Claude Fable 5 — What It Is and Why It Matters
Text: On June 9, 2026, Anthropic released a new model, Claude Fable 5 — and it's the most notable release in the Claude lineup over the past year. But to understand how Fable 5 differs from what you've used so far, you need to start with the previous flagship model — Claude Opus 4.8. Most Claude users have been working with it for the past few weeks, and it remains the baseline model for everyday tasks. Let's break it down in simple terms: what Opus 4.8 can do, what new features Fable 5 brings, the differences between them, and why June 22 is important.
At West Star Ltd, we use Claude models every day in all work processes — from development to analytics preparation — so we analyze not by press releases, but by what these models actually change in work.
CLAUDE OPUS 4.8 — WHAT IT IS AND WHAT IT CAN DO
Opus 4.8 was released on May 28, 2026, and until the release of Fable 5, it was Anthropic's most powerful model available to regular users. Anthropic described it modestly as a "modest but tangible improvement" over its predecessor, Opus 4.7. This is an honest statement: Opus 4.8 is not a revolution, but a refined version of an already strong model.
The main features of Opus 4.8, explained in simple terms.
First — honesty. This is perhaps the most useful improvement. Opus 4.8 is about four times less likely than its predecessor to miss a defect in the code without warning. In practice, this means: when you ask the model to write or check code, it warns about potential problems instead of silently issuing code with a hole. For example, if you ask to create an API route without input validation, Opus 4.8 will add a warning that it's unsafe and validation is needed in production. This applies not only to code — the model, in general, has become better at communicating its doubts and uncertainties, rather than presenting guesses as facts.
Second — better "judgment" in working with agents. In Claude Code, the model asks the right clarifying questions, catches its own mistakes, objects when a plan looks unsuccessful, and doesn't rush to make big changes until it understands the task. This is the behavior of an experienced engineer, not an executor who does what is asked.
Third — parallel sub-agents and dynamic workflows. Claude Code on Opus 4.8 can plan a large task and launch hundreds of parallel "sub-agents" in one session — for example, to carry out a migration through hundreds of thousands of lines of code, from planning to merging. This is no longer an assistant for one function, but a tool for tasks on the scale of an entire project.
Fourth — effort control. Next to the model selection, there is a switch for how diligently Claude works on the answer. A higher level means deeper thinking and better results, but more limit consumption. Lower means quick answers, more economical. By default, it's set to "high," for complex tasks "extra" or "max" is recommended.
Fifth — speed and cost. Fast Mode works 2.5 times faster and now costs three times less than previous models. The API price hasn't changed from Opus 4.7 — $5 per million input tokens and $25 per million output.
By the numbers: Opus 4.8 scores 88.6% on the SWE-bench Verified test and 69.2% on the more complex SWE-bench Pro, outperforming GPT-5.5 and Gemini 3.1 Pro in most categories. It has a context window of 1 million tokens (about 2000 pages of text). For everyday work — development, analytics, translations, document preparation — this is a very strong and reliable model.
CLAUDE FABLE 5 — WHAT'S NEW
And now the main event. On June 9, 2026, Anthropic released Claude Fable 5 — the first publicly available model of a new class that the company calls Mythos-class. To simplify, these are models that stand a step above the Opus class in capabilities.
Here you need to understand the structure. Anthropic has a closed super-powerful model, Claude Mythos (formerly known as Mythos Preview), which is not issued to the general public for cybersecurity reasons — it is so good at finding and exploiting vulnerabilities that it is used by a limited circle of organizations within the Project Glasswing program in collaboration with the US government. Fable 5 is essentially the same model but with additional protective mechanisms that make it safe for public release. Simultaneously with Fable 5, Mythos 5 was also released — the same engine with partially lifted restrictions, still only for a narrow circle of specialists.
What Fable 5 can actually do, according to Anthropic and early testers.
The main feature — the longer and more complex the task, the greater the advantage of Fable 5 over previous models. On short requests, the difference is not so noticeable, but on large projects, it becomes huge. A specific example: the company Stripe used Fable 5 to migrate across the entire codebase of 50 million lines of code in one day. The same work manually would take a whole team of engineers more than two months.
In benchmarks, Fable 5 noticeably outperforms everyone. On SWE-bench Pro (a complex programming test), it scores 80.3% against 69.2% for Opus 4.8, 58.6% for GPT-5.5, and 54.2% for Gemini 3.1 Pro. On the computer usage test (OSWorld-Verified) — 85%. On real expert work tasks (GDPval-AA), Fable 5 surpassed Opus 4.8. It is the first model to exceed the 90% mark on the analytical benchmark Hex.
