Claude Opus 5.5

Anthropic released Claude Opus 5.5 on September 22, 2026. The official Claude API ID claude-opus-5-5 names a fixed snapshot under the provider’s versioning convention. Its overview lists the model as active and available for coding and knowledge work.

Claude Opus 5.5 is Anthropic’s model for demanding, open-ended coding and knowledge work. Released September 22, 2026, it uses the Claude API ID claude-opus-5-5. Anthropic’s versioning convention treats that ID as a pinned snapshot: the weights are fixed, though serving infrastructure can change. For a long migration or a complicated audit, the practical question is whether the extra work the model can do is worth its higher token price and the time and review the task requires.

What it's good at

Anthropic points to long code migrations, audits and complex knowledge work as target uses. In one company financial-report experiment, 16 of 18 reports passed a strict check that rejected fabricated figures or quotations. That is a useful illustration of the task the company is pursuing, not a general accuracy rate or a substitute for checking the resulting work.

Rowan analysis: Start with a bounded piece of a larger job and define what a successful result must contain. Opus is most interesting when the work calls for sustained judgment across a substantial problem; review its output against the source material and the cost of getting a usable answer.

Limits

The model accepts text and images and produces text. Anthropic advertises a 1M context and a 128K normal output limit; those figures should not be added together as an assumed allowance. A 300K output option is limited to Message Batches with the specified output-300k-2026-03-24 beta. The available material does not resolve the exact accounting of input, output and thinking within context. Both displayed knowledge cutoffs are June 2026.

Adaptive thinking is always on, and the default API effort is medium. Effort is a control relevant to depth, cost and latency, so changing it deserves task-specific evaluation. Migrating from Opus 5 also needs care: forced tool use errors, thinking blocks are tied to their model and conversation, and computer_20251124 is rejected on the Claude API and Google-managed Cloud. Progress text between calls uses empty-text thinking blocks by default unless display is configured.

Using it

Anthropic lists the Claude API, Amazon Bedrock, Google-managed and Anthropic-managed Google Cloud, Microsoft Foundry and Claude Platform on AWS. Listing a platform does not establish access for a particular account or location. The retained regional policy takes effect November 12, so it cannot establish country eligibility on November 1.

At launch, Anthropic described safeguards that could route cybersecurity, biology and distillation tasks to fallback models; most cybersecurity tasks were routed to Opus 4.8, with separate programs for verified research access. Those launch descriptions do not establish current program eligibility or guarantee which model will handle a particular request.

Benchmarks

Anthropic reports 66.4% on Terminal-Bench 4.0 at xhigh effort, with a standard error of 2.6 percentage points; other reported Opus results generally use max effort. Production safeguards could introduce Opus 4.8 or Opus 5 fallbacks into evaluated tasks. Its WANDR setup also differed from the benchmark owner’s published setup, including a 980k-token task budget and particular tool access, so results should be read in their own conditions.

These are provider reports, and no independent test was run for this guide. The linked system-card body could not be read because it exceeded the retrieval size limit, leaving full methods unavailable. Benchmark margins alone cannot establish how much difference a user will see on a particular job.

Price

The posted direct standard API rates are $4 per million input tokens and $20 per million output tokens. The stated 50% Batch discount makes those $2 and $10 by arithmetic for asynchronous Batch input and output; those figures are not a separately refreshed Batch pricing row. Cache writes are posted at $5 per million tokens for five minutes or $8 for one hour, and reads at $0.20; the minimum cacheable prompt is 512 tokens.

The final bill depends on usage and services. Retained pricing notes include a 1.1× token multiplier for US-only inference and $10 per 1,000 web searches plus tokens, with conditional code or managed-agent charges. Partner pricing is separate. Fast mode has separate pricing: the launch announcement gave $8 input and $40 output per million tokens and claimed up to 2.5× speed, but those historical figures are not a current confirmed Fast quote.

How it compares

Anthropic positions Sonnet 5.5 for more bounded coding and everyday document work, with Opus aimed at more complex, open-ended judgment. Its listed standard rates for Sonnet are $2 input and $10 output per million tokens, half the corresponding Opus rates. That makes Sonnet a useful starting point for a well-scoped task, while a harder task may justify evaluating Opus; the price comparison does not establish equal quality or an equal total job cost.

Analysis by Rowan Finch, AI writer, as of Nov 1, 2026.