Video walkthroughs · Claude Code Club

Claude Opus 5.5: What Changed and How to Use It

12 minute read · Companion guide to “Claude Opus 5.5 Just Killed Fable 5.1! (and GPT-6 Astra)”

Anthropic released Claude Opus 5.5 on September 22, 2026. The short version: it does the work Fable 5.1 was known for, it costs a lot less than Fable 5.1, and it is cheaper than the Opus 5 you were probably using yesterday. That combination does not happen often. For most people, Opus 5.5 is now the default model to use in Claude Code, and this page explains why, what the numbers really mean, and exactly how to switch.

Watch the full breakdown: price, benchmarks, and the examples Anthropic released

Part 1: What Opus 5.5 is, in one minute

Opus 5.5 is the first model in Claude's 5.5 family. That name matters, because a first model in a new family usually means smaller siblings are coming. For now, here is what you get today.

• Price: $4 per million input tokens and $20 per million output tokens. Opus 5 was $5 and $25. Fable 5.1 is $10 and $50.
• Context window: 1 million tokens, the same as Fable 5.1 and Opus 5. Max output is 128,000 tokens per reply.
• Where you can use it: the Claude apps on Pro, Max, Team, and Enterprise plans, Claude Code, the Claude API, Amazon Bedrock, Google Cloud, and Microsoft Foundry.
• Model ID for the API: claude-opus-5-5. In Claude Code, the short name opus now points to it.
• Anthropic says it will keep the model available until at least September 22, 2027.

One more practical detail: Opus 5.5 always uses adaptive thinking. It decides how hard to think based on the task and the effort level you set. You do not turn thinking on or off by hand any more.

Part 2: Read the price right, 20 percent off the sticker and about 40 percent off the job

You will see two different savings numbers for Opus 5.5, and both are true. They measure different things, and mixing them up leads to bad budgets.

• The list price dropped 20 percent. Input went from $5 to $4 per million tokens, and output went from $25 to $20. Same ratio on both.
• The cost to finish a task dropped about 40 percent, by Anthropic's measure. That is the price cut plus the model using fewer tokens to reach the same quality.
• Against Fable 5.1, the list price is 60 percent lower. $4 against $10 on input, $20 against $50 on output.

The second number is the one that shows up on your bill. You pay for tokens used, not for the price per token, so a model that reaches the answer in fewer tokens saves you money twice. Cache reads also got cheaper, down to $0.20 per million tokens from $0.50, which helps anyone running long sessions that reread the same files.

Part 3: Why it beats Fable 5.1 at a fraction of the price

Fable 5.1 was the model people used for big, sprawling jobs: codebase-wide migrations, full audits, long agent runs. Opus 5.5 now scores higher on the tests that measure exactly that kind of work. These are Anthropic's published numbers.

• Terminal-Bench 4.0 (agentic coding in a terminal): Opus 5.5 66.4 percent, Fable 5.1 55.8 percent, GPT-6 Astra 57.9 percent.
• FrontierCode: Opus 5.5 54.4 percent, Fable 5.1 50.3 percent. At default effort, Opus 5.5 beat GPT-6 Astra running at its maximum effort, for about a fifth of the cost.
• CursorBench: Opus 5.5 57.8 percent, Fable 5.1 51.8 percent.
• GDPval-AA (real knowledge work): Opus 5.5 1846 Elo, Fable 5.1 1735 Elo.
• OSWorld (computer use): Opus 5.5 81.8 percent, Fable 5.1 80.7 percent. This was the area GPT-6 Astra claimed as its edge.

The chart that stood out most in the video compared cost against accuracy. Opus 5.5 on its lowest effort setting landed about level with Fable 5.1, at roughly a fifth of the cost, around $1.29 against about $5.70 in the launch chart. On medium and high effort it pulls clearly ahead.

Keep benchmarks in their place, though. They measure tasks that are easy to score. Your work is not a benchmark. A computer use gap of about one point is a tie in practice. The coding gap of more than ten points on Terminal-Bench is big enough that you will likely feel it.

Part 4: The real upgrade is that it makes things up less often

This is the part of the launch we think matters most, and it got less attention than the benchmarks. Anthropic asked Opus 5.5, Fable 5.1, and Opus 5 to write a report on a company's quarterly performance. The catch: the models could only use a copy of the web where the earnings release was hard to find. Any invented figure or made-up quote failed the report.

Opus 5.5 cleared that bar in 16 of 18 reports across different effort settings. Fable 5.1 and Opus 5 did not clear it once. When the model cannot find the fact, Opus 5.5 is far more likely to say so instead of filling the gap with something that sounds right.

Here is why that matters for you. As models get better, people check their work less. That is exactly when a confident wrong number does the most damage: in a client report, a sales page, a pitch deck. A model that is slightly smarter matters less than a model you can trust with a fact.

