Five developments in AI model pricing, consumer subscriptions, persistent agents, regulatory policy, and infrastructure capital expenditure.

At its annual I/O developer conference on 19–20 May 2026 in Mountain View, California, Google launched Gemini 3.5 Flash, the first model in its new 3.5 series. Gemini is Google's family of large language models – software systems that process text, images, and other data to generate responses and complete tasks. The new model surpasses Gemini 3.1 Pro, the previous leading model in the series, across coding, agentic task completion, and multimodal understanding benchmarks, while retaining the speed characteristics of the Flash tier: four times faster than other frontier models on output tokens per second, according to Google's own measurements.

Pricing has been confirmed at $1.50 per million input tokens, $9.00 per million output tokens, and $0.15 per million cached input tokens – roughly one-third to one-half cheaper than the previous tier. A token is roughly four characters of text; a million tokens is approximately 750,000 words. Gemini 3.5 Flash is rolling out immediately in the Gemini app, Google Search, Antigravity 2.0 (Google's consolidated developer platform), and the Gemini API. Gemini 3.5 Pro – targeting deeper reasoning tasks – was announced as in testing, with general availability expected next month. On benchmarks, confirmed scores include 76.2% on Terminal-Bench 2.1 (coding performance), 83.6% on MCP Atlas (tool-use reliability), and 84.2% on CharXiv Reasoning (multimodal understanding). Separately, Gemini CLI – Google's command-line interface for developers – is being retired, with users required to migrate to Antigravity CLI by 18 June 2026. One pricing detail worth flagging: independent analysts have noted that $1.50 per million input tokens represents ~5x the input cost of Gemini 2.5 Flash, the model it effectively supersedes, though Google frames the comparison on per-task cost rather than per-token cost given the speed advantage.

At Google I/O 2026, Google announced a significant restructuring of its consumer AI subscription ladder. A new "AI Ultra" entry tier priced at $100 per month was introduced, offering five times the usage limits of the $20-per-month AI Pro plan, along with 20 TB of cloud storage, YouTube Premium, and priority access to Antigravity – Google's consolidated developer and agent platform. Separately, Google reduced its highest-end AI Ultra plan from $250 to $200 per month, maintaining the same feature set at a 20% price cut; that tier retains a 20x usage limit over AI Pro. Google is also replacing flat daily prompt caps with a compute-based usage model that refreshes every five hours until a weekly ceiling is reached, with paid top-up credits planned for heavier workloads.

The pricing reset lands against a backdrop of significant usage growth. Sundar Pichai disclosed at the keynote that the Gemini app now has 900 million monthly active users – up from 400 million at I/O 2025 – while Google AI Overviews has reached 2.5 billion monthly active users. Google is now processing more than 3.2 quadrillion tokens per month, versus 480 trillion tokens per month at I/O 2025 – a ~6.7x increase in twelve months. The new $100 tier creates direct price competition with OpenAI's ChatGPT Pro at $200 per month and Anthropic's equivalent offerings, targeting developers and professional power users who need agentic capability without the full premium ceiling.

At Google I/O 2026, Google unveiled Gemini Spark, a persistent personal AI agent it describes as a "24/7 partner that does real work on your behalf." Unlike conventional AI assistants that respond to individual queries, Spark runs continuously on dedicated Google Cloud virtual machines – cloud-based computers maintained by Google – and can be accessed from any device via the Gemini app, email, text message, and Chrome. It is powered by Gemini 3.5 Flash with the full Antigravity 2.0 agent harness, enabling what Google calls "long horizon tasks" – multi-step workflows that involve writing code, parsing documents, and synthesising data across time. In its launch form, Spark integrates with Gmail, Docs, and other Google Workspace apps, with planned expansion to third-party tools via the Model Context Protocol (MCP) – a standard for connecting AI models to external software – over the summer.

