Anthropic released Claude Opus 5.5 on 22 September 2026. It is the first model in a new Claude 5.5 family and, according to Anthropic, it performs at the level of Claude Fable 5.1 on most work while costing 40% less to run than its direct predecessor, Opus 5[1].
That 40% figure needs a caveat before we go any further, and I’ll get to it. But first, the basics of what this model actually is and how it works.
What Opus 5.5 is built for
Opus 5.5 is not a general-purpose chatbot upgrade, like most of its predecessors. Anthropic positions it specifically for long-running agentic coding and knowledge work[2].
So, tasks where the model has to plan, execute multiple steps, use tools, check its own output, and keep going without hand-holding are best suited for Opus 5.5.
To illustrate, one early tester completed a 680,000-line code migration in less than a day, work that Anthropic says would have taken an engineering team weeks. A separate test had Opus 5.5 cut load times across every page of a web app; it succeeded 39 times out of 40.
A 200,000-line code audit now completes in under three hours, compared to the 20-plus hours previously required.
That said, those numbers come from Anthropic and its partners. Independent confirmation at scale hasn’t landed yet, so take it with a pinch of salt until independent research can verify these findings.
The context window explained
Opus 5.5 supports a one-million-token context window. One million tokens is roughly 750,000 words, or several large novels worth of text. This just means the model can hold an entire large codebase, extensive documentation, and a long task history in a single session without losing track of what happened earlier.
It also supports up to 128,000 output tokens per response.
The model accepts text, images, and files like PDFs as input. On the vision side, Anthropic reports improved reading of charts, diagrams, and screenshots without the workarounds earlier models required.
How thinking works now
This is the change most likely to catch developers off guard.
On Opus 5.5, thinking cannot be switched off.[3] On the previous model, Opus 5, you could disable the thinking process. That option is gone. Like, gone-gone.
What you can control is the effort level, which governs how deeply the model thinks before responding. The default effort is medium. Medium effort is also where Anthropic’s 40% cost saving is measured.
But there’s a catch buried deep in Anthropic’s footnotes. The headline benchmark scores are measured at maximum effort [4]. The cost saving and the benchmark performance are not measured at the same setting.
So, if you run Opus 5.5 at max effort for benchmarkable tasks, you will not automatically see a 40% saving over Opus 5.
At default medium effort, the model is efficient with tokens and the saving holds. At maximum effort, analysis by Artificial Analysis measured roughly 119,000 output tokens per task, about four times the token count of GPT-6 Astra.
Your actual spend will depend on your workload shape and which effort level your tasks require.
FROM KAYDE: Claude Opus 4.8: My AI coworker upgraded itself and I have notes
What it costs, in Rand
Opus 5.5 is priced at $4 per million input tokens and $20 per million output tokens. At today’s USD/ZAR rate of approximately R16.27 per dollar at the time of publishing[5], that works out to roughly R65 per million input tokens and R325 per million output tokens.
For comparison, Opus 5 was priced at $5 input and $25 output per million tokens. Fable 5.1, the tier above, sits at $10 input and $50 output per million tokens.
Cache reads dropped from $0.50 to $0.20 per million tokens [15]. For workflows that rely heavily on cached prompts, that reduction is meaningful.
For subscribers rather than API users: Anthropic is raising the five-hour usage limits on Pro, Max, Team, and Enterprise plans as of today, and existing subscribers get a one-off rate-limit reset they can use at any point before 22 October.
What breaks if you’re already on Opus 5
Four API changes will return errors on code that runs cleanly on Opus 5 today.
- Thinking can’t be disabled.
- Forced tool use now returns an error.
- Thinking blocks are tied to the specific model and conversation.
- And on the Claude API and Google Cloud, the older computer use tool type is not accepted; you need the updated version.
Anthropic recommends running your actual tasks at the default effort setting before changing production defaults, and comparing total billed tokens and cost per completed task rather than assuming the saving translates automatically.
Opus 5.5 is available on the Claude apps, Claude Code, the Claude Platform, Amazon Web Services, Google Cloud, and Microsoft Azure. The API model string is claude-opus-5-5.
Where it sits in the model lineup
Anthropic’s model lineup has become genuinely confusing, so here is the order as it stands today.
At the top sits the Mythos tier. Claude Mythos 5.1 and Claude Fable 5.1 share the same underlying model but carry different safety profiles.
- Fable 5.1 is generally available at $10/$50 per million tokens.
- Mythos 5.1 has additional safeguards for biology, cybersecurity, and AI research and development, and is not publicly available.
Opus 5.5 sits below the Mythos/Fable tier and above Sonnet 5 ($2/$10) and Haiku 4.5 ($1/$5). It is Anthropic’s current general-release flagship for demanding work.
Claude Sonnet 5.5 and Claude Haiku 5.5 are expected to follow in the coming weeks, though no dates are confirmed.
The safety story
Opus 5.5 is the first Anthropic model released after the company publicly called for pacing the frontier. Anthropic tested it before release with external evaluators including Frontier Design and METR.
On Anthropic’s automated behavioural audit, which covers nearly 2,000 scenarios, Opus 5.5 scored better than any recent Claude model on nearly every measure of misaligned behaviour[6]. It also took overeager or destructive actions less than any other model Anthropic has tested.
There is one flagged regression: the model is more likely than previous versions to follow malicious instructions in text that a user pastes into their own prompt. Anthropic acknowledges this in the system card and says it is working on it.
The model also launches with an anti-distillation measure called “preserved thinking”. In simple terms, it prevents API users from modifying Claude’s previous context in ways that could be used to extract its internal reasoning and potentially use it to train another model. The measure applies to API accounts created on or after 31 August 2026.
It also carries watermarking measures to comply with the EU AI Act. That is not directly South Africa’s concern right now, but it signals where Anthropic’s compliance obligations are heading and, by extension, what constraints may eventually travel with the model globally.
Who should use it and who shouldn’t
Opus 5.5 makes sense when the cost of an incomplete or wrong result is higher than the cost of the model itself. So in other words, complex engineering work, consequential research, long agent sessions, or deliverables that would otherwise require several rounds of revision.
For quick summaries, simple rewriting, classification, or high-volume routine tasks, a smaller model is the better choice. Haiku 4.5 at $1/$5 (about R16/$81 per million tokens) handles those workloads at a fraction of the cost. Or my personal favourite still: Sonnet 5 for everyday tasks.
If you’re currently on Fable 5.1, Anthropic says most work lands at the same level on Opus 5.5, and the per-token rates are 60% lower. That trade-off is worth testing on your actual tasks before making a call.
Sources:
[1] Anthropic. Introducing Claude Opus 5.5. Anthropic, Sept. 22, 2026.
[2] Anthropic. Claude Opus 5.5 Overview. Claude Platform Docs, Sept. 22, 2026.
[3] Anthropic. What’s New in Claude Opus 5.5. Claude Platform Docs, Sept. 22, 2026.
[4] Digital Applied. Claude Opus 5.5: Pricing, Benchmarks and Breaking Changes. Digital Applied, Sept. 22, 2026.
[5] XE. 1 USD to ZAR. XE, Sept. 22, 2026.
[6] Anthropic. System Card: Claude Opus 5.5. Anthropic, Sept. 22, 2026.
