Amazon CEO Andy Jassy may have been the source of security concerns that led Anthropic to cut off worldwide access to two models on Friday.

The Wall Street Journal reports that Jassy told Treasury Secretary Scott Bessent and other government officials that Amazon researchers used Anthropic’s Claude Fable 5 to obtain information that could be used in cyberattacks. The government subsequently imposed an export control ban on the Fable 5 and Mythos 5 models.

An Amazon spokesperson said in a statement that while it’s “not uncommon for governments to seek our counsel on potential security risks,” the company does not “share the details of those discussions.”

The spokesperson also pointed to an update stating that AWS has been affected by the model cut off.

The Information and Reuters similarly reported that Amazon (a major Anthropic investor) had communicated concerns about the security of Anthropic’s models.

David Sacks, Trump’s former AI czar who now co-chairs the President’s Council of Advisors on Science and Technology, offered his own account of the discussions, claiming that “a highly credible trusted partner of both Anthropic and the USG […] came forward with a jailbreak.”

Sacks added, “The Admin asked [Anthropic CEO Dario Amodei] to fix the jailbreak or de-deploy the model. Dario refused.”

Anthropic said in a blog post that the capabilities apparently causing government concern are already available in other publicly accessible models.

This post has been updated with a statement from an Amazon spokesperson.

Anthropic’s Fable 5 spent three days as the most capable AI model ever released to the public. It topped the Chatbot Arena leaderboard, crushed OpenAI’s GPT 5.5 on coding benchmarks by double-digit margins, and gave paying subscribers access to Mythos-class reasoning for the first time. Then, on June 12, the US government ordered Anthropic to shut it down.

The result is a strange moment in AI. The model that demonstrably outperforms everything else on the market is the one you cannot use. GPT 5.5, which OpenAI launched in late April under the internal codename “Spud,” is now the strongest model available to developers and consumers, not because it improved but because its only real competitor was removed.

The benchmark gap between the two is not close. On SWE-Bench Pro, which measures a model’s ability to resolve real software engineering issues across open-source codebases, Fable 5 scored 80.3% to GPT 5.5’s 58.6%, a 22-point difference. On SWE-Bench Verified, a curated subset of the same benchmark, Fable 5 reached 95.0%.

The coding benchmarks tell a similar story. Fable 5 leads the Code Arena by 98 Elo points, scoring 1,665 to GPT 5.5’s 1,501. On FrontierCode Diamond, a benchmark designed to test the most difficult programming tasks, Fable 5 scored 29.3% while GPT 5.5 managed 5.7%, and on the broader Chatbot Arena leaderboard Fable 5 sits at number one with GPT 5.5 in fourth.

GPT 5.5 does have one area of strength. On Terminal-Bench 2.0, which evaluates interactive terminal-based coding tasks rather than codebase-level issue resolution, GPT 5.5 scored 82.7% compared to Fable 5’s approximately 88.0%. The gap is narrower there, and the benchmark tests a different skill, executing commands and debugging in real time rather than reading and patching large repositories.

Pricing also favours OpenAI. GPT 5.5 costs $5 per million input tokens and $30 per million output tokens, half the price of Fable 5’s $10 and $50 respectively. For developers running high-volume applications where the performance difference is less critical than cost, GPT 5.5 is the more practical choice even when both models are available.

Fable 5 launched on June 9 as Anthropic’s first Mythos-class model made available to the general public. It offered a one-million-token context window and 128,000 output tokens. Anthropic made it available at no extra cost to Pro, Max, Team, and Enterprise subscribers until June 22, a promotional window that the government directive cut short after just three days.

The shutdown came via an export control directive issued on June 12. The government cited a jailbreak vulnerability as the reason for pulling both Fable 5 and the broader Mythos 5 model family. Anthropic has disputed the severity of the finding, saying the vulnerabilities identified are minor, publicly known, and achievable by GPT 5.5 without any bypass techniques, while reports indicate that Amazon CEO Andy Jassy played a role in triggering the government’s review.

The practical consequence is that developers and researchers who were evaluating Fable 5 for production use have had to revert to GPT 5.5 or Anthropic’s earlier Opus models. For coding-heavy workflows, the downgrade is significant. The 22-point gap on SWE-Bench Pro represents the difference between a model that can resolve four out of five real-world software issues and one that handles roughly three out of five.

Whether Fable 5 returns depends on Anthropic’s negotiations with the government over the export control classification. The company has publicly argued that the directive is disproportionate and that the cited vulnerabilities do not justify pulling the model entirely. Until that dispute is resolved, GPT 5.5 holds the top spot by default, the best model available not because it is the best model that exists.

TL;DR

Polish PM Donald Tusk announced a “sovereignty test” for significant government technology purchases and annual IT independence reports, warning that Poland’s dependency on foreign digital infrastructure demands urgent policy action.

Polish Prime Minister Donald Tusk has announced that Poland will introduce a “sovereignty test” for significant government purchases of technology solutions, warning that the country’s dependency on foreign digital infrastructure has reached a scale that demands a policy response. Speaking at the European Financial Congress in Sopot on Tuesday, Tusk said Poland will also publish annual reports documenting its progress toward IT independence, creating a public accountability mechanism for a priority he described as existential.

“At this point, the scale of this dependency, and I’m referring here to the relationship between the state and the digital sphere, has reached such proportions that it must prompt serious economic, institutional, and organisational decisions,” Tusk said. The statement places Poland among a growing number of EU member states that are translating tech sovereignty rhetoric into concrete procurement policy.

What the sovereignty test means

The details of the test have not been fully disclosed, but the framing suggests Poland will evaluate whether major government technology contracts create strategic dependencies on foreign providers, particularly in AI infrastructure, cloud computing, and telecommunications. Europe’s approach to technology governance has historically focused on regulation rather than procurement, but the AI era is forcing governments to consider who owns the infrastructure their citizens and institutions rely on, not just how it is regulated.

