Claude Opus 5 vs Fable 5 vs Sonnet 5 vs Opus 4.8: The Benchmark Comparison

A BitsMinds analysis. Claude Opus 5 landed on July 24, and it arrived into an unusually crowded family: it replaces Opus 4.8, sits below Fable 5 in Anthropic's own hierarchy, and costs more than Sonnet 5. So which one should you actually be calling? We pulled every published number and laid them side by side.

How we compared them — and what's missing

Two caveats matter more here than in a typical shootout, and they shape everything below.

First, Anthropic published its Opus 5 charts as images, not tables. The launch post states results in relative terms ("more than doubles Opus 4.8's performance," "within 0.5% of Fable 5's peak") rather than as a numeric grid. Where exact figures exist below, they come from Anthropic's own charts as transcribed by third parties, from the ARC Prize Foundation, or from public leaderboards — each labeled by source.

Second, these four models were never benchmarked against each other on one shared suite. Sonnet 5 launched June 30 against Terminal-Bench 2.1, SWE-bench Pro and OSWorld-Verified. Opus 5 launched July 24 against Frontier-Bench, CursorBench 3.2 and OSWorld 2.0 — different evals, different versions. The one model present in both sets is Opus 4.8, so we use it as the bridge: the honest way to compare Opus 5 and Sonnet 5 today is through how each one measures against the model they both replace or sit beside.

The spec sheet

Start with what isn't in dispute. These figures come straight from Anthropic's model documentation.

Spec	Claude Opus 5	Claude Fable 5	Claude Sonnet 5	Opus 4.8 (legacy)
API model ID	claude-opus-5	claude-fable-5	claude-sonnet-5	claude-opus-4-8
Price / MTok (in / out)	$5 / $25	$10 / $50	$3 / $15	$5 / $25
Context window	1M tokens	1M tokens	1M tokens	1M tokens
Max output	128k (300k via Batch beta)	128k	128k	128k
Comparative latency	Moderate	Slower	Fast	Moderate
Effort levels	All five (incl. xhigh)	All five	All five	All five
Reliable knowledge cutoff	May 2026	Jan 2026	Jan 2026	Jan 2026
Status	Current · default on Max	Current · top tier	Current	Legacy

Two things jump out. Every current model now shares the same 1M-token context — capacity has stopped being a differentiator across the lineup. And Opus 5 has the freshest knowledge cutoff in the family by four months, including a fresher one than the nominally more capable Fable 5. If your work touches anything that happened in early 2026, that gap is more practically useful than a benchmark point.

One pricing footnote: Sonnet 5's $3/$15 is its standard rate, but introductory pricing of $2/$10 runs through August 31, 2026.

Opus 5 vs Opus 4.8: the straight upgrade

This is the cleanest comparison of the four, because the price is identical — $5/$25 either way — so every gain is free. It is also the most lopsided.

Benchmark / measure	Claude Opus 5	Claude Opus 4.8	Source
Frontier-Bench v0.1 (coding)	43.3%	18.7%	Anthropic
ARC-AGI 3 (novel problem-solving)	30.2%	1.5%	ARC Prize Foundation
SWE-bench Verified	~97%	88.6%	vals.ai leaderboard
Organic chemistry	+10.2 pts	baseline	Anthropic
Protein tasks	+7.7 pts	baseline	Anthropic
Reasoning tokens (trading eval)	~1/7 as many	baseline	Anthropic
Latency (same eval)	under half	baseline	Anthropic
Token use, legal work at max effort	26% fewer	baseline	Anthropic
Misaligned-behavior score (lower is better)	2.30	higher	Anthropic audit
Price / MTok	$5 / $25	$5 / $25	Anthropic

The efficiency numbers deserve as much attention as the capability ones. Anthropic reports Opus 5 using roughly a seventh of the reasoning tokens at under half the latency of Opus 4.8 on a trading evaluation, and 26% fewer tokens on legal work at max effort. Since you pay per token, a model that is both better and more economical at identical list pricing makes staying on Opus 4.8 difficult to justify. Anthropic's own alignment audit also scores Opus 5 lower on misaligned behavior than Opus 4.8, Sonnet 5 and Fable 5 — the best of the family on that measure.

