AI Daily Brief: Washington Slows GPT-5.6, Gemini's Deep Think Takes the Lead

Today's frontier-AI stories are really about the same thing from three directions: how much capability labs are willing to hand out, how fast that capability is moving up the benchmark charts, and what it actually takes to build the silicon underneath it. Personnel reshuffles aside, the more durable story this week is about access tiers, reasoning benchmarks, and chips.

OpenAI built a ladder for its cyber models, and Washington just asked it to slow the next rung

OpenAI's cybersecurity push now runs on three distinct access levels. Plain GPT-5.5 ships with standard safeguards for general use. Verified defenders who sign up for Trusted Access for Cyber get a version with lighter restrictions on legitimate defensive work: vulnerability triage, malware analysis, patch validation. GPT-5.5-Cyber, the most permissive tier, is reserved for a smaller group of vetted defenders working on red-teaming and exploit validation inside controlled environments, paired with mandatory phishing-resistant account security. It's a deliberate escalation ladder: the more access a model grants, the more identity verification and monitoring comes with it.

That ladder is about to get a new rung, and this time Washington is involved before launch rather than after. Bitcoin World reported that the Trump administration asked OpenAI to limit GPT-5.6's initial release to a small group of partners approved customer by customer, with the Office of the National Cyber Director and the Office of Science and Technology Policy involved in the review. Sam Altman told staff the company hopes to follow with a broader release "a couple of weeks later" if the limited rollout goes smoothly. It's a notable reversal for a company that has historically shipped models widely and fast, and it puts OpenAI in roughly the same position Anthropic has occupied since its own frontier cyber model went through a similarly gated rollout. Traders are pricing the delay in real time: Polymarket odds on GPT-5.6 shipping by June 30 sit at 83 percent, down from 89 percent a week earlier as the government review adds friction to the timeline.

Gemini's Deep Think mode resets the reasoning leaderboard

While the cyber story is about access, Google's latest release is a pure capability jump. Gemini 2.5 Pro with Deep Think, an extended reasoning mode that runs internal chain-of-thought before producing an answer, posted 82.4 percent on GPQA Diamond, graduate-level physics, chemistry, and biology, ahead of both Fable 5 and GPT-5.5, and 89.8 percent on MMLU-Pro and 94.1 percent on HumanEval+, the highest scores yet recorded on either benchmark. The catch is that no single model leads everywhere anymore: Fable 5 still holds a wide lead on software engineering benchmarks like SWE-bench Verified, where Gemini's new mode trails well behind. Deep Think also costs roughly four times the standard per-token rate, the price of making a model think before it answers. The frontier has split into specialized leaderboards rather than one model winning across the board, and which one matters most now depends on whether the buyer is doing research or shipping code.

OpenAI's first chip, built with Broadcom

Underneath both of those stories sits a more basic constraint: compute. OpenAI and Broadcom this week unveiled Jalapeño, OpenAI's first custom chip, co-developed from initial design to manufacturing tape-out in nine months, which the companies describe as the fastest ASIC development cycle achieved in high-performance semiconductors. The chip is purpose-built for inference, the work of actually answering a prompt rather than training a model, and OpenAI says it used its own models to speed up parts of the design and optimization process. Initial deployment is slated for the end of 2026. It's a small data point on a bigger trend: the labs racing to lead on reasoning benchmarks are also racing to control the hardware those benchmarks run on, rather than rent it from one supplier alone.

What it means for India

The chip story has a direct India angle, even though India isn't the one building Jalapeño. The country's own bet is that it can't just be a buyer of frontier AI infrastructure built elsewhere. Google has committed $15 billion through 2030 for an AI hub and data center in Visakhapatnam, Andhra Pradesh, a 1-gigawatt facility whose foundation stone was laid in April. Reliance Industries is separately planning to spend roughly $17 billion on a 1.5-gigawatt data center cluster in the same city, which would become India's largest. That sits alongside Amazon's additional $13 billion AI and cloud commitment announced this week, pushing Amazon's total planned India investment to $48 billion by 2030. None of that closes the gap with a nine-month chip-to-tape-out cycle, but it's the clearest sign yet that India sees compute and data-center capacity, not just app-layer AI services, as the part of this race worth fighting for.

Markets and AI money

Indian markets were shut today for Muharram, so Thursday's close is the most recent read. The Sensex settled at 77,100, up 0.14 percent, with profit-taking in IT and metal shares trimming early gains; the Nifty 50 closed at 24,056, also up 0.14 percent.

The pattern across all three stories today is the same. Capability is fragmenting into specialized tiers, whether that's cyber-model access levels, reasoning-versus-coding leaderboards, or who controls the silicon underneath. The race is less about one model beating every other model, and more about who controls the layer everyone else has to build on top of.