Poland has already taken steps in this direction. Earlier this year, the government banned Chinese-made cars from entering military facilities, citing infrastructure security concerns. The country is also increasing the share of domestic companies in public procurements, a policy that parallels similar moves across the EU as member states grapple with dependency on American cloud providers and Chinese hardware.

The European context

Tusk’s announcement arrives as EU nations struggle to balance trade relationships with technology security. Germany and Spain are leading opposition to European Commission plans to ban Chinese technology suppliers from telecom networks as part of new cybersecurity rules. The intersection of trade policy and technology security is creating fractures within the EU, with larger economies that have deep commercial ties to China resisting restrictions that smaller or more security-focused members support.

The EU’s own AI infrastructure plans are stumbling, with the €20 billion gigafactory programme facing delays and funding gaps. That vacuum makes national-level action like Poland’s sovereignty test more likely: if the EU cannot deliver a coordinated infrastructure strategy, individual member states will build their own frameworks for managing technology dependency.

The AI dimension adds urgency that previous technology sovereignty debates lacked. Europe’s inability to access Anthropic’s Mythos cybersecurity model until this week demonstrated that regulatory power alone does not guarantee access to the most consequential AI capabilities. A government that depends on foreign AI systems for critical functions, from defence analysis to healthcare delivery, faces risks that procurement rules alone may not address. European efforts to build sovereign AI alternatives, such as BNP Paribas and Mistral’s cyber-focused model, are gaining political support precisely because of these dependency concerns.

Poland holds the rotating presidency of the Council of the EU in the first half of 2025 and has used the position to push digital sovereignty higher on the bloc’s agenda. Tusk’s Sopot speech signals that Poland intends to lead by example, implementing domestic procurement reforms while advocating for EU-wide standards. Whether other member states follow will depend on whether they view technology sovereignty as a security imperative or a barrier to accessing the best available tools, a tension the EU has not resolved.

A new AI weather forecasting tool released today by the startup WindBorne Systems offers more frequent and accurate predictions on key variables than the world-leading system developed by European governments, thanks to advancements in how sensor readings are fed into deep learning models.

Founded by a group of Stanford students in 2019, WindBorne began by building a better weather balloon, with the idea of selling weather data. But with the arrival of the weather-forecasting deep learning models in 2022, the team realized they could capture more value by building their own model as well.

Today marks the release of the sixth version of that model, WeatherMesh, which the company says is more accurate than traditional and AI forecasts produced by the European Centre for Medium-Range Weather Forecasts (ECMWF), the European intergovernmental organization seen by meteorologists as the leading provider of accurate weather prediction.

One simple way to understand it, WindBorne’s chief product officer Kai Marshland says, is that WeatherMesh-6 “is as accurate five days out as a traditional forecast is the day before,” particularly on surface temperature measurements.

WeatherMesh-6 produces a forecast every hour, as opposed to every six hours, as traditional models do. Its resolution is now down to 3 km in Europe and the continental U.S., where the quality of data is highest.

Traditional weather forecasts are generated by complex physics models that require expensive super computers to run and take a long time to do it. AI models — being built by startups and major labs like Google DeepMind — tend to move faster than physics models, but for now don’t have as high a resolution or predict as accurately over longer time horizons.

Still, weather AI is improving rapidly and already being used at major government agencies around the world. Researchers are working to integrate it into the systems used to aggregate weather data and produce public forecasts.

WindBorne benefits from its unique combination of model-building and data collection. The company now has about 400 balloons in flight gathering sensor readings at any given time, launched from 15 sites around the globe. The advances in its current model come from improvements in how the data collected by the balloons is fed into the models.

“I don’t understand, personally, the business model of being [an] AI based weather company without a dataset advantage,” WindBorne CEO John Dean told TechCrunch.

The ECMWF’s superiority is attributed to the organization’s skills at “data assimilation,” the work of turning disparate sensor readings into a comprehensive, machine-readable picture of the world. For now, AI weather models depend on datasets produced by the ECMWF and the U.S. National Oceanic and Atmospheric Administration (NOAA).

But WindBorne and other organizations are working to feed data directly into the models, and the company’s head of AI, Joan Creus-Costa, says the direct ingestion of data from their balloons and other sources is the key reason for improvement in the new version of WeatherMesh. It’s taken a year of tuning and re-architecting the transformer-based model for the model to deliver these forecasts without losing stability.

“When we started doing [data assimilation], we were still very heavily reliant on ECMWF,” Dean said. “I predict today, if we removed ECMWF’s initial conditions, we would actually still do pretty good.”

The company suffered a scare last year when a United Airlines jetliner flew into one of its balloons. While the plane suffered minor damage, no one was hurt, in part because WindBorne followed U.S. regulations about how large its sensor package could be. Now, however, the company uses the global aviation surveillance system ADS-B to move its balloons out of the way of passing aircraft, in an effort to reduce the odds of another crash.

WindBorne, which has raised $25 million in venture funding with a reported valuation of $85 million in 2024, sells its balloon data to NOAA, where it is used in the American weather forecasting enterprise, and the U.S. Air Force and Navy. The company also sells its forecasts to investors and commodity traders, but Dean says the company remains focused on building out its model and data infrastructure over commercial products, in part because of the changing nature of the information environment.

“I’m not trying to invest a massive team into building a SaaS product, if the way people want consumer information two years from now is through an agent, right?” Dean said.

Correction: This story misreported how WindBorne’s balloons use ADS-B to avoid air traffic; the company monitors air traffic and maneuvers its balloon around it, but hasn’t yet added ADS-B transponders to its sensor platforms.