Anthropic just shipped Claude Fable 5. What it means for a small business.

On June 9, 2026, Anthropic released Claude Fable 5, the most capable model it has ever made generally available, alongside a locked-down sibling called Claude Mythos 5. The benchmark claims are genuinely big. The price is also big. For a small business, both halves of that sentence matter, so here is the plain version.

What Fable 5 is

Fable 5 and Mythos 5 are the same underlying model. The difference is safety: Mythos 5 has safeguards removed in specific domains and is restricted to vetted partners, while Fable 5 is the version with those guardrails on, released for general use. (Anthropic’s naming nods to this: fabula, “that which is told,” next to the Greek mythos.)

The capability claims are not marketing fluff. Anthropic reports state-of-the-art results on nearly every benchmark it tested, and the proof points are concrete:

• Coding at scale. Stripe said Fable 5 compressed months of engineering into days, completing a 50-million-line Ruby migration in a single day that would have taken a team roughly two months by hand.
• Vision. It is positioned as the new best model for visual tasks, including rebuilding a web app’s source code from a screenshot.
• Finance and research. It topped Hebbia’s finance benchmark, and in blind tests scientists preferred the model’s molecular-biology hypotheses around 80% of the time over the previous Opus-class model.

It is available now on the Claude API and the major clouds (AWS Bedrock, Google Vertex AI, Microsoft Foundry), and in tools like GitHub Copilot. Paid Claude subscriptions include it at no extra cost through June 22, after which it draws on usage credits.

The number that matters for you: the price

Update, September 4, 2026: Fable 5.1 shipped on September 1 and the headline price did not move, but the effective price did. Anthropic released Claude Fable 5.1 alongside Mythos 5.1 with the same $10 and $50 per million token rates, a 1M-token context window, 128K maximum output, always-on adaptive thinking and a June 2026 knowledge cutoff. The change that matters is underneath: cache reads dropped 75 percent to $0.25 per million tokens. Anthropic estimates that works out to roughly 25 percent cheaper for typical workloads and up to 45 percent cheaper for highly agentic ones, where the same large context gets re-read on every step of a long task. That does not change the advice in this post, because the tasks that did not justify a frontier model at $50 per million output tokens still do not. It does change the arithmetic for one specific shape of work: long-running agentic jobs against a fixed, large codebase or document set, where caching does most of the lifting. If you priced that kind of job in June and walked away from it, it is worth re-pricing. For reference on where the frontier has settled, OpenAI launched GPT-6 Astra on September 3 at exactly the same $10 and $50 headline rate, which we covered in the post on AI-discovered vulnerabilities.

Fable 5 costs $10 per million input tokens and $50 per million output tokens. That is less than half the price of Anthropic’s earlier Mythos Preview, but it is still a frontier price: by Anthropic’s own comparison it runs about double the Opus 4.8 flagship, and it is many times the cost of the cheaper Sonnet and Haiku tiers that most everyday tasks run on.

This is the part a small business should sit with. The most powerful model is almost never the one you should be paying for. We made the same point when Google shipped its cheaper Gemini 3.5 Flash: for the work most businesses actually automate, a support assistant, a document summarizer, a quoting helper, a draft-writer, a mid-tier model clears the quality bar at a fraction of the cost. Paying $50 per million output tokens to answer a customer’s “what are your hours” is setting money on fire.

When Fable 5 is worth its price

There is real work that justifies a frontier model, it is just narrower than the launch-day excitement suggests:

For most small businesses, Fable 5 is the engine your vendor or developer might reach for on the hardest jobs, not a subscription you need to go buy. For a studio like ours it matters directly: a model that turns a two-month migration into a one-day job changes what we can take on for clients. For the business owner, the practical question is unchanged: is the task hard enough to need it?

The safety design is worth a note

Fable 5 ships with three classifiers covering cybersecurity, biology and chemistry, and model distillation. When one trips on a risky request, the prompt is quietly answered by Claude Opus 4.8 instead of Fable 5, rather than being refused outright. Anthropic says more than 95% of sessions never hit a fallback, and an external red-team found Fable 5 complied with zero harmful single-turn cyberattack requests across 30 jailbreak techniques.

The restricted sibling, Mythos 5, is the same model with those guardrails removed in specific domains, handed only to vetted cybersecurity defenders (under a program Anthropic calls Project Glasswing) and, later, approved biology researchers. Anthropic shipped all of this just days after publicly warning that AI capability is starting to outrun safety, which tells you how seriously it is taking the dual-use risk.

None of that changes your day-to-day, but it is a useful signal: the frontier is now capable enough that the company building it is gating its full power behind partner programs.

What to do about it

1. Don’t switch your stack to the newest model by reflex. The right model for a job is the cheapest one that clears your quality bar, and that is usually not the frontier tier.
2. Match the model to the task. Route routine work to cheap models and reserve the expensive one for the genuinely hard jobs. If you’re paying one flat premium rate for everything, you’re overpaying.
3. Watch the bill, not the benchmark. A model being “state-of-the-art” is irrelevant if a model costing a fraction as much already does your specific task well. We’ve argued this about self-hosting and cloud APIs alike: measure real usage, then choose.

Fable 5 is an impressive piece of work, and on the hardest problems it will earn its keep. For everything else, the discipline is the same as it was last week.

Update, August 2026: the cheaper rungs arrived, and they matter more

Two months on, the point above has a much better illustration than it did at launch, because Anthropic filled in the tiers below Fable 5. Claude Sonnet 5 shipped on June 30 and Claude Opus 5 on July 24, so the price ladder now looks like this, per million tokens:

One correction to that table since we first published it. Sonnet 5’s $2/$10 was announced at launch as introductory pricing, set to expire on August 31, 2026 and revert to $3/$15. On August 11, Anthropic made the introductory price permanent and confirmed the September 1 increase will not happen. So the row above is now the standing price rather than a promotional one, and anyone who budgeted for a 50 percent jump this autumn can take it back out.

Read the output column, because that is where the money actually goes. Fable 5 costs five times what Sonnet 5 does per million output tokens, and ten times Haiku. Anthropic’s own framing of Sonnet 5 is that it performs close to the previous Opus-class flagship at a fraction of the price, which is the whole argument of this post stated by the vendor.

Nothing here changes the advice. It sharpens it. Sonnet 5 is the model most business automation should be running on, Haiku handles the genuinely simple classification and routing work, and the two expensive tiers are for the jobs where a wrong answer costs more than the tokens. If your vendor has everything pointed at one premium model because it was the best one available in the spring, that is now a line item worth revisiting.

A second correction, as of September 4. The Fable 5 row is now Fable 5.1 and the per-token numbers in it are unchanged, so the ratios above still hold. What is not in the table is cache pricing, and that is where the September 1 release actually moved: cache reads are now $0.25 per million tokens, down 75 percent. Cached input is not a line most small businesses ever look at, and for a chatbot answering one question at a time it barely registers. For an agent that re-reads the same 200,000 tokens of context on every step of a two-hour job, it is most of the bill. If your vendor quotes you for that kind of work, ask specifically what they assumed for cache reads.

If you want help picking the right model for a specific use case, instead of defaulting to the most expensive one, that’s a conversation we have all the time.