Michael H

Posted on Oct 16 Originally published at agent5.news

Claude Fable 5.1 Explained: Price, Changes and the Ban Story

#ai #machinelearning #claudefable51 #anthropic

Originally published on agent5.news.

Claude Fable 5.1 is an AI model that Anthropic released on September 1, 2026. It follows Fable 5 at the same base prices, and Anthropic says it costs less to run on long tasks. Anthropic describes it as its most capable model open to all customers. This article explains what changed, what Fable 5.1 costs, and how it relates to Mythos 5.1 and Claude Opus 5.5. It also explains why a US government order took Fable 5 offline for 19 days in June 2026.

What is Claude Fable 5.1?

Claude Fable 5.1 is a large language model, which means software that reads and writes text and computer code. It sits in Anthropic's top tier, known as Mythos-class, above the company's Opus models.

Anthropic built it for long jobs: coding sessions that run for hours, research with many steps, and finished documents, spreadsheets and slide decks. It can keep 1 million tokens in view at once. A token is a small chunk of text, and Anthropic says 1 million tokens is roughly 555,000 words.

Fable 5.1 comes with built-in safeguards. Automated filters, known as classifiers, watch for requests about hacking and risky biology. When a filter flags a request in the Claude apps, an older and less capable Opus model answers instead.

Fable 5.1 is available in the Claude apps on paid plans. Developers can reach it through Anthropic's API, Amazon Bedrock, Google Cloud and Microsoft Foundry. An API is the connection that lets other software send requests to a model.

You may not need it. For most work, Anthropic's own documentation says to start with the cheaper Opus 5.5.

Fable 5.1 vs Fable 5: what changed?

The base prices stayed the same. What changed is the cost of reused input, the test scores, and how often the safety filters interrupt normal work.

Running cost: Reading stored input, known as a cache read, dropped from $1.00 to $0.25 per million tokens. That is a 75 percent cut.
Test scores: In Anthropic's own testing, Fable 5.1 scored 52.6 percent on Terminal-Bench-Science 0.1, a test of scientific research tasks, against 24.7 percent for Fable 5. On Terminal-Bench 4.0, a test of multi-step tasks in a text-only command window, it scored 55.8 percent against 42.0 percent. These figures come from Anthropic, not an independent lab, with a stated uncertainty of 3.5 to 4.5 points on the science test.
Fewer false alarms: Fable 5.1 may now be used to find security flaws in software, though not to write attacks that exploit them. Anthropic says people using its coding tool, Claude Code, can expect around 60 percent fewer interruptions per session. Its biology filters now trigger 85 percent less often on harmless basic biology and medical questions than at Fable 5's launch.
Data rules: For business customers, Fable models still come with a 30-day data retention rule for safety monitoring.
Watermark: Text from Fable 5.1 carries an invisible watermark, a hidden pattern in the wording that a detection tool can check. It follows a European Union code of practice on marking AI-generated content, covered in our explainer on how AI watermarks work.

Anthropic's documentation also lists trade-offs: fewer progress updates during long runs, more whole-file rewrites for small edits, and three changes that can break code written for Fable 5. Its launch post adds that, in testing, the model could still sometimes bypass approvals and automated checks.

Fable 5 remains available, and Anthropic says it will not retire it before June 9, 2027. For background, see our explainer of the original Fable 5.

Fable 5.1 pricing: what does it cost?

On Anthropic's API, Fable 5.1 costs $10 per million input tokens and $50 per million output tokens, the same as Fable 5. Input is what you send. Output is what the model writes back.

Anthropic's price list, per million tokens:

Input: $10
Output: $50
Cache read: $0.25 (Fable 5: $1.00)
Cache write: $12.50 for five minutes of storage, $20 for one hour
Batch jobs that can wait: $5 for input, $25 for output

A cache read happens when the model reuses material it has already processed, such as a long document or the earlier part of a conversation. AI agents, which are programs that let a model work through many steps with little supervision, reread the same material again and again.

That explains the claim that Fable 5.1 is up to 45 percent cheaper. That number is Anthropic's estimate of the total bill for tool-heavy agent work, where cache reads make up most of the cost. For typical workloads, the company estimates about 25 percent. Both figures are based on four weeks of real usage in August 2026, and both apply only where usage is billed by token. The prices for new input and output did not move.

In the Claude apps, Fable 5.1 is limited to paid plans:

Free plan: not available
Pro, and standard seats on Team plans: pay-as-you-go usage credits on top of the subscription, billed at standard API rates
Max, and premium seats on Team plans: included, for up to 50 percent of the weekly usage limit

Fable 5 took the same path. Pro plans included it during a promotion that ended on July 19, 2026, and it then moved to usage credits.

Our guide on why newer AI models sometimes cost more, not less shows how list prices and real bills can move in different directions.

What is Mythos 5.1 and who can use it?

Mythos 5.1 is the same model as Fable 5.1 with looser safeguards, and only vetted organizations can use it. It launched on the same day at the same price.

Fable is the version for general use. Mythos is for vetted security teams and life science researchers whose work Fable would block or reroute. One example is penetration testing, which means authorized attack tests on computer systems.

On Terminal-Bench 4.0, Anthropic reported 60.9 percent for Mythos 5.1 and 55.8 percent for Fable 5.1. The company says the difference reflects tasks where Fable's security filters stepped in.

