Fable 5 vs Opus 4.8: which one to use (and the 10 tasks where the difference shows)

Anthropic shipped Claude Fable 5 and everyone is asking the same question: is it worth swapping Opus 4.8 for Fable 5?

Short answer, before the list: it depends on the task, and the difference isn't linear. In four categories Fable solves what 4.8 couldn't. In plenty of others, switching just burns budget.

Quick verdict: Fable 5 or Opus 4.8?

Rule of thumb: Fable wins where the task is long, visual or dense. Opus 4.8 still makes sense where the task is short and the volume is high.

Below are the ten concrete situations where the difference shows. Each one is the evidence behind the table.

TL;DR

• What it is: Claude Fable 5, Anthropic's Mythos-class model, released for general use through the API and paid plans.
• Models/Stack: Fable 5 (with safeguards) and Mythos 5 (same model, reduced safeguards, restricted to cyber defense partners and biology research).
• Cost/Access: $10 per million input tokens, $50 per million output tokens. Less than half the price of Mythos Preview.
• Useful link: Anthropic's official announcement
• Update: since this post was published, Anthropic released Claude Opus 5. If you're choosing between newer generations, the Fable 5 vs Opus 4.8 comparison below still holds up as a read on where the jump happens, but check current pricing before you decide.

This comparison covers the previous generation (Fable 5 vs Opus 4.8, June 2026). If you're picking a model today, the current decision is between Fable 5.1, Opus 5 and Sonnet 5, with benchmarks, price per task and a decision tree in Claude Fable 5.1 vs Opus 5: which one to use and what it costs.

The context: why does Claude Fable 5 change the game?

Before jumping into the list, understand what changed underneath.

Fable 5 and Mythos 5 are the same model under the hood. The difference is the safeguard layer: Fable is the one you use day to day, with classifiers covering cybersecurity, bio/chem and distillation. Mythos is the raw version, released through Project Glasswing to people working on cyber defense and biological research. For 95% of Fable sessions, none of that shows up. There's no fallback to Opus 4.8, it's Fable handling it directly.

And the trick here is genuinely technical. According to Anthropic, Fable 5 has state-of-the-art performance on "nearly all benchmarks tested" and beats every model they had made available before, including Claude Opus 4.8, which we covered recently and which was the ceiling for anyone shipping AI to production. Big gains in reasoning over documents, reading tables and charts, and, what matters most to people who write code, software engineering.

But switching models because a better one came out is a reaction, not engineering. People who decide well have an eval for their own domain and measure before migrating. That's the difference between using AI and building with AI, which is what we practice every week, with running code, in the Clã Beer and Code.

Next up, the ten things this unlocks in practice.

1. Migrating a giant codebase in a day, not two months

The case that opens the announcement is Stripe's: Fable 5 ran a migration on a 50-million-line Ruby codebase in a single day. The same work would take a team two months.

This isn't "smarter autocomplete." It's the model holding the context of an entire codebase, understanding dependencies and applying a consistent transformation at scale. With Opus 4.8, a large migration was a puzzle of context windows, manual chunking and exhaustive review to make sure step 4,000 didn't contradict step 12.

2. Rebuilding an app's code from screenshots

Fable 5's vision capability is state of the art, and here's the wildest example: it rebuilt the source code of a web application from screenshots alone.

Think about what that means for your workflow. Got a Figma mock, a screenshot of a competitor, a photo of a legacy screen with no code behind it? The model can now go from pixel to component. Opus 4.8 helped with an approximate layout; faithfully reconstructing structure from an image was hallucination territory.

3. Reading exact numbers from charts and scientific figures

One of the most underrated gains is in document interpretation. Fable 5 extracts precise numbers from scientific figures. It doesn't estimate "around 40%," it reads the right value off the chart.

For anyone building a product on top of reports, papers, dashboards or scanned spreadsheets, that's the difference between a reliable pipeline and one you can't leave unattended. Reading tables and charts was exactly one of the spots where Opus 4.8 slipped.

4. Keeping the thread across millions of tokens

Fable 5 sustains focus on a single task across millions of tokens without getting lost. That's long-context memory actually working, not just a big window on paper.

In practice: agents that run for hours, refactors that cut across dozens of files, analysis of huge logs. The classic long-context problem was never fitting the text in. It was remembering the beginning by the time you reached the end. Opus 4.8 would start to skid. Fable holds the reasoning from start to finish.

5. Performing 3x better on long-running agentic tasks

There's a hard number for this one. Running Slay the Spire with persistent memory, Fable 5 performed 3x better than Opus 4.8.

Why does a game matter to you? Because it's an honest proxy for a long agentic task: chained decisions, evolving state, long-term consequences. It's the same shape as an agent operating a system for hours while making dependent decisions. Tripling performance here is a direct signal of a more reliable agent in production.

6. Playing (and finishing) on vision alone, with no helper tools

Fable 5 completed Pokémon FireRed using vision only, with no supporting tools at all. No game API, no memory reading. Just looking at the screen and acting.

