Anthropic's Claude Fable 5 Is Back After a 19-Day US Export Control Shutdown

Title: Anthropic's Claude Fable 5 Is Back After a 19-Day US Export Control Shutdown

Date: 17 July 2026

Type: Blog

Author: SAASiQ (contact@saasiq.ai)

Word count: 1490 words

Reading time: 6 min

Published: 17-07-2026

Tags: #AI #BigAI #Strategy #Architecture

Anthropic switched off Claude Fable 5, its most capable public model, for every customer in the world on 12 June, after the US Department of Commerce placed it under export controls. The controls were lifted on 30 June and the model came back on 1 July with an added safety classifier. For 19 days, anyone who had built on it had to fall back to something older.

What happened on 12 June

Anthropic released two models on 9 June. Fable 5 was for general use and shipped with strong safeguards. Mythos 5 had fewer safeguards and was restricted to partners in Project Glasswing, Anthropic's programme for defensive cybersecurity work.

Three days later, at 5:21pm Eastern on Friday 12 June, Anthropic received a directive from the Commerce Department suspending all access to both models by any foreign national, inside or outside the United States, including Anthropic's own foreign-national employees. The letter was signed by the Commerce Secretary, Howard Lutnick. The law firm Mayer Brown describes it as an 'is-informed letter' under the Export Control Reform Act of 2018, which imposes a licence requirement on the model, and notes that it treats remote access to a model through an API as a controlled 'release', something export rules had not previously been read to cover.

Anthropic had no reliable way to check a user's nationality in real time, so it disabled both models for all customers. Its other models, including Claude Opus 4.8, were not affected.

Why the government acted

The letter did not set out the national security concern. Anthropic said the government's understanding was that it had learned of a way to bypass, or 'jailbreak', Fable 5's safeguards. Fortune and The Wall Street Journal reported that the information came from Amazon's chief executive, Andy Jassy, who told administration officials that Amazon researchers had prompted the model into identifying software vulnerabilities, and in one case into producing code that showed how a vulnerability could be exploited. Amazon is one of Anthropic's largest investors, with at least $13 billion put in since 2023.

Anthropic said the technique amounted to asking the model to read a specific codebase and fix any flaws in it, and that the same capability was widely available from other models. It disagreed that the finding of "a narrow potential jailbreak" was grounds for recalling a model used by hundreds of millions of people. When it restored access on 1 July, it added that its own testing had found less capable models, among them Opus 4.8, GPT-5.5 and Kimi K2.7, could identify the same vulnerabilities.

What customers lost for 19 days

The suspension covered every route to the models. MarketScale reported that access was pulled at the same time from the direct Claude APIs, AWS Bedrock, Google Cloud, Microsoft Foundry, Snowflake and Box, and Claude Code and the Claude apps lost the model too. Buying through a cloud provider made no difference, because the control applied to the model itself, whichever channel it was bought through.

Fable 5 had been live for three days, and some customers had already built on it. VentureBeat cited a Stripe case study from the launch in which Fable 5 completed a codebase-wide migration across a 50-million-line Ruby infrastructure in a single day, work estimated to take a team more than two months by hand. It reported that workflows which had adopted Fable 5 or Mythos 5 were forced back onto older models such as Opus 4.8.

Because the directive covered foreign nationals anywhere, it reached UK and EU users directly. Fortune reported on 16 June that Al Carns MP had said British hospitals, companies and researchers had lost access. Agence Europe reported that the European Commission was assessing the consequences and saw the case as a reason to strengthen Europe's technological sovereignty, through instruments including the AI Act, the Cyber Resilience Act and the NIS2 Directive. The Commission had published a technological sovereignty package covering cloud, AI and semiconductors on 3 June.

How it came back

On 26 June the government approved Mythos 5 for a set of US organisations that operate and defend critical infrastructure. On 30 June Anthropic received notice that the controls on both models had been lifted, and on 1 July Fable 5 returned worldwide on the Claude Platform, Claude.ai, Claude Code and Claude Cowork. Perplexity reinstated it the same day as the 'orchestrator' model in its Computer product, the model that breaks a request into steps and hands them to other models and tools.

