Claude Opus 5 for Coding: Price, Tokens and Setup

Claude Opus 5 dropped yesterday and my timeline filled up with the usual stuff: benchmark screenshots, people saying it’s insane, people saying it’s hype. What almost nobody covers is the only thing that actually matters if you code with AI every day: should I migrate my workflow or not?

I went through the official docs, the migration guide and the independent numbers from Artificial Analysis, and here’s the summary with verifiable data. Spoiler: the answer isn’t “yes” for everyone.

What Opus 5 is, in hard numbers

According to the official model table, Opus 5 (claude-opus-5) is this:

Pricing: $5 per million input tokens, $25 per million output tokens. Exactly the same as Opus 4.8. This is the most important part of the launch and the most overlooked: more model for the same money.
1M token context window by default (no beta header) and up to 128k output tokens.
May 2026 knowledge cutoff, the most recent in the family.
Adaptive thinking: the model decides how much to think on each request. Manual extended thinking with budget_tokens is gone.
Available on the Anthropic API, Bedrock, Google Cloud and Microsoft Foundry. It’s also the default model on Claude Max and the strongest model available on Claude Pro.

There’s also a Fast mode that runs it at roughly 2.5x the default speed, at twice the base price - same as Opus 4.8 had.

Benchmarks: what you gain

Here are the numbers from the official announcement and from third parties, each with its source:

On Frontier-Bench v0.1, Opus 5 beats every model on the market and more than doubles Opus 4.8’s score at a lower cost per task.
On CursorBench 3.2, at max effort it lands within 0.5% of Fable 5’s peak score (Anthropic’s top-tier model), at half the cost per task.
On ARC-AGI 3 it triples the next-best model, and on OSWorld 2.0 (computer use) it beats Fable 5’s best result at just over a third of the cost.
On the Artificial Analysis Intelligence Index (data as of July 25, 2026), Opus 5 at max effort scores 61, number 1 out of 190 models. For context: the median in its class is 32.

Anthropic’s own benchmarks should always be taken with a grain of salt. But when an independent third party ranks it first out of 190 models the day after launch, the signal is pretty clear: for agentic coding it’s the new ceiling at this price point.

The real cost isn’t the price per token

And here’s the part the headline doesn’t tell you. The per-token price matches Opus 4.8, sure. But your bill doesn’t depend only on the per-token price: it depends on how many tokens the model burns, and Opus 5 burns more:

Thinking is on by default. On Opus 4.8, a request without a thinking field ran without thinking. On Opus 5, that same request thinks. It’s a documented breaking change in the migration guide: max_tokens is a hard limit on thinking + response combined, so if you had max_tokens tuned for no-thinking workloads, the reasoning will eat it.
It’s more verbose. In Artificial Analysis measurements, Opus 5 generated 100M output tokens to complete their Intelligence Index, versus a 63M median across comparable models. Anthropic’s own migration guide warns that visible responses run longer, and that lowering effort reduces thinking volume without reliably shortening the response: you need to prompt for conciseness explicitly.
Time is also a cost. At max effort, Artificial Analysis measures 52.8 output tokens per second and a time-to-first-token over a minute. If your workflow depends on fast iteration, that hurts.

The practical takeaway: on a well-scoped task, Opus 5 can end up cheaper than 4.8 because it solves things sooner with fewer round trips. On a vague prompt, it can cost you more because it thinks and writes more. Effort settings and prompting discipline are what decide. If the spend worries you, I already covered why Claude Code burns through your limits and how I track mine with PocketMeter.

Setup: what to change in Claude Code

First, the obvious but necessary bit: Opus 5 requires Claude Code v2.1.219 or later. Update:

claude update

From there, per the model configuration docs:

# Switch in the current session
/model opus        # the "opus" alias already resolves to Opus 5 on the Anthropic API

# Or at launch
claude --model opus

If you want to pin the exact version so a future alias change doesn’t move you without warning, use the full model ID in your settings or via environment variable:

{
  "model": "claude-opus-5"
}

Watch out for one detail: if you’re on Bedrock, Google Cloud or Foundry, the opus alias resolves to Opus 5 on Bedrock and Google Cloud, but on Microsoft Foundry it still resolves to Opus 4.6. There you’ll need to specify the full deployment name or set ANTHROPIC_DEFAULT_OPUS_MODEL.

