Claude Opus 5 Released: How It Compares to Opus 4.8 and Fable 5

On July 24, 2026, Anthropic released Claude Opus 5. It lands roughly two months after Opus 4.8, and the version number jumps a whole generation, from 4.8 to 5. Anthropic's own documentation calls it a step-change rather than an incremental improvement over Opus 4.8.

Here is the short version. Pricing does not move ($5 input / $25 output per million tokens), yet on agentic benchmarks it matches or beats the flagship Claude Fable 5 — which costs $10 / $50, so Opus 5 does that work at half the price. On the other hand, the API ships two breaking changes, and code that moves over untouched can hit 400 errors or truncated output. This article cross-checks the official announcement and documentation against several news reports, and lays out performance, pricing, and the migration traps from a practitioner's point of view.

1. Opus 5 in three lines

In one sentence: Opus 5 does close to Fable 5's job at Opus 4.8's price. Which also means the reverse — when you need raw maximum capability, Fable 5 and Mythos 5 still have their place (see the losses below).

2. Core specs and where you can use it

Start with the numbers. Here are the headline specs, from Anthropic's official model list.

Source: Anthropic, "Models overview" and "What's new in Claude Opus 5" (as of July 2026)

The line people skim past but feel in daily work is the knowledge cutoff moving forward. Opus 4.8 and Fable 5 stop at January 2026; Opus 5 reaches May 2026. That is several extra months of library versions and recent spec changes it can answer from plain knowledge. Even so, it still does not know "today" — never let it assert current events or fresh price changes without a web search.

Where you can run it

3. Pricing — unchanged, and half of Fable 5

The single most practical fact about this release is that capability went up and the price did not move. Anthropic's documentation states the pricing plainly: "$5 per million input tokens and $25 per million output tokens, unchanged from Claude Opus 4.8."

The point worth internalizing is that price per token and cost per task are different things. In its Opus 5 material, Anthropic leads with charts of cost per task, not unit price. On the computer-use benchmark OSWorld 2.0, for instance, the official announcement says it beats Fable 5's best score at roughly one-third of the cost. Half the unit price plus fewer wasted steps compounds into a much wider real-world gap.

4. Benchmarks in detail

Anthropic's announcement is built around charts rather than a table of numbers, so what the page actually states in prose is mostly relative: "more than double," "within 0.5%." This article therefore separates the comparisons Anthropic states in its own text from figures that multiple outlets report consistently.

4-1. Comparisons Anthropic states directly

4-2. Figures from press coverage

The numbers below come from the decoder (July 24, 2026) and tech-ish (same day), both reading them off Anthropic's charts. Only items where the two outlets agree are listed. Note that these figures are not written as numbers in the official page's text (marked with a yellow dot for "press-sourced").

The pattern is not subtle. The longer the model has to work on its own, the bigger the gap — terminal work, computer use, business automation, novel problems. Conversely, on single-patch style coding (DeepSWE) it loses to GPT-5.6 Sol. The same shape we mapped in our GPT-5.6 Sol comparison repeats itself one tier up.

5. Where it still loses

Launch write-ups tend to list only the wins, so let's deliberately do the opposite. Below are the weaknesses Anthropic itself acknowledges in the announcement, plus the numbers reporters flagged.

6. Two breaking changes (thinking on, effort limits)

If you call the API directly, this is the one section not to skip. The two items Anthropic lists as "behavior changes" can break things if your code moves over unmodified.

Change 1: thinking is on by default

On Opus 4.8, omitting thinking meant it ran without thinking. On Opus 5, omitting it means thinking runs — the same behavior as writing thinking: {"type": "adaptive"}.

# The same code that ran without thinking on Opus 4.8...
client.messages.create(
    model="claude-opus-5",
    max_tokens=1024,          # now caps thinking + answer together
    messages=[...],
)
# On Opus 5 it thinks first, so the answer can be cut off

max_tokens is a ceiling on thinking tokens plus answer text combined. Anything that sized max_tokens tightly around the expected answer length will now have that budget eaten by thinking and get truncated. When you migrate, you have to decide between explicitly turning thinking off and giving max_tokens more headroom.

Change 2: thinking can only be disabled at effort high or below

And that "turn it off" option now comes with a limit. Combining thinking: {"type": "disabled"} with effort xhigh or max returns a 400 error. Because the check happens per request, code that raises effort partway through fails in the worst way: it works at first, then suddenly starts erroring.

Anthropic goes further and documents what goes wrong when you do run with thinking off. With thinking disabled, Opus 5 sometimes writes tool calls out as body text — meaning the tool never runs while the turn still looks successful — and can leak internal XML tags into the output. If you are building agents, keeping thinking on and dialing effort down is the safer path.

7. Choosing among the five effort levels

Opus 5 supports all five levels, low / medium / high / xhigh / max, and the default in both the API and Claude Code is high. For the Claude Code side of the settings, see our guide to the effort setting.

8. Other new features

9. A different personality, and prompt fixes

You will probably notice this before you notice any benchmark. The documentation lists four behavior changes you will see without touching a line of code.

10. Who should migrate right now

Migration steps (for API users)

Summary

Claude Opus 5 is the kind of release users can simply enjoy: a clear step up at the same price. The gap is widest in agents that work autonomously for long stretches, and Anthropic leads with cost per task to make its case — frontier-class performance at half the price of Fable 5.

At the same time, the API has two changes you cannot ignore (thinking on by default; thinking can only be disabled at effort high or below). And the key to migrating well is that you remove prompt text rather than add it. Instructions written to patch an older model's weak spots are exactly what makes Opus 5 verbose and over-cautious.

Finally, remember that a benchmark number is only ever the value measured on that task. Opus 5 loses to GPT-5.6 Sol on DeepSWE, and there are areas where max effort is not the best setting. Measure it again on your own workload — in the end, that is the only number that settles anything.

FAQ

Note: figures in this article come from Anthropic's announcement "Introducing Claude Opus 5", the official documentation "What's new in Claude Opus 5" and "Models overview", plus reporting from the decoder and tech-ish (all July 24, 2026). Values marked 🟡 are not stated as numbers in the official text and were read off charts by reporters, so they carry less certainty. Specifications and pricing can change; confirm anything decisive against the official documentation.