Anthropic shipped two new Claude models six days apart. Claude Opus 5.5 arrived on 22 September 2026 and Claude Sonnet 5.5 on 28 September. Since then the question in every one of our Claude classes has been the same: which one should my team actually use?
The honest answer is that for most office work the gap between them is now small, and the bigger difference is cost and speed. Here is what changed, what the numbers say, and the rule of thumb we give Malaysian teams.
The two models side by side
| Opus 5.5 | Sonnet 5.5 | |
|---|---|---|
| Released | 22 September 2026 | 28 September 2026 |
| API price (per million tokens, in / out) | US$4 / US$20 | US$2 / US$10 |
| Context window | 1M tokens | 1M tokens |
| Max output | 128K tokens | 128K tokens |
| Anthropic’s own label | Long-running agentic coding and knowledge work | Best combination of speed and intelligence |
| Claude plans | Pro, Max, Team, Enterprise | All plans, including Free |
| Default in Claude Code (paid plans) | Yes | Available as the “sonnet” option |
Source: Anthropic model overview, Opus 5.5 and Sonnet 5.5 announcement pages, claude.com/pricing, Claude Code docs. Checked 4 October 2026. API prices exclude tax.
Two things stand out. First, Opus got cheaper. Its list price fell from US$5/US$25 to US$4/US$20, and Anthropic says it costs about 40% less than Opus 5 on typical workloads because it also uses fewer tokens. Second, Sonnet 5.5 now has the same 1M-token context window as Opus, so “the document is too long” is no longer a reason to reach for the bigger model.
What the benchmarks actually say
| Benchmark | Opus 5.5 | Sonnet 5.5 |
|---|---|---|
| GDPval-AA v2.1 (real knowledge-work tasks) | 1846 | 1844 |
| OSWorld 2.1 (using a computer like a person) | 81.8% | 80.1% |
| Terminal-Bench 4.0 (agentic terminal work) | 66.4% | 70.6% |
| FrontierCode 1.1 (hard coding) | 54.4% | 46.2% (max effort) |
Figures as published by Anthropic on the Opus 5.5 and Sonnet 5.5 pages, September 2026.
Look at the first row. GDPval-AA scores models on the kind of tasks office workers do: analysis, drafting, spreadsheets, reports. Opus 5.5 and Sonnet 5.5 are effectively tied. On Terminal-Bench, Sonnet is actually ahead. Opus pulls clearly in front only on the hardest coding work, which is FrontierCode here, and on long multi-step jobs that run for an hour or more.
Benchmarks are a starting point, not a verdict. But they line up with what we see in class: a finance executive summarising a 90-page tender document will not notice a difference. A developer asking Claude to migrate a legacy codebase will.
When to use Sonnet 5.5
- Everyday drafting and analysis. Emails, proposals, meeting notes, policy summaries, first drafts of board papers.
- Documents, slides and spreadsheets. Anthropic specifically calls out polished documents, slides and spreadsheets as a Sonnet 5.5 strength.
- Anything you do twenty times a day. It is over 30% faster than Sonnet 5, and speed matters more than the last few points of quality when you are iterating.
- Automations that call Claude in volume. At half Opus’s API price, Sonnet is the default for n8n workflows that classify emails, extract invoice fields or draft WhatsApp replies.
When Opus 5.5 is worth it
- Long agentic jobs. Cowork or Claude Code tasks that read a whole folder, make a plan and work through dozens of steps. Anthropic’s showcase example is a 680,000-line code migration finished in under a day.
- Hard reasoning where a wrong answer is expensive. A tricky tax treatment, a contract with interacting clauses, a financial model with circular references.
- Using a computer for you. It edges Sonnet on OSWorld, the computer-use benchmark, which matters for browser and desktop tasks.
A 20-second rule for business teams
- 1Start every task on Sonnet 5.5It is fast, it handles long documents, and for most office work the output is as good.
- 2Switch to Opus 5.5 when Sonnet strugglesIf the answer is shallow, misses a step or the job is long and multi-stage, rerun it on Opus.
- 3Use Opus by default for Claude Code and long Cowork jobsThat is already the default in Claude Code on paid plans, and long agentic work is where it earns its keep.
- 4In automations, price it per taskRun a sample of 50 real cases on both. If Sonnet’s accuracy is acceptable, the 50% saving compounds every month.
What this means on each Claude plan
On the Free plan you get Sonnet, not Opus. On Pro, Max, Team and Enterprise you can switch between both in the model menu, and Anthropic raised usage limits on paid plans when Opus 5.5 launched. If your company is still on personal Pro accounts, that is the bigger issue to fix: for company data, use Team or Enterprise. Our Claude Pro vs Team vs Enterprise guide explains why.
Developers get both models on Anthropic’s API and through AWS Bedrock, Google Vertex AI and Microsoft Foundry, and Opus 5.5 is also selectable in GitHub Copilot. API model IDs are claude-opus-5-5 and claude-sonnet-5-5.
The model is rarely the bottleneck. In our classes, the same person gets dramatically better output by giving Claude the right files, a clear brief and an example of what good looks like. That skill transfers across every model release. Picking Opus over Sonnet does not.
What about Fable 5.1 and Haiku?
Fable 5.1, released on 1 September, is Anthropic’s most capable model, priced at US$10/US$50 per million tokens. Anthropic says Opus 5.5 performs at Fable 5.1’s level on most work for well under half the price, so for business teams Fable is now a specialist choice. At the small end, Haiku 4.5 is still the current Haiku for very cheap, high-volume calls. We will update this guide when a newer Haiku ships.
Two days on Claude, Cowork and Claude Code with the 5.5 models: when to use which, skills, MCP connectors and multi-agent workflows on your own files.
Next public class: 22–23 October 2026 at Worq TTDI · seats available
Preparing for the exam rather than the job? See our Claude Certified Architect prep. Everything else is on the Claude training in Malaysia page.
HRD Corp claimable. AITraining2U is an HRD Corp registered training provider. Our public and in-house courses are claimable under the HRD Corp Claimable Courses (SBL-KHAS) scheme, and we prepare the grant paperwork with you. See HRDC-claimable AI training: courses, costs and how to claim.