On Pro, Max, Team and some Enterprise plans, Fable 5 could use up to 50 per cent of weekly usage limits until 7 July, and after that it moved onto usage credits. Anthropic said it would re-enable the model on AWS, Google Cloud and Microsoft Foundry as soon as possible, and Amazon has since said it is back on Bedrock. Mythos 5 remains limited to approved US organisations.

Anthropic added a classifier aimed at the technique in the Amazon report, which it says blocks it in over 99 per cent of cases. It also committed to give designated government partners access to models before release where they advance national-security-relevant capabilities, to notify government counterparts when it finds significant jailbreaks or misuse, and to work with Amazon, Microsoft, Google and other Glasswing partners on a shared scale for rating how serious a jailbreak is.

OpenAI held its new models back as well

OpenAI announced its GPT-5.6 family, Sol, Terra and Luna, on 26 June, but at the US government's request opened it only to a small group of trusted partners whose names it had shared with the government. OpenAI said it did not want that kind of government access process to become the long-term default, and that it was going along with it as the best route to wider release.

The public release followed on 9 July, across ChatGPT, the API and Codex. OpenAI launched ChatGPT Work alongside it, an agent built to work through projects lasting several hours. API prices at launch, per million input and output tokens, were $5 and $30 for Sol, $2.50 and $15 for Terra, and $1 and $6 for Luna.

On 14 July the White House launched Gold Eagle, a clearinghouse run by the Treasury, the Cybersecurity and Infrastructure Security Agency and the Pentagon, working with industry on a voluntary basis. It takes in software vulnerabilities found using frontier AI, verifies them and coordinates fixes across US critical infrastructure. It was set up under Executive Order 14409, which President Trump signed on 2 June.

The cheaper models kept arriving

Anthropic released Claude Sonnet 5 on 30 June, while Fable 5 was still offline, at $2 per million input tokens and $10 per million output tokens until 31 August, rising to $3 and $15 after that, and made it the default model for Free and Pro users. TechCrunch reported a score of 63.2 per cent on agentic coding against 69.2 per cent for Opus 4.8, and Anthropic says it has a much lower ability to carry out dangerous cybersecurity tasks than its Opus models.

Google released Nano Banana 2 Lite, an image model priced at $0.034 per image at 1K resolution, and Gemini Omni Flash, a video model in public preview at $0.10 per second of video, on the same day. On 1 July xAI opened a beta of its Grok Voice Agent Builder, which builds phone agents without code, at $0.05 per minute plus $0.01 per minute for telephony.

Our view, from the finance and HR systems we work on, is that most of the AI work inside them is extraction, classification, reconciliation and drafting, and models priced like Sonnet 5 handle that. None of the models in this section was caught by either government action.

What buyers can take from it

The fallback VentureBeat described during the 19 days was Opus 4.8, the previous Anthropic model, which the directive did not touch. Moving work onto an earlier model at short notice only goes smoothly if the fallback has been tested on the organisation's own tasks, with its prompts and evaluations kept current, and if switching can be done by changing configuration, without rebuilding anything.

OpenAI's newest models were held back for two weeks at the government's request, so a second provider's frontier model was not a sure fallback either. In practice the fallback for a frontier model is usually the previous generation, from either vendor. Both government actions left those models alone.

Anthropic's statement on 12 June gave customers no instructions beyond saying it was working to restore access. Buyers can ask their AI vendors, and the cloud providers that resell them, how customers are told when a model is suspended for legal or regulatory reasons, and whether fees are credited for the time it is unavailable.

Fable 5 is available again on Anthropic's own services and on Bedrock, Mythos 5 remains limited to approved US organisations, and OpenAI's GPT-5.6 models have been on general release since 9 July.

SAASiQ - Intelligent Solutions for SaaS ©