Translated to the real world: automating interfaces you don't control. Legacy systems with no API, third-party screens, flows that only exist in the GUI. An agent that sees and acts on pixels opens up an entire category of automation that, with Opus 4.8, depended on duct-taped OCR plus brittle scripting.

7. Reflecting on and validating its own work

At maximum effort, Fable 5 reflects on and validates its own work before delivering. It reviews what it produced instead of spitting out the first plausible answer.

That goes after the number one pain of AI in production: trusting the output. An Anthropic customer described the result on legal redlines as "noticeably different." For devs, it's less "generate and pray," more "generate, check and justify." Opus 4.8 did this in a shallow way; here it's become default behavior at high effort.

8. Senior-level reasoning over dense financial documents

On the Hebbia Finance Benchmark, Fable 5 posted the highest score of any model on senior-level reasoning.

This isn't pulling a number out of an invoice. It's the analysis an experienced analyst would do: cross-referencing tables, understanding business context, drawing a defensible conclusion from a dense document. This kind of document-based reasoning saw a substantial gain over Opus 4.8, and it's exactly what you need to build a product on real financial data.

9. Delivering frontier code at medium effort (and spending fewer tokens)

On FrontierCode, Cognition's benchmark, Fable 5 had the best score among frontier models even while running at medium effort.

This is where cost engineering lives. You don't need to max out effort, and your token budget, to get top-tier code. The model delivers state of the art with less computational effort, which changes the math for any product that runs generated code at scale. At $10 per million input and $50 per million output, medium effort beating frontier is money in your pocket.

10. Generating genuinely new scientific hypotheses

This is the biggest outlier. Fable 5 is the first model to consistently produce new scientific hypotheses. In blind comparisons, scientists preferred the model's hypotheses about 80% of the time. One of them, a previously unknown mechanism for an E. coli protein, was corroborated in a study.

For devs, the message generalizes beyond biology: the model went from "synthesizes what exists" to "proposes what hasn't been written yet." In system design, debugging a bizarre bug, exploring architecture, that's a collaborator that suggests a path, not one that only follows orders. Opus 4.8 was great at retrieving and combining; creating a genuinely new hypothesis was the limit.

When Opus 4.8 is still the right choice

The list above is long, but it doesn't say "switch everything." The cases where staying on 4.8 is the correct decision:

• High volume of short calls. $50 per million output tokens is unforgiving at scale. Classification, routing, simple extraction: the quality gain doesn't cover the bill.
• A stable flow that already passes your eval. A model that solves the problem doesn't get replaced by a model that solves a problem you don't have better.
• Tight latency. Maximum effort with self-validation is slower. In an interactive interface, that shows.
• Cases that brush up against safeguards. Fable routes certain requests to Opus 4.8. If your domain is offensive cyber or dual-use bio/chem, expect blocks.

Limitations and things to watch

It's not all rocket ships. Before you throw Fable 5 into production:

• Safeguards and fallback. The safeguards trigger in fewer than 5% of sessions. That's small, but it exists.
• Mythos 5 isn't for you (yet). The version without safeguards is restricted to Glasswing partners and biology researchers. What you use is Fable.
• Output cost. $50 per million output tokens isn't trivial at volume. Maximum effort with self-validation consumes more. Measure before you scale.
• A benchmark isn't production. 3x on a game and the top of FrontierCode are great signals, but your domain has its own traps. Evaluate on your own data.

Quick FAQ

Fable 5 or Opus 4.8: which is better? Fable 5 is superior on practically every benchmark, but "better" isn't the useful question. For long, visual or dense-document tasks, Fable solves what 4.8 couldn't. For short calls at high volume, 4.8 still makes more sense on the math. See the verdict table at the top.

What's the difference between Fable 5 and Mythos 5? Same model under the hood. Fable 5 has safeguards and is available to everyone through the API and paid plans. Mythos 5 has reduced safeguards and is restricted to cyber defense partners (through Project Glasswing) and biology research.

How much does it cost? $10 per million input tokens and $50 per million output tokens, less than half the price of Mythos Preview.

Do I need to move everything from Opus 4.8 to Fable 5? No. For long tasks, vision and reasoning over documents, the gain is clear. For cheap, simple flows that already run well, measure the cost-benefit before migrating blind.

Does it work through the API? Yes. Fable 5 is available through the API and on Anthropic's paid plans.

Conclusion

Claude Fable 5 isn't "Opus 4.8, slightly better." It's a model that unlocks entire categories of work (migration at scale, vision-driven automation, long-running agents, senior-level reasoning over documents) that used to be a hack or flat-out impossible.

But that doesn't turn "switching models" into a strategy. The leap isn't having access to Fable 5: it's knowing when it's worth the $50 per million and when Opus 4.8 delivers the same result for less. You decide that with an eval on your own domain, not with a vendor's benchmark table.

The tool evolved again. The question that remains is the usual one: are you going to use AI, or are you going to build with it?