It is Friday, June 27, 2026. OpenAI launched its most powerful model of the year and approximately 20 people on earth are currently allowed to use it. Alphabet lost $269 billion in market value this week and it was not because of a bad quarter. The week that was supposed to deliver GPT-5.6 to everyone instead delivered a preview to a government-approved guest list and a statement from OpenAI saying it is not happy about this arrangement either. And GPT-4.5 officially retired from ChatGPT today, which means this is also a funeral. Here is everything.
OpenAI launched GPT-5.6 and the US government is personally approving who gets to use it
On June 26, OpenAI previewed GPT-5.6 — its most powerful series to date — in three tiers named Sol, Terra, and Luna. The catch, disclosed in the same breath as the announcement: at the Trump administration’s explicit request, access is currently restricted to roughly 20 government-approved partner organizations. Their names were individually cleared by the White House. Everyone else is still on GPT-5.5.
The model itself is a meaningful step up. Sol is the flagship — built for long-horizon coding, agentic security workflows, and frontier reasoning, with a new “max” reasoning effort mode and an “ultra” mode that spins up sub-agents to tackle complex tasks in parallel. Sol hit 91.9 percent on Terminal-Bench 2.1, ahead of Claude Mythos 5 at 88 percent and GPT-5.5 at 83.4 percent. On ExploitBench — the cybersecurity vulnerability reasoning test — Sol matched Mythos Preview’s performance while using roughly one-third of the output tokens, which is the efficiency claim OpenAI is emphasizing most prominently heading into an IPO that needs a credible path to lower serving costs. Pricing for Sol: $5 per million input tokens and $30 per million output — same as GPT-5.5, suggesting OpenAI is absorbing compute costs to defend market share. Terra delivers GPT-5.5-level performance at half the cost. Luna targets high-volume workloads at $1 input and $6 output.
The government dimension is the actual story. This is the first time an American AI company has launched a frontier model under a government-managed access list — a step beyond the voluntary pre-release review the June 2 executive order established. OpenAI said plainly in its announcement that it does not believe “this kind of government access process should become the long-term default” because it “keeps the best tools from users, developers, enterprises, cyber defenders, and global partners who need them.” The company cooperated anyway, framing it as a short-term concession to avoid a more coercive outcome. Commerce Secretary Howard Lutnick met with Sam Altman before the announcement to confirm all relevant agencies had reviewed the model. Broad availability in ChatGPT, Codex, and the open API is targeted for “the coming weeks.”
The pattern this establishes matters more than this specific launch. Both Anthropic’s Fable 5 and now OpenAI’s GPT-5.6 have had their availability directly shaped by government action in the same month. The administration must formalize its frontier AI evaluation framework by August 2026 under the executive order. Until that framework exists, every major model launch is subject to the same ad hoc negotiation that produced a 15-day global shutdown for Anthropic and a government-approved guest list for OpenAI.
Read the official announcement: OpenAI’s announcement of GPT-5.6 Sol, Terra, and Luna
Read the coverage: Axios on the government restriction and what it means for future launches | The Next Web on OpenAI’s decision to cooperate and publicly push back simultaneously | VentureBeat on the enterprise compliance implications of all three tiers being rated “High” risk | Engadget on Sol, Terra, and Luna explained in plain terms | Memeburn on the pricing, benchmarks, and what the Sol name is actually a reference to
Alphabet lost $269 billion in market value this week and it started with a talent problem and ended with an infrastructure one
The week of June 22 to 27 was the worst for Alphabet stock in more than a year. The company shed approximately $269 billion in market value across five trading sessions. The decline unfolded in two chapters with very different causes.
Chapter one, June 22: two senior Google DeepMind researchers — Noam Shazeer, co-author of the Transformer architecture paper that underpins every AI model in existence, and John Jumper, the Nobel Prize-winning researcher who led AlphaFold — announced they were leaving Google for OpenAI and Anthropic respectively. CNBC confirmed the market treated their exits as a signal about something larger: the competitive intelligence of Google’s most important division walking out the door and into rival labs. Alphabet shares fell 5.7 percent on June 22, its worst single-day performance in over a year.
Chapter two, June 24: a separate and ultimately larger worry took over. Investors processed a specific number buried in Alphabet’s earlier SEC filings — $180 to $190 billion in 2026 capital expenditures, with CFO Anat Ashkenazi confirming that 2027 capex will “significantly increase” beyond that. The market began asking whether AI infrastructure spending at that scale can generate the returns being modeled. Goldman Sachs’ David Solomon told clients the markets are in “greed mode.” The Indmoney analysis of Alphabet’s AI spending cycle is the clearest available breakdown of the thesis on both sides: the bull case rests on $460 billion in contracted Cloud backlog and a 2028 free cash flow recovery; the bear case rests on the possibility that AI compute pricing compresses faster than the infrastructure investment pays off.
