents because the same repository information may be needed across many tool calls.

A small cost calculator makes the pricing easier to reason about:

def opus_55_cost(input_tokens, output_tokens,
 cache_reads=0, cache_writes=0):
 return (
 input_tokens * 4
 + output_tokens * 20
 + cache_reads * 0.20
 + cache_writes * 5
 ) / 1_000_000

estimate = opus_55_cost(
 input_tokens=200_000,
 output_tokens=40_000,
 cache_reads=800_000,
)

print(f'Estimated cost: ${estimate:.2f}')

This example estimates about $1.76, assuming each token category is counted separately. Opus 5.5 also has a fast mode in Claude Code and the Claude Platform, with up to 2.5 times the speed at $8 per million input tokens and $40 per million output tokens.

Safety has to follow capability

A model that can edit code and operate tools for hours needs more than a refusal message. It needs controls around the actions it takes. Opus 5.5 includes an action classifier that screens tool actions before they run, an auditable sandbox for isolated execution, and code review intended to catch vulnerabilities before changes are merged.

It also has stronger defenses against prompt injection. Prompt injection is an attempt to hide instructions inside a webpage, document, issue, or other input so that an AI system follows the attacker’s directions instead of the user’s. The model was tested against these attacks across coding, browsing, computer use, and tool workflows.

The safeguards become stricter in high-risk areas. Many cybersecurity tasks are routed to Opus 4.8, while broader biology research access is reserved for verified organizations. That may make some workflows less direct, but it reflects an important design choice: raw capability is not the only measure of a production model.

The release also describes better results on behavioral evaluations, including tests for crossing containment boundaries and taking hard-to-reverse actions. Still, no evaluation covers every real-world situation. A model can perform well in a controlled test and encounter a novel failure once it is connected to unfamiliar tools and data.

Clearer communication is a technical improvement

Opus 5.5 is also meant to be easier to work with. Its responses put the main conclusion earlier, use less unexplained jargon, and follow writing instructions more consistently. That sounds like a presentation detail, but it affects engineering quality: a readable plan is easier to review, and a clear explanation makes a questionable assumption easier to spot.

For long sessions, communication becomes part of control. Developers need to understand what an agent changed, which tests passed, where uncertainty remains, and what should happen next. A model that produces shorter, better-structured updates reduces the time spent reconstructing its decisions.

Where Opus 5.5 fits

Claude Opus 5.5 is most compelling for large migrations, codebase audits, research reports, financial analysis, and business workflows that require many connected steps. Smaller models may remain a better choice for short answers, routine summarization, or high-volume tasks where the highest level of reasoning is unnecessary.

At launch, Opus 5.5 is available through the Claude Platform and major cloud platforms, including Amazon Web Services, Google Cloud, and Microsoft Azure. Developers using the platform can select it with the model identifier claude-opus-5-5.

The broader shift is easy to miss if you look only at leaderboard scores. Opus 5.5 is not merely trying to answer harder questions. It is trying to waste less motion while doing difficult work. For coding agents and other long-running systems, that combination of capability, speed, clearer communication, and lower token use may matter more than a small increase in any single benchmark.