AI News Week 27: US export controls tighten, Fable 5 returns, chip and power shortages bite

1. GPT-5.6 has launched – but only with US government approval

On Friday 26 June, OpenAI introduced GPT-5.6: three models named Sol (flagship), Terra (all-rounder), and Luna (fast and cheap). The catch: you cannot use them yet.

For now, OpenAI is only releasing GPT-5.6 to around 20 partners approved by the US government, and only within the US. The reason is government intervention. Secretary of Commerce Howard Lutnick called OpenAI CEO Sam Altman, warning him not to release the model without official clearance. Access will be decided on a case-by-case basis.

OpenAI is open about its stance: "We do not believe this government approval process should become permanent. It keeps the best tools away from the people who need them." The company is complying anyway – as the fastest route to broad availability in the coming weeks.

Regarding the models: OpenAI states that Terra delivers the performance of its predecessor, GPT-5.5, at half the price. Pricing per million tokens: Sol $5 / $30 (input/output), Terra $2.50 / $15, Luna $1 / $6.

The takeaway: This is the second time Washington has intervened, and the pattern is clear: from now on, governments will decide who gets to use the latest models. It is therefore wiser to build business-critical processes on open-weight models or European alternatives like Mistral.

2. Fable 5 is returning – Mythos 5 remains exclusive

We must distinguish between the two, as the situation is often confused. Fable 5, the smaller sibling of Mythos, remains offline as of 29 June but is close to a broad return. Mythos 5 has been partially back since 26 June – but only for a select few.

Mythos 5: only for selected organisations. In a letter to Anthropic co-founder Tom Brown on 26 June, Secretary of Commerce Howard Lutnick allowed the delivery of Mythos 5 without an export licence to a specified group of US organisations. This includes critical infrastructure operators, federal agencies, and national laboratories (listed under "Annex A"). Mythos 5 is the version without safety filters and was never publicly available.

Fable 5 is close to a general return. Axios reported on 27 June that the US government is about to lift the ban on Fable 5 – potentially as early as this week. The Pentagon and the NSA must still approve, whilst other agencies have already cleared it. It remains unclear whether paying subscribers will regain the promised free access, or if Fable 5 will require payment or identity verification in future.

Debunking a rumour: The brief appearance of Fable 5 in the Android app's model selector on 21/22 June was a display error. Anthropic clarified that they routing "exactly zero queries" to Fable 5. It was never actually available.

The difference between the two: Fable 5 and Mythos 5 are the same model. Fable 5 has additional safety filters that redirect sensitive queries to the weaker Opus 4.8. This is precisely why the protected version gets to return first, whilst the unfiltered Mythos 5 remains restricted to a few.

3. Swiss job market: AI squeezes entry-level roles

The first jobs.ch AI Report has been published. It analysed 7.3 million job advertisements on jobs.ch, jobup.ch, and JobScout24.ch, supplemented by surveys of over 3,600 employees and around 850 companies. The study covered 18 job categories and 19 regions.

The key finding: in 2025, the share of job ads for entry-level candidates was around one-third lower than the average from 2019 to 2022 – the period before the generative AI breakthrough. In professions where tasks are highly suitable for AI support or semi-automation, the proportion of junior roles fell by 16 percent compared to the pre-AI era. Affected sectors include administration, HR, banking, finance, marketing, procurement, sales, IT, and telecommunications.

At the same time, the share of senior profiles rose by 26 percent. Businesses continue to seek experienced specialists who can deliver value immediately.

The takeaway: What we have already seen in software development is now spreading: entry-level jobs are being replaced by AI, whilst senior roles remain in demand. The question remains: how will we find tomorrow's seniors if we stop training juniors today?

4. The bottleneck isn't models – it is chips and power

Events this week highlighted where the AI boom is really struggling: not with models, but with computing capacity, memory, and energy.

Google rations supply – even for Meta. Google has restricted Meta's use of its Gemini models because it cannot supply the computing capacity Meta demands (Financial Times and Bloomberg, 28 June). Several customers are being throttled, and Meta is particularly hard hit. As a result, Meta has put internal projects on hold and ordered staff to cut costs. Even a tech giant investing billions in its own data centres must buy from a competitor – and is now facing rationing.

Oracle suffers its worst week since the dot-com crash. Oracle shares lost 19 percent in a single week, closing at $148.53 on 26 June – the steepest weekly drop since August 2001. The stock is down around 29 percent for June. The cause is financial concern: debt rose to $156.2 billion (by end of May, up from around $87 billion a year prior), investments in AI data centres surged to $55.7 billion (nearly tripling), and free cash flow stood at minus $23.7 billion. Oracle plans to raise another $40 billion through debt and new shares by 2027. Markets are openly questioning whether this costly expansion will ever pay off.

You are co-funding the boom. Hardware memory is running short as AI firms buy up vast quantities for their data centres. The cost is being passed on to consumers. Microsoft is raising Xbox prices by $100 to $150 from 1 August. Apple followed suit on the same day: the MacBook Neo rose from $599 to $699, the iPad Air (128 GB) from $599 to $749, and the iPad Pro (256 GB) from $999 to $1199. PlayStation 5 and Switch 2 prices rose even earlier. Memory prices could increase by 40 to 50 percent in the third quarter, and by another 30 to 40 percent in the fourth. Analysts do not expect relief until around 2028.

And the climate is suffering. Whilst Europe swelters in heatwaves, data centres are struggling to keep cool. Cooling accounts for around 40 percent of their energy consumption under normal conditions, and even more during extreme heat – precisely when the power grid has the lowest reserves. According to a study by the UN University, AI water consumption could equal the annual basic needs of 1.3 billion people by the end of the decade.

The takeaway: The real bottleneck is not AI, but chips and power. I see two solutions. First, specialised models running on small servers, which are often better suited to specific tasks than massive, general-purpose engines. Second, decentralised, local models running on powerful local hardware rather than expanding central data centres. In the long run, physics could help: researchers at Kiel University have shown that electrons in graphene can be controlled with light pulses rather than electricity. This could make chips up to 10,000 times faster whilst using less power. This is still basic research, and integrating electrons reliably into a circuit remains unsolved. But the direction is correct: the future lies in local models on powerful hardware, not ever-growing data centres.

5. In brief

Mistral upgrades for enterprise use. At the end of June, the French provider expanded its integration with corporate software. The updates offer greater control for administrators, tighter access rights, and simultaneous connections to multiple accounts. This mirrors Anthropic's centralised integration updates from week 26. Providers are making their AI enterprise-ready, moving away from individual user setups.

Claude Design shares limits – and exports cleanly. Claude Design now shares the same usage quota as Chat, Cowork, and Claude Code, giving most users more flexibility. Export to PDF and PowerPoint is reliable, and the integration list is growing: Adobe, Canva, Gamma, Miro, Replit, Vercel, Wix, and more. This is highly useful for anyone building designs in Claude and wanting to transfer them to existing applications.

Three things to watch this week

1. Expect regulatory delays for top US models. GPT-5.6 is launching only for approved partners, and Fable 5 is still awaiting its return. Both restrictions are regulatory, not technical. Plan and test a backup provider for critical processes. And check closely whether the return of Fable 5 is tied to fees or identity verification.

2. Invest in your juniors. The jobs.ch report shows that entry-level roles are shrinking whilst senior profiles are in higher demand. Companies that train young talent and equip them with AI tools now will secure the skilled workforce of tomorrow.

3. Evaluate specialised and local models. The bottlenecks are chips, memory, and power – which drive up costs and waiting times. For many tasks, a small, focused model on your own hardware is cheaper and better than a large, cloud-based general-purpose model.