The Decoder

Microsoft CEO Satya Nadella warns against "token-maxing," using the most powerful AI models for every problem. He says frontier models shouldn't be wasted on everyday tasks, and the marginal cost of productivity gains must match the token cost. Yet he admits, "I'm like a token-maxer too. So it is addictive."

Google Research's Gemini-SQL2 turns natural language into executable SQL queries. Built on Gemini 3.1 Pro, it tops the BIRD benchmark at 80.04 percent accuracy, well ahead of OpenAI and Anthropic. Google says the technology could improve natural language features across its data services.

Microsoft and three Chinese universities have developed SkillOpt, a method that optimizes instruction documents for AI agents using principles from traditional model training. A simple Markdown file is enough to boost GPT-5.5 by about 23 points on procedural tasks, and the same file transfers across models and agent environments like Codex and Claude Code.

Anthropic's Claude Fable 5 hits 88 percent accuracy on the hardest FrontierMath tier, a massive jump from Opus 4.5, which sat below 10 percent in early 2026. OpenAI's GPT-5.5 reaches about 75 percent on the same tier. The pace of improvement in AI math keeps accelerating.

An internal memo to 6,000 employees reveals Meta is heading toward billions in AI costs from internal use alone. Starting in 2027, budgets, allocations, and a central dashboard called "AI Gateway" will govern token consumption. CTO Andrew Bosworth put it bluntly: "All motion is not progress and token usage alone is not a measure of impact of any kind."

Moonshot AI has released Kimi K2.7 Code, an open-weights model with one trillion parameters built for programming. It still trails GPT-5.5 and Claude Opus 4.8 in coding benchmarks but costs a fraction of the price. So the key question isn't whether it's the best model, but whether the extra runs you get for the same budget make up for the gap in quality.

The US government has ordered Anthropic to shut down global access to Fable 5 and Mythos 5, citing alleged jailbreak risks. Anthropic is complying but pushing back publicly: the vulnerabilities are minor and exist in competing models like GPT-5.5, the company says, an ironic turn after the company spent months hyping the cybersecurity risks of its own Mythos class. Anthropic warns the move could set a precedent that halts all frontier deployments.

Anthropic surveyed nearly 52,000 Americans about their hopes and fears around AI. Sixty-four percent fear job losses, and 56 percent worry about losing the ability to think for themselves. Daily AI users are far less concerned. Still, most people reject AI in their own workplace, even for tasks they think it can handle.

OpenAI now lets Codex users bank their rate-limit resets and trigger them manually instead of watching them expire on a fixed schedule. If you hit your usage cap mid-session, you can cash in a saved reset right away instead of waiting. Users on the Go, Plus, Pro, and Business plans each get one free reset to start. Plus and Pro users can also invite friends to unlock extra resets.

Claude Fable 5 tops the Artificial Analysis Intelligence Index with 64.9 points and sets records in five of ten benchmarks. But the gain over Opus 4.8 is just 5.7 percent at double the token price. Safety filters with fallback routing push costs even higher.

Within days of each other, Google and OpenAI separately exposed operations allegedly originating in China that use AI for fraud and covert influence campaigns. Both target US infrastructure and political debates.

French AI startup Mistral AI is negotiating a new funding round of around 3 billion euros at a valuation of approximately 20 billion euros.

Anthropic is throttling its new Mythos model for certain tasks while building apps that directly compete with its largest customers. Customers, partners, and investors are pushing back.

OpenAI is acquiring Ona, formerly Gitpod, a startup founded in Kiel, Germany in 2020 that specializes in AI agents and secure cloud development environments for software development.

Jeff Bezos' AI startup Prometheus has closed a $12 billion funding round at a $41 billion valuation. The company launched just last November with $6.2 billion in seed funding. No products yet, because Bezos says sharing details would be 'premature.'

Deezer now offers a free AI music detector that lets users on any major streaming platform check whether AI-generated songs are hiding in their playlists.

OpenAI is considering cutting API token prices to win customers from Anthropic, according to the Wall Street Journal, signaling a potential price war in the AI industry.

Anthropic publishes a sweeping essay and two policy frameworks. The company calls for binding audits of frontier models and paints a picture of AI as a strategic weapon wielded by nation-states.

Anthropic reverses course on a policy that would have secretly undermined AI researchers, but another point of contention persists.

Google released DiffusionGemma, a 26-billion-parameter model that generates text via diffusion, achieving 1,000 tokens per second on an H100 GPU—four times faster than autoregressive models, but with lower quality. It's currently experimental.

Sam Altman told employees he expects an OpenAI IPO "within the next year," but a delay to 2027 is possible. He frames it as caution around self-improving AI, though Anthropic's stronger growth numbers and imminent IPO may be the real reason to wait.

SpaceX wants to launch data centers into space, and Elon Musk is pitching it as a near-trivial engineering problem ahead of the company's IPO. A first AI satellite would match the output of a single Nvidia GB300 rack. But Google's own research suggests real AI training would require about 10,000 tightly coupled satellites.

A German regional court has ruled that Google is directly liable for the content of its AI search overviews. According to the court, previous limited liability protections for search engine operators don't apply to AI overviews. In this case, Google's AI had falsely linked two publishers to fraud and made claims that didn't appear in any of the linked sources. The ruling could set a precedent for AI-generated content liability worldwide.

China plans to invest roughly $295 billion in a nationwide AI data center network over the next five years, with at least 80% of technology from domestic suppliers like Huawei. Meanwhile, Taiwan is considering criminalizing AI chip smuggling to China for the first time.

At WWDC 2026, Apple showed off a rebuilt version of Siri. The assistant runs on foundation models developed with Google. For complex queries, it taps Nvidia GPUs.

OpenAI steps back from fully autonomous AI by 2028, now advocating human-machine collaboration. Altman and Pachocki propose an international body to potentially slow frontier AI development.

OpenAI has confidentially filed an S-1 registration with the SEC, taking the first formal step toward an IPO. There's no set timeline, and the company calls it "a complicated set of tradeoffs." Rival Anthropic recently filed its own IPO paperwork, which likely adds to the pressure.

Microsoft Research introduces Lens, a 3.8B parameter text-to-image model that rivals much larger models by training on 800M detailed captions generated by GPT-4.1. It requires a fraction of the compute. Lens-Turbo generates images in under a second. Open source under MIT.

Google has ordered more than three million AI chips from Intel for 2028. Nvidia is testing Intel's manufacturing tech for its upcoming Feynman architecture. Both moves come as TSMC can't keep up with AI chip demand. Intel's long-struggling foundry division is getting a rare second chance.

Only 26% of companies have full visibility into their AI costs, a KPMG survey finds.