Opus 5 vs Fable 5: Fable's power at Opus 4.8 pricing

Anthropic just shipped Opus 5, the successor to Opus 4.8. We were waiting for it, because ever since Fable 5 came out, we all loved that model: it does incredible things. Its problem is that it costs an arm and a leg in tokens. It was even pulled for a while, came back, then had its usage cut in half. A model you want to keep using, but a limited one.

Opus 5 fixes that: you get almost Fable 5's power at Opus 4.8's cost.

Four frontier models in two months

Anthropic is shipping fast. The current lineup:

Sonnet 5 — for volume, when you have a lot to process and want speed.
Fable 5 — the absolute top, but it costs a fortune.
Mythos 5 — specialized in cybersecurity. We don't have access; it's reserved for certain companies. Note: Fable 5 is a derivative of Mythos.
Opus 5 — the brand-new default model, nearly equivalent to Fable 5, but not cybersecurity-oriented like Fable and Mythos.

Price is the real argument

Opus 5 is $5 input, $25 output per million tokens. Half of Fable 5, and the same price as Opus 4.8. Frontier intelligence at Opus pricing.

That's far from a detail: Fable is billed against your subscription and it burns fast. Run three requests with a few sub-agents and your credit is gone. Now you can work without torching your tokens on every task.

The benchmarks

On agentic coding in the terminal, we go from 21 (Opus 4.8) to 43 (Opus 5): a big efficiency jump. On knowledge work, a score of 1861.

On agentic coding — what matters most to us — Opus 5 is neck and neck with Fable 5: the gap comes down to almost nothing. On DeepSWE and FrontierCode, scores nearly equal to Fable. And versus Opus 4.8, you gain around 10 points (68 vs 59): a real gap.

On SWE-bench Verified (solving software engineering problems), Opus 5 scores 97, ahead of GPT-5.6 Sol, with Fable 5 at 95. Looking at resolution by task difficulty: on tasks under 15 minutes it's good at 98%, and it stays number one on 15-min-to-1h tasks as well as 1-to-4h tasks.

The takeaway: a model nearly equivalent to Fable 5, at Opus 4.8's price.

Not better everywhere

There are still cases where Fable and Mythos keep the edge:

On SWE-bench Pro, Fable 5 pulls ahead, but only slightly (80 vs 79).
For anything domain-specific — legal, medical, truly specialized fields — Fable 5 is above.
In cybersecurity, Mythos is better. Fable is based on Mythos but with more restrictions. Opus 5 has fewer guardrails than Fable 5.

Which model to use, and when

The goal is not to torch your tokens. Here's how I split it:

Sonnet 5 for volume: autocomplete, refactors, simple things. On my side, I have many technical sub-agents that create workspaces and open pull requests — they default to Sonnet 5 so they don't consume too much.
Opus 5 becomes my default model. It's a serious model that can do a huge amount: I push the bulk of the work onto it and run sub-agents with dynamic workflows.
Fable 5, whose usage stays restricted but is still included in the subscription, I keep for the truly specific things. In my case, not legal or medical, but rather advanced business strategy, large architectures from scratch, deep foundational thinking where you lay big foundations — which then get executed on Opus. In short: long tasks and deep thinking for Fable, execution on cheaper models.

In short

Opus 5 becomes the default: almost Fable 5's power, at Opus 4.8's price. For most people who were on Opus 4.8, the gain is substantial.

If you want more on agentic dev, I share it all regularly in my newsletter The Agentic Dev.

And the full video is here: watch on YouTube.