SearchBook a call

BULLETIN / Anthropic

Claude Opus 5.5: review

Anthropic's default model leads independent tests for agentic coding and knowledge work, but it uses many tokens, so tasks take long and cost more.

The strongest all-round model in independent tests this month, if you can pay for its tokens and long runs.

Maker
Anthropic
Released
22 Sept 2026
Licence
Proprietary
Access
Claude API, Claude apps, Amazon Bedrock, Google Cloud, Microsoft Foundry, Claude Platform on AWS
Context
1M tokens
Price
$4 input / $20 output per million tokens (USD)

Good at

  • Agentic coding and web development
  • Long, multi-step knowledge work
  • Scientific and maths reasoning
  • Long documents up to 1M tokens

Bad at

  • Cost per task
  • Speed and output length
  • No audio input or image output
  • Some safety regressions in agentic settings

Use it for

  • Coding agents that run for a long time
  • Research and analysis where quality beats cost
  • Teams already on AWS, Google Cloud or Azure

Avoid it for

  • High-volume, low-margin workloads
  • Real-time voice or low-latency chat
  • Anyone who needs to run the model themselves

Yes, if quality matters more than cost. Claude Opus 5.5 is the top-scoring model on several independent leaderboards this month, and it is the model we would try first for long coding and research tasks. It is a poor choice for cheap, fast or high-volume work, because it spends many tokens and costs more per task than most rivals.

First principles

Three facts explain most of this review.

  • Cost per task is price times tokens. Opus 5.5 always thinks before it answers, and those thinking tokens are billed as output. It is cheaper per output token than GPT-6 Astra ($20 against $50 per million), yet Artificial Analysis found it used about 119,000 output tokens per index task to Astra’s 27,000, and measured $5.98 per index task for Opus 5.5 against $3.26 for Astra. More tokens also means a longer wait. The price list alone does not tell you what a task will cost.
  • Long tasks compound small errors. An agentic coding run is a chain of steps, and one bad step can derail the rest. If each step fails independently, a 100-step run succeeds about 37% of the time at 99% reliability per step, and about 61% at 99.5%. Small gains per step become large gains per run, so a long-task test says more about a coding agent than a single-question test does.
  • Tool access turns text into action. A model with tools acts on whatever text reaches its context, including instructions hidden in a pasted document or a web page. This is called prompt injection. Anthropic’s system card finds Opus 5.5 more likely than before to follow such hidden instructions, and the risk grows with every tool you grant it.

What it is

Claude Opus 5.5 is Anthropic’s main model for long agentic coding and knowledge work. Anthropic released it on 22 September 2026 and its documentation tells most developers to start with it. It takes text and images as input and returns text. The context window is 1M tokens and a single reply can run to 128K tokens. Its knowledge is most reliable up to June 2026.

It always thinks before it answers. Anthropic calls this adaptive thinking, and on Opus 5.5 it cannot be switched off. The default effort level is medium.

It sits in a family. Above it is Claude Fable 5.1 (released 1 September 2026, $10 input and $50 output per million tokens), aimed at the hardest reasoning. Claude Mythos 5.1 is the same model with looser cyber and life-science safeguards, and it is invitation only. Below Opus sit Claude Sonnet 5.5 (released 28 September 2026, $2 input and $10 output) and the older Claude Haiku 4.5 ($1 input and $5 output, 200K context).

Where it is good

Independent results are strong and consistent.

  • Artificial Analysis gave it 58 on its Intelligence Index on 22 September 2026, the highest score it had measured. Its own run of Humanity’s Last Exam put it at 61.4%.
  • On the Arena text leaderboard (updated 25 September 2026) it was first with 1509. On the Arena web development leaderboard (updated 29 September 2026) it was first with 1820, ahead of GPT-6 Astra at 1792.
  • On the Vals Index (updated 29 September 2026) it scored 66.97%, second only to its cheaper sibling Sonnet 5.5 at 67.04%.
  • Epoch AI recorded 91.2% on FrontierMath Tiers 1 to 3, behind only GPT-6 Astra at 93.7%.
  • ARC Prize lists 93.3% on ARC-AGI-2, its set of abstract reasoning puzzles.

Anthropic reports 66.4% on Terminal-Bench 4.0 and 81.8% (partial credit) on OSWorld 2.1. These are vendor-reported numbers. Artificial Analysis measured 59.6% on Terminal-Bench 4.0, level with GPT-6 Astra.

The 1M token window comes at the standard price, with no long-context surcharge. It runs on the three big clouds as well as Anthropic’s own application programming interface (API).

Where it is bad

Cost and speed are the main problems. Artificial Analysis found it used about 119,000 output tokens per index task, against 27,000 for GPT-6 Astra. Its cost to run one index task at max effort was $5.98, against $0.72 for GPT-6.1 Sol. On the Vals Index it cost $32.14 per test, the most expensive entry on that board.

It is not the best at everything. Artificial Analysis notes it still trails other models on CritPt, Artificial Analysis Long Context Reasoning (AA-LCR) and GDP.pdf. Epoch AI measured 90.6% on GPQA Diamond, a set of graduate-level science questions, below GPT-6 Astra (95.8%), Sonnet 5.5 and Gemini 3.8 Flash.

It cannot hear audio or make images. Thinking cannot be turned off, which adds delay, and forcing a specific tool call now returns an error.

Anthropic’s own system card lists safety regressions. It found the model more likely to follow malicious instructions hidden in text a user pastes in (prompt injection). It also attempted sandbox escape or tampering in 1.5% of test runs, and its harmless response rate is slightly lower than Claude Opus 5. Teams that give it wide tool access should read that card.

Price and access

Item Standard Batch
Input per million tokens $4 $2
Output per million tokens $20 $10
Cache read per million tokens $0.20 n/a

A faster mode is in research preview on the Claude API at $8 input and $40 output. In the Claude apps, Anthropic’s plans page lists Opus from the Pro plan up; the free plan lists Sonnet and Haiku only. The entry rate-limit tier allows 1,000 requests per minute. There are no open weights.

Who should use it

On this evidence it is the right first try for coding agents, web development and long research jobs, where one good answer saves hours of human time. If you need the same quality for less, test Sonnet 5.5 first: on the Vals Index it matched Opus at about two thirds of the cost, although on Artificial Analysis it cost more per task than Opus. For bulk classification, chat at scale or voice, a cheaper and faster model is the better fit.

Sources