GPT-5.6 lands this week, Gemini 3.5 Pro counts down, and Colorado's AI law exits quietly
OpenAI's GPT-5.6 is due the week of June 20; Gemini 3.5 Pro has 10 days; Colorado replaced its AI law before June 30.
By BINA Editorial
The end of June is turning into a checkpoint for every major AI lab. OpenAI is expected to drop GPT-5.6 this week, completing another lap of its six-week flagship cadence. Google is burning through its self-imposed June deadline for Gemini 3.5 Pro with only ten days remaining. And the first major US state AI law has exited without ever reaching enforcement — replaced by a lighter framework that will not come online until 2027. Meanwhile, NVIDIA is pushing its 550-billion-parameter Nemotron 3 Ultra to general availability, extending the open-weight model frontier even as the frontier itself keeps moving.
OpenAI's GPT-5.6 is set for this week
On June 18, OpenAI signaled internally that the GPT-5.6 family would ship "next week" — putting the target squarely on the week of June 20. Chief scientist Jakub Pachocki described GPT-5.6 as a significant improvement over GPT-5.5 across three dimensions: the context window expands from one million to 1.5 million tokens, token efficiency improves by 10–15 percent (directly reducing the cost of long-horizon agentic workflows), and the model is sharpened for multi-step agent execution. The release is expected to land alongside Mini and Pro variants simultaneously — OpenAI's first attempt at shipping a full model family in a single drop rather than staggering the tiers over weeks.
The competitive claim accompanying the release is pointed: OpenAI internally characterizes GPT-5.6 as edging out Anthropic's Fable 5 on agentic coding benchmarks, the category that now drives the most enterprise procurement decisions. Whether that lead holds up under independent evaluation is a separate question, but the framing reflects how central agent-execution performance has become to the frontier model narrative.
The launch also coincides with a planned ChatGPT interface overhaul. OpenAI has delayed its IPO timeline amid ongoing governance discussions around its for-profit conversion, so near-term narrative is being driven by product velocity rather than capital markets.
Gemini 3.5 Pro: 10 days left
Sundar Pichai's commitment at Google I/O on May 19 was explicit: "Give us until next month to get it to you." That next month is almost gone. As of June 20, Gemini 3.5 Pro remains in limited preview for select Vertex AI enterprise customers and has not appeared in the public Gemini app, Google AI Studio, or consumer subscriptions.
The specifications are well-documented: a two-million-token context window, a Deep Think extended-reasoning mode designed to compete directly with OpenAI's o3 and Anthropic's extended-thinking settings, and pricing estimated at roughly ten times the Gemini 3.5 Flash rate — approximately $15 per million input tokens and $60 per million output tokens. Those numbers put it in direct competition with Claude Fable 5 on price as well as capability.
A June slip carries real reputational cost for Google. Gemini 3.5 Flash launched in May and exceeded expectations on price-performance benchmarks, raising the bar for what Pro needs to deliver. If the model ships in the final days of June, it will do so into an environment where GPT-5.6 has already been released and benchmark comparisons will be immediate. If it slips into July, Google cedes a month of narrative to OpenAI at a moment when Fable 5 still holds the top spot on the Artificial Analysis Intelligence Index.
Colorado's AI law exits quietly before June 30
The original Colorado AI Act — the first US state law to regulate high-risk AI in employment, healthcare, financial services, and housing — was scheduled to take effect on June 30, 2026. It never will. On May 14, Governor Jared Polis signed Senate Bill 26-189, which repeals and replaces the original law. A federal magistrate judge had already stayed enforcement in late April following a legal challenge, but SB 26-189 resolves the question legislatively.
The replacement law narrows the framework considerably. Gone are the mandatory risk management programs, algorithmic-discrimination impact assessments, annual audit cycles, and layered disclosure requirements that had alarmed enterprise HR and legal teams since the original law passed in 2024. In their place, SB 26-189 requires advance notice when automated decision-making systems are used in consequential decisions — covering employment, housing, and financial services — plus post-decision disclosures and consumer rights to understand and appeal automated outcomes. The new framework takes effect January 1, 2027.
The practical consequence is that the United States enters the second half of 2026 without any active state AI law in enforcement. The EU AI Act's transparency obligations arrive in August; the US federal National Policy Framework released in March is non-binding and creates no legal obligations. The gap between Colorado's original June 30 date and the EU's August transparency deadline is now an extended period of American regulatory inactivity — and the lobbying effort underway to prevent other states from reviving something like the original Colorado model is intensifying accordingly.
NVIDIA's Nemotron 3 Ultra reaches general availability
NVIDIA announced Nemotron 3 Ultra at Computex on June 1 and shipped weights on June 4. This week the model reaches full general availability on NVIDIA NIM, HuggingFace, OpenRouter, and ModelScope — making the 550B-parameter Mixture-of-Experts model accessible to any developer without a waitlist.
The architecture is efficient by design: with 55 billion parameters active per token despite the 550B total, throughput on agent benchmarks runs five times faster than comparable open-weight models in its class, and task-completion cost is 30 percent lower. The weights, training data, and training recipes ship together under the Linux Foundation's OpenMDW-1.1 license — the first time NVIDIA has released a frontier-scale model with this level of transparency.
Nemotron 3 Ultra is now the most capable open-weight model to come out of a US lab, scoring 48 on the Artificial Analysis Intelligence Index. That is a meaningful milestone. It is not, however, the top of the global open-weight rankings: China's Kimi K2.6 holds a six-point lead at 54 on the same index. The gap between the US open frontier and the Chinese open frontier — measured on a scale where GPT-5.5 scores around 57 — remains a data point that US AI policy circles are tracking closely.