Access runs through two verification programs, one for cybersecurity and one for life sciences. When Mythos 5.1 launched, Anthropic said only a set of US organizations had it. On October 6, 2026, the company opened an expanded Cyber Verification Program with three access tiers, and each tier includes Mythos 5.1. For the least restricted tier, Anthropic reviews every organization together with the US government.

Opus 5.5 vs Fable 5.1: which one should you pick?

For most work, Anthropic itself points people to Claude Opus 5.5, which costs less. Fable 5.1 is meant for the hardest and longest tasks.

Anthropic released Opus 5.5 on September 22, 2026, three weeks after Fable 5.1. It costs $4 per million input tokens and $20 per million output tokens. For input and output, Fable 5.1 costs 2.5 times as much.

Anthropic says Opus 5.5 "performs at the level of Claude Fable 5.1 on most work." In the company's published results, Opus 5.5 scored higher on several tests, for example 66.4 percent against 55.8 percent on Terminal-Bench 4.0. Anthropic adds that in its own use, the gap between the two is narrower than those scores suggest.

Should you care about Fable 5.1? A short guide based on Anthropic's documentation:

Most workloads: start with Opus 5.5. In the Claude apps, Opus is part of the Pro plan, while Fable costs extra there.
Agent sessions that run for hours and deep multistep research: consider Fable 5.1, or try it when Opus 5.5 still falls short at a higher effort setting. Effort is a control for how much the model thinks before it answers.
Strict privacy needs: Opus 5.5 is available with zero data retention, an arrangement in which Anthropic does not store prompts and answers. Fable 5.1 keeps data for 30 days by default.

Both models carry similar safeguards for cybersecurity and biology. Our model release news tracks the rest of the lineup, including Sonnet 5.5 and Haiku 5.5, which followed on September 28 and October 7, 2026.

Why was Fable 5 banned in June 2026?

Fable 5 was never banned outright, but for 19 days nobody could use it. On June 12, 2026, three days after launch, the US Commerce Department placed export controls on Fable 5 and Mythos 5. The order barred access by any foreign national, inside or outside the United States, including Anthropic's own foreign employees.

In this case, the controls meant a government license was required before the models could be exported. Under US rules, sharing controlled technology with a foreign person inside the United States also counts as an export. Anthropic said it had no reliable way to check nationality in real time, so it switched both models off for every customer. Its other models kept running.

The stated reason was national security. Anthropic said the letter, which arrived at 5:21 pm Eastern time, gave no specifics. According to Anthropic's later account, the trigger was a report by Amazon researchers. They had found a way to get around Fable 5's filters, a trick known as a jailbreak. With it, Fable 5 identified a number of software flaws and, in one case, wrote code showing how a flaw could be exploited.

The two sides describe the episode differently:

A senior administration official told Fox Business that Anthropic's "recklessness" in responding to the concerns led to the controls. The outlet reported that Amazon and five other companies had found a workaround that opened the full hacking abilities of the more advanced model.
Anthropic said the technique exposed nothing unique. In its tests, less capable models, including rival products, found the same flaws. A source close to the company told Fox Business that officials gave no details at first, only a request to pull the model within 90 minutes.

Our explainer on US AI export controls covers how these rules work in general.

How did the Fable 5 export controls end?

The Commerce Department withdrew the controls on June 30, 2026, and Fable 5 returned for users worldwide on July 1, 2026.

June 26, 2026: The government approved the return of Mythos 5 for a set of US organizations, about 100 according to Nextgov.
June 30, 2026: Commerce Secretary Howard Lutnick announced that the controls in the June 12 letter were withdrawn and that a license was no longer required.
July 1, 2026: Fable 5 came back in the Claude apps and on Anthropic's API.

Anthropic changed two things. First, it trained a new filter aimed at the reported trick. The company says the filter blocks that technique in over 99 percent of cases, at the cost of flagging more harmless coding and debugging requests. Second, it promised closer work with the government: early access to its most advanced future models for official testing, fast reports on new jailbreaks, and joint research.

Anthropic complied but objected. Its June 12 statement said the government should be able to block unsafe AI releases through a transparent process based on technical facts. This order, it said, did not meet that standard. On July 2, 2026, it published a draft scale for rating how serious a jailbreak is.

Two months later, Fable 5.1 arrived with the more precise filters described above.

What happens next

Private data storage for large customers: Anthropic says a system named Enterprise Frontier Safeguards will roll out in phases starting in fall 2026. It lets companies keep monitoring data in their own cloud accounts instead of on Anthropic's systems. Until it is ready, eligible customers can use Fable 5.1 with zero data retention.
Wider access to Mythos 5.1: The expanded Cyber Verification Program opened on October 6, 2026, and members of its entry tier have until December 15, 2026 to adopt stronger login security. At launch, Mythos 5.1 was limited to US organizations. Anthropic said it was working with the US government to add international partners and has not named a date.
Government review of the next top model: An executive order signed on June 2, 2026 gave agencies 60 days to design a voluntary framework for the most advanced models. Under it, developers would give the government access for up to 30 days before a model goes to other trusted partners. Anthropic has promised such early access, so the next Mythos-class launch will show how the review works in practice.

DEV Community