Open Season: The Week AI Stopped Competing on Who's Biggest

Open Season: The Week AI Stopped Competing on Who's Biggest

From Mira Murati's open-weight debut to a looming regulatory deadline, the AI industry's center of gravity just shifted — and it has nothing to do with parameter counts.

Murati's Inkling Gambit: The Anti-ChatGPT Manifesto

It has been nearly two years since Mira Murati departed OpenAI as its chief technology officer. On July 15, her startup Thinking Machines Lab quietly dropped one of the week's most provocative AI releases: Inkling, a 975-billion-parameter, open-weight model that the company says is "not the strongest overall model available today, open or closed." That's not a disclaimer — it's the point.

Inkling is a mixture-of-experts system, meaning it draws on only about 41 billion parameters for any given task despite its enormous total footprint. It was trained on 45 trillion tokens spanning text, images, audio, and video, and reasons natively across all four. Full weights are already live on Hugging Face. But the release is less about the model itself and more about what you're supposed to do with it: fine-tune it. Thinking Machines is pushing Tinker, its customization platform, as the real product — Inkling is the raw material.

The philosophy is a direct rebuttal to the "one model to rule them all" architecture that OpenAI, Google, and Anthropic have spent years building. Murati is betting that enterprises don't want frontier-level performance across the board; they want a precise tool they can own and adapt. That's a bet that requires the open-weight ecosystem to mature quickly. For now, it's a compelling thesis backed by limited evidence — but Inkling is its most concrete proof of concept yet.

GPT-5.6 Went Three Ways — and Microsoft Is Going Inward

OpenAI didn't stand still this week either. On July 9, the company publicly launched GPT-5.6 as three distinct models: Sol, Terra, and Luna. Sol leads with advanced agentic capabilities suited to coding, biology, and cybersecurity tasks. Terra delivers GPT-5.5-level performance at roughly half the cost. Luna optimizes for speed and affordability. Three models, three price points, one message: the era of a single flagship is over.

The strategic logic is straightforward. As AI becomes infrastructure, buyers want the right tool for the job — not necessarily the most powerful one. Charging enterprise customers Sol prices for tasks that Terra handles just as well is a value proposition that doesn't survive long in a competitive market. OpenAI is hedging before someone else forces its hand.

Meanwhile, Microsoft announced it has been quietly replacing select OpenAI and Anthropic frontier model calls in its own products — specifically inside Excel and Outlook — with its own MAI family. Tens of thousands of weekly prompts previously routed to third-party APIs are now processed internally. Microsoft isn't publicizing this as a rivalry; it's framing it as efficiency. But the implications are significant: the company that made a $13 billion bet on OpenAI is now actively reducing its dependence on the models that bet was supposed to bring to market. AI partnerships, it turns out, have the same shelf life as any vendor relationship.

Meta moved in the opposite direction this week, pushing outward rather than inward. The company announced it is integrating its latest image and video generation tools — including Muse Image, its most advanced generator to date — directly into Instagram, WhatsApp, and Meta AI. Muse Image can understand complex compositional prompts, blend source photos, and generate functional QR codes embedded in visuals. Meta's strategy is distribution over differentiation: it doesn't need to win on raw model capability if it can make generative AI native to apps that 3 billion people already use every day.

What 4,000 Researchers Found in Seoul

The International Conference on Machine Learning ran July 6–11 in Seoul, drawing thousands of researchers to present findings that will shape what models look like in 2027 and beyond. Two threads dominated.

The first is "selective activation sparsity," a training method that teaches models to engage only the parameters most relevant to a given task. On reasoning benchmarks, models trained this way matched the performance of models three times their size — a finding with enormous implications for inference costs and edge deployment. If smaller models can reason at frontier level, the economics of AI shift dramatically away from massive centralized compute and toward lean, distributed alternatives.

The second thread builds on that: reasoning capability isn't primarily about model size or even raw training data. MIT and Stanford researchers presented evidence that the crucial variable is how a model learns to self-correct during its reasoning chain. Models trained to identify and fix errors mid-reasoning outperformed models that generated longer but uncorrected chains of thought. The quality of the internal dialogue, not its length, is what matters.

Combined, these findings push toward the same conclusion that Murati's Inkling represents from the product side: the race to scale may have hit its most useful ceiling. Diminishing returns are real, and the research community is now deeply focused on getting more out of less.

The Regulatory Reckoning No One Can Outrun

While labs were launching models, regulators were sharpening pencils. In Brussels, the EU's AI Act transparency rules are set to take effect in August 2026, and the formal publication of its simplification package in the Official Journal of the European Union is expected before the August 2 deadline. The Act's full rollout has been marked by political negotiation and repeated deferrals, but the transparency obligations — requiring AI systems to clearly identify themselves as such and to disclose their capabilities — are now effectively locked in for the EU's 450 million consumers.

Across the Atlantic, the picture is patchwork but accelerating. The Federal Trade Commission set a July 31 deadline for public comment on a new policy statement addressing AI accuracy — an early signal that the agency intends to treat misleading AI outputs as a consumer protection issue rather than a technical novelty. At least 35 AI-related bills have been enacted at the state level in 2026 alone, with Alabama and Georgia passing laws prohibiting health insurers from using AI to make adverse coverage decisions without meaningful human review.

Apple's position in Europe grew more precarious when the EU General Court upheld the designation of iOS and the App Store as gatekeeper services under the Digital Markets Act, dismissing Apple's legal challenge. The ruling reinforces the EU's ability to mandate interoperability and restrict anticompetitive bundling — conditions that could reshape how AI features are distributed on mobile devices in the continent's market.

The regulatory pattern is the same everywhere: what began as voluntary ethics frameworks is hardening into enforceable law. Labs that treated compliance as a PR exercise are now encountering legal obligations with teeth.

The Bottom Line

This week crystallized something the industry has been circling for months: the AI competition has shifted planes. The question is no longer which lab can build the biggest model — it's who can build the most useful one, at the right price, in a form developers and enterprises can actually own and customize. Murati's bet on open weights, OpenAI's three-way product split, Microsoft's quiet insourcing, and ICML's efficiency findings all point in the same direction. Scale was the weapon of the last war. Fit, efficiency, and legal durability are the weapons of the next one. The labs that internalized this early have a head start. The rest are running the wrong race.

Sources

• Thinking Machines amps up its bet against one-size-fits-all AI with its first open model, Inkling — TechCrunch
• Mira Murati's Thinking Machines debuts its first AI model — Axios
• Inkling: Our open-weights model — Thinking Machines Lab
• Latest AI Model Releases — July 2026 (GPT-5.6) — AI Release Tracker
• Vector researchers advance generative AI at ICML 2026 — Vector Institute
• FTC Seeks Public Comment on Policy Statement Addressing AI Accuracy — FTC
• 2026 AI Laws Update: Key Regulations and Practical Guidance — Gunderson Dettmer
• U.S. Tech Legislative & Regulatory Update – Q2 2026 — Inside Global Tech
• Top Tech News in July 2026 — StyleTech
• Every major tech layoff in 2026 — TechCrunch