Tuesday opened with one announcement and one absence. Google shipped its Gemini 2.5 Pro Deep Think reasoning mode yesterday, pairing a two-million-token context window with a benchmark lead in science and math. Claude Fable 5 remains offline, pulled by a government export-control directive on June 12 and still banned as of this morning. The thirteen-day free window Anthropic advertised at launch is beside the point now; Anthropic cannot flip the meter back on while the ban holds, and the company has said it intends to restore access once regulators allow it.

The collision is unintentional. The timing is not. The same week that one lab cannot legally serve its best model, the other lab shipped its best reasoning mode and put it behind a $250 subscription. That decision is the news today, and Google’s Deep Think is the headline part of it.

The Lead: Gemini 2.5 Pro Deep Think Lands With a 2M-Token Window and a $250 Subscription Toll

Google shipped Gemini 2.5 Pro with Deep Think on June 22, putting a reasoning mode with a two-million-token context window and a benchmark lead in science and math behind its $250-a-month Ultra subscription tier.

Deep Think is DeepMind’s answer to the slow-thinking models OpenAI and Anthropic shipped first. The model decomposes a problem, walks multiple reasoning paths, and checks its own logic before responding. The pattern is what AI labs internally call test-time compute, spending more inference budget per query to lift accuracy on hard reasoning, and the Deep Think rollout post shows it leading on AIME, GPQA, and Humanity’s Last Exam. Anthropic’s Claude Fable 5 still leads on software engineering benchmarks and long-horizon agentic coding, but it has been inaccessible to subscribers since June 12. What changes this week is which frontier model operators can actually reach.

That second half is the Anthropic story. Fable 5 launched June 9, was included free on Pro, Max, Team, and Enterprise plans, and went dark three days later under a Trump administration export-control directive over jailbreak concerns. Most subscribers got roughly five usable days before the ban landed. The model remains offline. Anthropic has said it will restore Fable 5 to standard subscription access once the regulatory path is clear. In the meantime, Opus 4.8 is the ceiling for Anthropic subscribers.

What It Means for You

The best model each lab makes now costs differently: one costs a $250 subscription, the other costs access you do not currently have.

If you have been waiting for Fable 5 to return, that timeline is not in Anthropic’s hands. The cleanest operating posture is to treat Opus 4.8 as your Anthropic ceiling for planning purposes and evaluate Deep Think against the $250 Ultra toll in the meantime.

Google’s Deep Think is the available option for operators who need a hard-reasoning model today. The two-million-token context window is double what Fable 5 shipped with on its standard window. The same shift shows up in workspace tools operators already pay for: Notion released its External Agents API this week, letting Claude, Codex, and custom agents work directly inside a workspace with a 91% reduction in token usage on database operations. Underneath the consumer apps you actually use, Samsung Electronics rolled ChatGPT Enterprise and Codex out to every employee in Korea and its entire Device eXperience division globally, three years after banning the tool over a data leak.

“The cheap frontier moves every quarter. The team workflow you built does not move with it.”

The bills land first, then the choices.

What’s Moving Underneath

The week’s macro story is talent. The frontier labs spent the spring buying chips and the summer buying people, and three of those moves landed in the last six days. Noam Shazeer, who co-authored the 2017 “Attention Is All You Need” paper that introduced the transformer, joined OpenAI as Lead for Architecture Research on June 18. Google had paid roughly $2.7 billion in 2024 to bring Shazeer back through its Character.AI licensing deal; that investment now reads as a placeholder.

DeepMind answered the same week by hiring more than twenty researchers out of Contextual AI under an $80M to $90M licensing structure that mirrors the Character playbook. The Shazeer-out, Contextual-in trade looks like a draft swap; the asymmetry is that a Turing-class architect is harder to replace than twenty product researchers. Underneath both moves, Getty Images announced a multi-year display deal with OpenAI on Sunday that puts licensed photos inside the ChatGPT search results panel, and Getty Images stock jumped 145% in pre-market trading on the news. The talent flows go one direction; the licensing dollars flow the other.

“Compute is a capex line. People are a multi-year hire. The market is pricing the second one above the first.”

None of these reach your screen this week. All of them are the scaffolding that decides which lab’s app you will be paying for next year.

One Tool Worth Knowing

Gemini 2.5 Pro Deep Think (Google Gemini Ultra subscription, $250/month)

Deep Think is the new top tier of Google’s Gemini app, gated to Ultra subscribers and the Gemini API. It is a slower, more deliberate query mode: the model spends more compute deciding how to answer before it answers. Google publishes the benchmark wins in science, math, and reasoning; expect the per-query cost to look more like a thinking-model rate than a flash-model rate.

The practical routing question is whether to use it at all, or when. Here is one explicit workflow: you are reviewing a supplier contract with conditional pricing clauses that trigger at different volume thresholds. Load the full contract into a single Deep Think query, ask it to map every conditional clause, flag any that conflict, and return a decision tree. That is the kind of multi-step, cross-referenced task where Deep Think earns its toll. For the follow-up email to the supplier, route that to Gemini Flash. You are doing model routing whether you name it or not: reserve the expensive reasoning step for the problem that actually requires it, and let the cheaper model handle everything else. A recent survey of agent loop design from the Berkeley AI Research Lab (May 2026) found that routing the planning step to a reasoning model and the execution steps to a faster model cut overall cost by 40% with no measurable quality loss on the planning output. That result is the principle in numbers.

Wisdom Speaks

“For which of you, intending to build a tower, sitteth not down first, and counteth the cost, whether he have sufficient to finish it?” Luke 14:28, KJV

The Greek behind “counteth” is psephizo: to compute with pebbles. Jesus is not warning his disciples off the building. He is warning them that the meter runs whether they watched it start or not. The discipline this passage commends is arithmetic before commitment: build the tower or do not, but know the figure first. The Fable 5 subscribers who reached for the most capable model on June 9 did the right kind of pebble-counting for two weeks. The ban interrupted the arithmetic before the bill could. That is the rarer failure mode, a cost you cannot control rather than one you failed to calculate. Epictetus puts the same lesson in colder terms: what is not in your control is not in your control; what remains is to know the difference before you commit. Pricing moves. Regulatory access moves. The discipline that holds in either case is sitting down to count before the foundation is poured, not after the scaffold comes down.

---

Yesterday’s digest: John Jumper Leaves DeepMind for Anthropic, on the AlphaFold creator anchoring Anthropic’s biology push. Friday: Anthropic Makes the OAuth Click Disappear, on enterprise MCP going single sign-on with Okta. Earlier in the month: Anthropic Disables Fable 5 and Mythos After White House Export Ban, the export-control story whose effects are still running this morning. The talent war Jumper joined last week and the ban that keeps Fable 5 dark this week are the same story told in different ledgers.