A couple more striking examples of capabilities. Fable 5 completed the video game Pokémon FireRed using only raw screen screenshots — without maps, hints, and auxiliary tools. Previous Claude models required complex auxiliary setup for the same game. And file memory improved Fable 5's results on the game Slay the Spire three times more than the same conditions improved Opus 4.8.
Technically: a context window of 1 million tokens, generation of up to 128 thousand tokens in one response, support for text, images, and PDFs as input. Model identifier in the API — claude-fable-5. The model is available through Claude API, Claude Platform, Claude Code, as well as on AWS Bedrock, Google Cloud, and Microsoft Foundry.
THE DIFFERENCE BETWEEN FABLE 5 AND OPUS 4.8 — SIMPLY
If you remove the technical details, the difference boils down to a few things.
Model class. Opus 4.8 is the pinnacle of the Opus class, an evolutionary improvement of a familiar line. Fable 5 is a higher-class model (Mythos), released to the public for the first time. It's not "Opus, but slightly better." It's the next step.
Strength on complex tasks. On simple everyday requests — writing a letter, explaining a concept, drafting a small script — the difference between them is small, and Opus 4.8 will do just fine. But the longer, more complex, and more "agent-like" the task (large code migrations, multi-hour autonomous work, projects on the scale of an entire system), the more Fable 5 pulls ahead.
Price. And here's the main practical difference. Opus 4.8 costs $5 per million input tokens and $25 per output. Fable 5 costs $10 for input and $50 for output — exactly twice as expensive as Opus, though still less than half the price of the closed Mythos Preview. Fable 5 is the most expensive of the major models on the global market.
Security and routing. Fable 5 has an interesting built-in protection: if you ask a sensitive question (e.g., about offensive cybersecurity or biology), the system automatically redirects the request to the less risky Opus 4.8 model. According to Anthropic, this triggers in less than 5% of sessions, so the average user hardly notices the difference, and potentially dangerous applications are filtered out. Before launch, the model underwent over 1000 hours of stress testing, and no universal ways to bypass the protection were found.
IMPORTANT — FABLE 5 IS FREE UNTIL JUNE 22, THEN FOR TOKENS
The most practical thing to know right now. From June 9 to June 22, 2026, inclusive, Fable 5 is available for free (no extra charges) on Claude Pro, Max, Team, and seat-based Enterprise subscriptions. This means that if you have any of these subscriptions, for two weeks you can use Anthropic's most powerful public model without paying anything extra beyond the subscription.
Starting June 23, 2026, the picture changes. On these same subscriptions, using Fable 5 will start consuming paid credits (usage credits) — meaning you'll have to pay separately for it, by tokens. And for those working directly through the API, payment by tokens starts from day one: $10 per million input and $50 per million output tokens.
Practical advice: if you want to understand whether Fable 5 is worth the money for your tasks, test it during these two weeks while it's free. Run your real work scenarios on it — the most complex, voluminous ones, where the regular model struggles. If you see that Fable 5 solves what Opus 4.8 handled worse, after June 23, you can consciously decide which tasks are worth paying extra for.
HOW TO CHOOSE — PRACTICAL CONCLUSION
For most everyday tasks, Opus 4.8 remains the right choice. It's fast, significantly cheaper, very reliable, and for letters, texts, analytics, regular development, its capabilities are more than enough. There's no need to pay twice as much for tasks that Opus already handles excellently.
Fable 5 makes sense where the task is truly complex and lengthy: large-scale code migrations, multi-hour autonomous agent work, projects where you need to keep a huge context in mind and not lose the thread over hundreds of steps. Here, the price difference pays off because Fable 5 does in a day what would otherwise take weeks.
A few honest caveats to keep in mind.
Some of the impressive figures in the system card refer to Mythos 5, not the public Fable 5. For example, record figures for offensive cybersecurity and biology were measured on the Mythos version with lifted restrictions. Fable 5 itself in lockdown mode does not perform these tasks — and this is done intentionally. Therefore, when comparing, it's important to see which version a specific figure refers to.
The price is really high. Fable 5 is the most expensive large model on the market. For a solo entrepreneur or a small company, this means that using it for everything is economically unreasonable. A sensible strategy is to conduct the main work on Opus 4.8 (or even cheaper models like Sonnet) and connect Fable 5 selectively, for specific heavy tasks.
The free period is marketing, not a gift forever. Two weeks of free access are made so that you get used to the model and then pay for it. This is normal, but it's worth entering this period with a clear head: test, not build work processes that will suddenly become expensive after June 23.
For us at West Star Ltd, the approach is simple: each model is a tool for its class of tasks. Opus 4.8 is the workhorse for every day. Fable 5 is heavy machinery for those rare tasks where maximum power is needed and where saving weeks of work justifies the higher price. The main thing is not to pay for power that is not needed for a specific task. And the next two weeks are a good opportunity to understand for free where exactly this power will really come in handy.