On effort: the default is high, and levels run from low to max. Change it with /effort in-session or --effort at launch. My recommendation after reading the guide: keep high for daily work and only go up to xhigh or max for tasks where capability matters more than the bill. If you run at xhigh or max, raise max_tokens to 64k as a starting point, because the model needs room to think.

And if you pay per use in Claude Code, Fast mode toggles with /fast and costs double, billed through usage credits. Handy when you know exactly what you want and just need it written fast.

Setup: what to change in the API

If your code is on Opus 4.8, swapping the ID is trivial, but there are two breaking changes that will blow up on you if you skip them:

# This worked on Opus 4.8 and returns a 400 on Opus 5:
client.messages.create(
    model="claude-opus-5",
    max_tokens=16000,
    thinking={"type": "disabled"},
    output_config={"effort": "xhigh"},  # thinking disabled + xhigh/max = 400
    messages=[{"role": "user", "content": "..."}],
)

Disabling thinking is only allowed at effort high or below. And the other breakage: requests without a thinking field now think, so revisit your max_tokens.

Two more things that break or disappear: the web fetch tool is not available on Opus 5, and Priority Tier is not supported. If your organization has a Priority Tier commitment, that traffic stays on Opus 4.8, which keeps it.

And one improvement: the prompt caching minimum drops from 1,024 to 512 tokens, so prompts that were too short to cache before now create cache entries, with no code changes.

If you don’t want to migrate by hand, Anthropic ships a skill in Claude Code that automates it:

/claude-api migrate this project to claude-opus-5

It swaps the ID, applies the breaking parameter changes and leaves you a checklist of what to review manually.

When NOT to migrate to Claude Opus 5

This is the part that actually saves you money:

Your workload uses the web fetch tool. It doesn’t exist on Opus 5. Stay on 4.8 or redesign that piece.
You have a Priority Tier commitment. Opus 5 doesn’t support it.
You iterate fast on simple tasks. For renaming things, writing simple tests or mechanical changes, Sonnet 5 at $3/$15 (with introductory pricing of $2/$10 through August 31, 2026) or Haiku 4.5 at $1/$5 are faster and far cheaper. That’s exactly why the opusplan alias exists: Opus for planning, Sonnet for execution.
Your prompts are full of “verify your work” instructions. Opus 5 verifies on its own; the migration guide explicitly tells you to remove those instructions because they cause over-verification. If you’re not going to rework your prompts, it’ll waste tokens. And if your problem is the AI messing things up, a well-written AGENTS.MD pays off more than repeating instructions in every prompt.
You work in offensive security. Opus 5 ships with safety classifiers that block pentesting and exploit generation, and flagged requests get re-routed to Opus 4.8. You’ll spend your day bouncing off fallbacks.

My verdict

If you code with Claude daily, migrating is almost a no-brainer: same price, better model, and in Claude Code it’s a /model opus after updating. But migrate understanding that default thinking and verbosity change your bill even though the per-token price doesn’t. Pin the exact model, revisit your max_tokens, drop the effort on simple tasks, and leave Sonnet for mechanical work.

The model is good. The bill is on you.

FAQ

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens, exactly the same as Opus 4.8. Fast mode costs double. What changes your bill isn’t the price but how many tokens it burns: thinking is on by default and responses run longer.

Is it worth migrating from Opus 4.8 to Opus 5?

For agentic coding, yes: it’s number 1 on the Artificial Analysis Intelligence Index at the same price. The clear exceptions are if you use the web fetch tool, have a Priority Tier commitment or work in offensive security - in those cases, staying on Opus 4.8 makes more sense.

Can you disable thinking in Claude Opus 5?

Yes, but only at effort high or below: combining disabled thinking with effort xhigh or max returns a 400 error on the API. Also note that requests without a thinking field now think by default, so revisit your max_tokens when migrating.

What do you think?

Leave your opinion, question or suggestion. Comments are synced with GitHub Discussions .