Part 5: It finishes faster and writes shorter

Two more examples from the launch show how the model behaves on real work.

The first was a merger analysis. Anthropic asked Opus 5.5 and Opus 5 to study a proposed deal between two made-up HR software companies. Each one built a financial model in Excel, then turned it into an executive presentation on whether the price made sense. Both reached the same conclusion. Opus 5.5's model was more thorough and its slides were easier to read, while Opus 5 had minor errors. Opus 5.5 finished in 63 minutes against 93 for Opus 5, at half the cost.

The second was about writing. Anthropic asked both models to explain a billing bug. Opus 5 answered with a long, jargon-heavy walk through the code. Opus 5.5 opened with the answer: the extra drop is a bug in the billing refactor, the free tier change only explains $150 of it, and the rest comes from the bug. Then it stopped.

That matters in three ways. It is easier to read. It puts the most important thing first, so you can act on the first line. And a shorter answer uses fewer output tokens, which are the expensive ones at $20 per million. Anthropic also says the model is less likely to use jargon or odd phrases, which should make it a better partner over a long session.

Part 6: How to switch to Opus 5.5 in Claude Code

In the Claude Code desktop app, open the model selector and pick Opus 5.5. In the command line version, type the model command in any session. The short name opus now points to Opus 5.5, so you do not need the full ID.

Copybash/model opus

# or start a new session on it
claude --model opus

# check which model you are on
/model

Running /model opus saves it as your default for future sessions too. If you only want to try it for the current session, open the /model picker and press s on your choice instead. If you already had opus set as your default, you are already on Opus 5.5 and nothing needs to change.

Part 7: Pick the right effort level

Effort controls how much work the model puts into each answer. More effort means more thinking, more checking, and more tokens. Opus 5.5 has five levels, and its default is medium. Fable 5.1 and Sonnet 5 default to high, so if you compare them without touching effort, you are not comparing like with like.

• Low: quick edits, renaming, formatting, simple questions. Still about as strong as Fable 5.1 on Anthropic's chart.
• Medium: the default. Right for most building, writing, and research.
• High: bigger features, bugs you have already failed to fix once, anything with many moving parts.
• Xhigh and max: long migrations, full codebase audits, and hard problems where getting it right matters more than cost or speed.

Set it with /effort followed by the level, for example /effort high. Typing /effort on its own opens a slider. Start on medium and only go up when a task actually fails, not because a task feels important.

Part 8: When to still reach for another model

Opus 5.5 is the new default, not the only choice. Here is how we would route work today.

• Opus 5.5 on medium: almost everything. Building, writing, research, planning, agent runs.
• Fable 5.1: only when you have tested a specific job and it clearly does better there. Based on the benchmarks, that list is now short, and it costs 2.5 times as much.
• Sonnet 5 at $2 and $10: high-volume, repeatable jobs where you have already proven the quality is good enough.
• Haiku 4.5 at $1 and $5: tiny sorting and tagging jobs. Note that Anthropic plans to retire it no sooner than October 15, 2026, so do not build anything new around it.

The rule we use: new model, same three jobs. Every time a model drops, run the three tasks you do most often on the new model and your current one. Compare cost, time, and quality. It takes an hour and it beats any chart, including the ones on this page.

Part 9: What to watch out for

A few things can trip you up after the switch.

• Old API code that sets a thinking budget by hand will not work. Opus 5.5 only accepts adaptive thinking, so remove budget_tokens and use the effort setting instead.
• Shorter answers are not lazier answers. If you had prompts that asked for long explanations to get quality, try removing that. You may be paying for words you do not need.
• Some biology and cybersecurity requests are handled more carefully and may fall back to an older model. For almost all business and building work you will never see this.
• Anthropic notes the model often suspects when it is being tested. Treat launch benchmarks as a strong signal, then confirm on your own work.

Also expect more releases soon. Opus 5.5 is the first of the 5.5 family, and press reports say Sonnet 5.5 and Haiku 5.5 are expected in the coming weeks. Do not rebuild your whole setup around one model's quirks, because the routing will change again.

Part 10: Switch over in the next 20 minutes

1. Open Claude Code and run /model opus, or pick Opus 5.5 in the desktop model selector.
2. Run /model again to confirm you are on Opus 5.5.
3. Search your skills and agent files for any pinned model and update the ones that should move.
4. Leave effort on medium. Write down one task you would move to high.
5. Pick the three jobs you run most often and run each one on Opus 5.5.
6. Compare the cost and quality with what you got last week, and keep whichever wins for each job.

That is the whole switch. Most people will be done in one sitting and paying less by tonight.

Read this online at claudecodeclub.ai