The Gemini 2.5 Pro Deep Think launch on June 22 — the strongest model Google has ever shipped, with benchmark numbers that beat Claude Fable 5 on graduate-level science — landed in the middle of this same week and barely moved the stock. Investors are not currently pricing Alphabet on its model quality. They are pricing it on whether $180 to $190 billion a year in data center construction generates a return before the competitive landscape shifts again.
Read the coverage: CNBC on Alphabet’s worst single-day drop in over a year and the DeepMind departures behind it | Quartz on Shazeer and Jumper leaving and what the market read into it | IndMoney’s full breakdown of the three-chapter Alphabet decline — talent, capex, and model regression risk
The US government is now personally approving which companies get early access to frontier AI — and that is a new thing
This one deserves its own section because the coverage has largely embedded it inside the GPT-5.6 story without pulling out what is actually being established as precedent.
In June 2026, the US government took three distinct actions against frontier AI models: it ordered Anthropic to shut down Fable 5 and Mythos 5 globally on June 12, it partially restored Mythos 5 to a vetted set of US organizations on June 27, and it approved a specific list of roughly 20 companies allowed to access GPT-5.6 Sol at launch. These three actions together constitute something new: the US government is now functioning as an access gatekeeper for the most capable AI models in the world.
OpenAI said explicitly that it does not want this to become the norm. Anthropic called the June 12 shutdown disproportionate. Both companies cooperated. The executive order signed June 2 requires the administration to establish a formal, classified evaluation framework for frontier AI capabilities by August 2026. Until that framework exists, the arrangement is ad hoc — Commerce Secretary Lutnick personally signing off on GPT-5.6 access, White House officials reviewing Anthropic’s model capabilities, individual company names appearing on government-managed lists. The question that defines the second half of 2026 for the entire AI industry is whether August’s framework regularizes this into a predictable process or entrenches it as an ongoing negotiation that each lab navigates separately.
Read the analysis: FindSkill’s clear-headed explainer on what the government access framework means for regular users and developers | Build Fast With AI on the full pattern across Fable 5, Mythos 5, and GPT-5.6 Sol and what the precedent establishes
GPT-4.5 retired from ChatGPT today and almost nobody noticed because the week was extremely loud
OpenAI announced on May 28 that GPT-4.5 would retire from the ChatGPT consumer interface after a 30-day sunset period. That period ended today, June 27. Existing GPT-4.5 conversations have been automatically migrated to the corresponding GPT-5.5 model. The API continues to support GPT-4.5 separately from the consumer interface for developers who built workflows on it.
GPT-4.5 launched in February 2026 as OpenAI’s first model to prioritize emotional intelligence and conversational fluency over raw reasoning performance. It was also the most expensive model OpenAI had ever offered at launch — $75 per million input tokens and $150 per million output — pricing that was broadly described as absurd and which OpenAI quietly dropped over subsequent months. Its four-month lifespan in the consumer product makes it the shortest-tenured major GPT model since GPT-3.5. The retirement is a clean signal about where OpenAI’s product strategy has landed: the market does not pay a premium for a model that is good at feelings when the model that replaced it is better at feelings and also significantly better at code, reasoning, and cost.
Read the coverage: Build Fast With AI on the GPT-4.5 retirement and what migrated where
Also worth reading today
- The EU AI Act general-purpose AI obligations deadline is three days away. June 30 is the compliance cutoff for any company with a frontier model deployed in Europe. The list of companies not ready is long. Brussels has not publicly announced an enforcement grace period. (Build Fast With AI June 27 roundup)
- Andrej Karpathy joined Anthropic this week as a research advisor. Karpathy co-founded OpenAI, spent years at Tesla leading Autopilot, returned to OpenAI briefly in 2023, and has since been one of the most influential AI educators and commentators online. His arrival at Anthropic is a significant signal about where he thinks the most interesting research is happening. (Build Fast With AI June 25 roundup)
- OpenAI’s ChatGPT now serves more than 900 million weekly active users, with roughly one-fifth of queries expressing direct commercial intent. The company confirmed advertising has become a core part of its business strategy — positioning sponsored experiences around usefulness rather than attention-based metrics. This is the clearest signal yet that OpenAI is building toward an ad-supported tier ahead of its IPO. (MarketingProfs AI Update)
- Google, Microsoft, GitHub, NVIDIA, Salesforce, Snowflake, and Hugging Face published a draft open standard called Agentic Resource Discovery — a specification that lets AI agents locate, verify, and connect with tools, APIs, MCP servers, and other agents at runtime without preconfigured integrations. It is the first serious attempt to build interoperability infrastructure for the multi-agent ecosystem before it fragments into incompatible silos. (MarketingProfs AI Update)
That is your Friday. GPT-5.6 exists and 20 companies are currently allowed to use it. Alphabet lost $269 billion in a week because investors are doing math on data center construction costs. GPT-4.5 retired quietly while everyone was looking at Sol. And the US government is now approving access lists for the most powerful AI models on earth. Have a good weekend. The circus resumes Monday.