One caveat on the SWE-bench Verified figure: independent trackers listed Opus 5 between 96% and 97% within a day of launch, and at least one leaderboard carried a "last updated" date that predates the launch itself. Treat it as directionally right, not settled.

Opus 5 vs Fable 5: near-frontier at half the price

Here's the awkward one. Anthropic still describes Fable 5 as its "most capable widely released model" and prices it at double. But Opus 5 beats it on several published measures.

Benchmark / measure	Claude Opus 5	Claude Fable 5
Frontier-Bench v0.1 (coding)	43.3%	33.7%
SWE-bench Verified	~97%	95.0%
CursorBench 3.2 (Opus 5 at max effort)	within 0.5% of peak, half the cost per task	peak score
OSWorld 2.0 (computer use)	beats Fable 5's best, ~1/3 the cost	prior best
Price / MTok	$5 / $25	$10 / $50
Comparative latency	Moderate	Slower
Reliable knowledge cutoff	May 2026	Jan 2026
Anthropic's positioning	"complex agentic coding and enterprise work"	"most capable widely released model"

Read that last row against the rest and the tension is obvious. On the coding and computer-use evals Anthropic chose to publish, the cheaper model won or drew — while the docs still route you to Fable 5 for "the highest available capability." Both can be true: Fable 5 was built for long-running agentic work and retains headroom on tasks that don't appear on these charts, and Anthropic notably did not publish a Fable 5 number for ARC-AGI 3, where Opus 5 posted its most striking result.

The practical read: Fable 5 is no longer the automatic choice just because it costs more. For coding and computer-use work specifically, Opus 5 is the better-value default, and Fable 5 becomes the escalation path you reach for when your own evals show it earns the premium.

Opus 5 vs Sonnet 5: no shared benchmark, so mind the bridge

These two have never been run against each other on a common eval. What we can do is line each up against Opus 4.8, which both were measured against, and read the gap.

Benchmark	Claude Sonnet 5	Opus 4.8 (shared baseline)	Claude Opus 5
Terminal-Bench 2.1	80.4	74.6	not published
SWE-bench Pro	63.2	69.2	not published
Humanity's Last Exam (with tools)	57.4	57.9	not published
OSWorld-Verified	81.2	83.4	measured on OSWorld 2.0 instead
GDPval-AA v2	1,618	1,615	"state of the art," no figure given
Frontier-Bench v0.1	not measured	18.7%	43.3%
Price / MTok	$3 / $15	$5 / $25	$5 / $25
Comparative latency	Fast	Moderate	Moderate

Sonnet 5 traded roughly evenly with Opus 4.8 — ahead on Terminal-Bench and GDPval, behind on SWE-bench Pro and OSWorld — at 60% of the price and noticeably faster. Opus 5 then more than doubled Opus 4.8 on Frontier-Bench. Chain those together and Opus 5 is clearly the stronger model, but Sonnet 5 remains the value and latency play, and the gap on general reasoning is far narrower than the gap on hard agentic coding.

The verdict: which one should you use?

If you're…	Use	Why
Still on Opus 4.8	Switch to Opus 5	Same price, better on every published measure, and materially cheaper per task in tokens.
Doing agentic coding or computer use	Opus 5 (try xhigh)	Its strongest published category, at half Fable 5's rate.
Paying for Fable 5 out of habit	Re-run your evals	Opus 5 matched or beat it on the coding and computer-use charts Anthropic published.
On genuinely frontier, long-horizon agent work	Keep Fable 5	Still Anthropic's designated top tier, with headroom these charts don't capture.
Running high volume or latency-sensitive work	Sonnet 5	Fast tier, $3/$15 standard — and $2/$10 through Aug 31, 2026.
Building chat, classification, or subagents	Sonnet 5 at low/medium effort	Effort is the cost dial; frontier capability is wasted here.

For how Anthropic's lineup stacks up against the rest of the field, see our cross-lab benchmark comparison.