Google disclosed specific initial use cases: automatically parsing monthly credit card statements to flag hidden subscription fees; synthesising raw meeting notes from across emails and chats into polished Google Docs; and extracting deadline information from school communications to deliver consolidated daily digests. High-stakes actions such as sending emails or spending money require explicit user confirmation. Gemini Spark will be available to Google AI Ultra subscribers in the US starting the week of 19 May 2026, before expanding more broadly. A companion feature, "Daily Brief," pulls from Gmail, Calendar, and Tasks to present a prioritised morning digest of the day's obligations and suggested next steps. Gemini Spark's functionality depends on persistent access to Gmail, Calendar, Docs, and Chrome browsing context. Google has not yet disclosed retention policies for the cloud VM sessions, nor specific data residency terms – a gap that enterprise legal and compliance teams will need to evaluate before broad deployment.

Axios reported on 20 May 2026, citing two sources familiar with the plans, that a Trump administration AI executive order is imminent. In its current draft form, the order creates a voluntary framework under which AI labs would share their frontier models – their most capable and powerful AI systems – with the federal government at least 90 days before public release, and would also provide access to certain critical infrastructure providers. The order's primary stated purpose is to bolster cybersecurity around advanced AI models. Both national security and civilian agencies appear to have enforcement roles under the draft. Some details were previously reported by Bloomberg.

The draft represents a notable shift for the Trump administration, which repealed the Biden-era AI executive order on Inauguration Day and had characterised heavy AI regulation as a barrier to American competitiveness. Axios reported that the shift was accelerated by concerns about Anthropic's Claude Mythos Preview model – the first model verified by the UK AI Security Institute (AISI) to autonomously complete a 32-step corporate-network penetration simulation, achieving success in 3 of 10 runs. Mythos also achieved a 73% expert-level task success rate on AISI's broader benchmarks. Sources at top AI companies confirmed to Axios that they are cooperating with the White House. The Department of Commerce had already announced separate agreements with Google DeepMind, Microsoft, and xAI to test powerful models ahead of the expected order. The draft's "voluntary" characterisation is significant and its scope contains key unresolved questions: which agencies would participate in the 90-day review, what capability or compute threshold triggers the framework, and whether open-weight model releases – such as Meta's Llama series, which are freely downloadable – fall under its scope.

At the Google I/O 2026 keynote, Alphabet CEO Sundar Pichai publicly confirmed that Alphabet – the parent company of Google – expects to spend $180–190 billion in capital expenditure (capex) in 2026, describing this as roughly six times its $31 billion annual capex in 2022. Capital expenditure refers to spending on physical infrastructure such as data centres, computing hardware, and power systems. The figure had been previously guided during Alphabet's Q1 2026 earnings call, where CFO Anat Ashkenazi raised guidance from $175–185 billion to $180–190 billion following a single quarter in which Google Cloud revenue grew 63% year-over-year to $20 billion. Pichai confirmed at I/O that the majority of 2026 capex is directed at custom silicon – specifically eighth-generation Tensor Processing Units (TPUs), Google's proprietary AI chips: the TPU 8t for large-scale AI training and the TPU 8i for inference workloads.

To contextualise the scale: Alphabet's $180–190 billion is one component of a broader hyperscaler infrastructure commitment. Amazon projects $200 billion in 2026 capex, Microsoft is tracking toward $190 billion, and Meta has guided $115–135 billion. Combined, the four largest US cloud platforms have committed to between $680–720 billion in 2026 infrastructure spend – nearly double 2025 levels, per analysis from Fortune citing company earnings reports. Pichai also disclosed that Google Cloud ended 2025 at an annual run rate of over $70 billion. The aggregate commitment is occurring against documented supply constraints: Pichai acknowledged in the Q4 2025 earnings call that Alphabet expects to "go through the year in a supply constrained way," citing power, land, and chip manufacturing bottlenecks. At I/O, he specifically noted custom silicon as the primary near-term scaling lever, signalling that TPU investment is intended to reduce dependency on external GPU supply chains.