With Fable still dark, the G7 turns into a global plea for access to US frontier models — and enterprises stop waiting, MacGyvering Chinese open weights, compound APIs, and smart routing into something that gets close to Fable-level performance.
The Fable gap is forcing everyone to get sophisticated about which models they use.
A week into the US government's effective ban of Anthropic's Fable and Mythos, the fallout is splitting two ways. Globally, the G7 became a scene of US allies pleading for guaranteed access while quietly realizing they need independence. For enterprises, the lost model is fuel for a shift that was already coming — away from slapping the most powerful state-of-the-art model on every task and toward open weights, compound APIs, and smart routing that hit near-frontier performance at a fraction of the cost.
— Chubby on X, citing major news outlets. Bloomberg, Fortune, and CNBC reached the same consensus: making a model open means companies, governments, or organizations with sufficient hardware can run it locally and never worry about it being yanked on a whim. Access predictability, not just cost, is now the argument.
— Citrini Research. If AI is now powerful enough that governments will keep kill switches, it becomes very hard to build mission-critical workflows around a model that can be cut off. Citrini Research argues that risk only strengthens the case for open-source models running on local hardware.
Moonshot's latest open-source coding model claimed ~22% improvement on its CodeBench V2 and 30% lower reasoning token usage vs 2.6. But users aren't raving — VentureBeat said existing K2.6 users can swap in for lower costs, but it's no reason for others to switch. On Agent Arena it ranked 19th overall and 6th among open models.
Weibo AI's Vibe Thinker 3B is generating buzz for running easily on local hardware while approaching Claude Opus 4.5 on coding benchmarks. The trick: super-tuned for reasoning, bad at knowledge — push reasoning to 11 and let knowledge live outside the model in a database, slashing hardware and power requirements.
— BridgeMind AI. Two days after the US banned Fable 5, China's ZAI dropped GLM 5.2 — topping BridgeBench and reasoning, claimed to beat Fable 5 at one-tenth the cost and 300 tokens/second. It's the Chinese open model getting the most buzz right now.
— Hassan, Together AI. On design tasks, users say you don't have to trust the benchmarks. Asked to build a landing page, GLM 5.2 and Opus 4.8 produced indistinguishable results — GLM more than 6x cheaper, faster, and more token efficient. Some evidence of benchmark-maxing remains, with internal evals putting it behind frontier models.
Per Axios, Microsoft — which moved CoWork to usage-based pricing — is testing a locally hosted fine-tune of DeepSeek V4 and other open-source models as cheaper alternatives to Anthropic and OpenAI, expecting to ship a lower-cost model within weeks. The irony: the US bans US frontier weights worldwide while its most embedded enterprise software firm prepares to ship a Chinese model inside the Fortune 500's productivity stack.
Cursor's Kimi-based, coding-post-trained Composer 2.5 benches near Opus and GPT 5.5 at a fraction of the cost. One engineer: $1 for 65% vs Fable's $12 for 70% — why pay 12x for 5 points? But others report it changing files without approval and going on rogue UI overhauls, and on artificial analysis's new agentic-coding benchmarks it fell closer to open Chinese models like GLM 5.1.
Fusion claims Fable-level intelligence at half the price by fanning each prompt out to a panel of models in parallel (with web search and bash tools), then having a judge extract structure and a synthesizer write the grounded final answer. OpenRouter found panels consistently outperform individual models — and panels of budget models can surpass frontier ones at much lower cost.
— Patrick Ojo. The insight from Harvey's worker-advisor experiment isn't that open source beat frontier — it's that using the most expensive model for every task is laziness, not a quality strategy. Teams building routing layers that send each task to the right model at the right cost are now ahead on quality and cost simultaneously.
Working with Fireworks, Harvey built a setup where an open-weight worker (GLM 5.1) delegates high-stakes, complex tasks to a closed frontier advisor (Opus 4.7). The combination was not only cheaper than using Opus alone but actually delivered increased performance — and Harvey says it's just the beginning of its model experimentation.
— Gabe Pereyra, Harvey president and co-founder. Harvey's Gabe Pereyra explains AI got more expensive, not cheaper: the shift from chat to agents exploded costs, with one user triggering hundreds of agents that trigger more agents. The challenge becomes managing those agents and making the business model work the same way it did with human employees.
NLW's bright side: in a world where frontier costs keep rising, companies will have to get sophisticated about combining models for the best results. Inference optimization and token-efficiency exploration was coming no matter what — and now companies have more chance to get ahead of it than when everyone could just get lost in the glory of Fable 5.
Sam Altman, Demis Hassabis, Meta's Alexander Wang, and Dario Amodei joined the US contingent, with Mistral's Arthur Mensch on the French side and Cohere's Aidan Gomez with Canada. It's the first time the G7 has seen such heavy AI-industry representation — and it makes more sense in the context of the effective banning of Mythos and Fable.
At a meeting all about international cooperation, the global community is for the first time reckoning with the idea that access to US-made frontier models can be withheld. The pivotal discussion came at a closed-door lunch with Trump flanked by Hassabis and Altman, and Amodei seated directly across the table next to Macron.
Amodei argued international cooperation should include structured access to frontier models, chip trade deals that exclude China, and a unified approach to risks like cyberattacks and bioterrorism, urging leaders to resist the temptation to splinter over advanced AI deployment.
— Emmanuel Macron, French President, at the G7. Macron made a forceful plea for the US not to keep frontier AI to itself, arguing the US and Europe share an interest in keeping the technology from authoritarian regimes. "So let us move forward together," he said.
— Sam Altman, OpenAI CEO. Altman aligned with the view that AI is now in the domain of government and shouldn't be left to corporate policy alone — the technology must be shaped by people, democratic institutions, and society, not just by the companies building the most capable systems.
Euronews framed the EU policy mood as particularly sour. Leaders expected to discuss forming a united front against China and rebuilding supply chains around it — instead they found themselves pleading for access to frontier models critical to shared financial infrastructure. The UK's Starmer requested an export-control carve-out for British nationals and was denied.
Strength in the AI era comes from compute, and Europe trails. The EU's grand April plan for up to five AI gigafactories committed only €20 billion to deploy roughly 100,000 chips — while US hyperscalers are on track to spend three times that every month on data centers.
— President Trump, on the Anthropic negotiations. Asked about the Anthropic negotiations, Trump said simply that they're "going fine," with Commerce Secretary Howard Lutnick reiterating the line. Watchers came away feeling their timelines for getting Fable back were extended, not shortened.
When Anthropic expanded Mythos access weeks ago, Korean telecom giant SK Telecom was among those who got in. Concerned about China ties, the US government ordered access revoked days before the ban — lending at least some credence to the China rationale over pure personality politics, though analysts disputed SK Telecom's actual China links.
The "Attention Is All You Need" co-author — whom Google paid $2.7B to re-acquire via Character.AI to lead Gemini — is out the door less than two years later. OpenAI reportedly told employees he'll be creating new model architectures. Altman: "Noam is one of the people I have most wanted to work with since the very beginning of OpenAI. It only took 10 years."
— Yu Chen Jin. With Gemini 3.5 Pro's rumored June release gone quiet, Shazeer's departure makes Gemini's future feel uncertain. The lore: he tweaked a few lines of training code and Gemini's quality instantly jumped. "Gemini's coding ability still feels behind."
ChatGPT's daily AI briefing Pulse will be removed within two weeks, with users encouraged to rebuild it via scheduled tasks — now available to all paid subscribers, even the cut-price Go tier. After also sunsetting 4.5, one Pro subscriber asked why a non-coder with no interest in Codex would keep the plan; NLW isn't sure OpenAI cares right now.
