What Is DeepThink — And Why Is Everyone Talking About It?
Last week Google quietly toggled a new switch inside its Gemini 2.5 Pro model and called it DeepThink. Flip that switch and the chatbot changes character: instead of blurting out the first plausible answer, it takes a deep breath, spawns multiple lines of reasoning in parallel, cross-checks itself with external tools, then hands you a single, highly-reasoned response. Google claims the same architecture let a research build of Gemini earn a gold medal score at the 2025 International Mathematical Olympiad—an unprecedented feat for any AI.
How Does “Parallel Thinking” Work?
- Mixture-of-Experts core routes each token to a small set of specialized sub-nets, keeping runtime reasonable.
- An extended inference budget gives the model more “thinking ticks” on hard prompts—several seconds instead of milliseconds.
- Novel reinforcement learning teaches the model to use that extra time productively, generating many candidate chains of thought, pruning bad ones, merging the best, and only then writing the answer.
- Built-in tool calls let DeepThink run live Google Search or spin up a sandbox to execute code, injecting real data into the reasoning loop.
The result feels less like chatting with autocomplete and more like sitting across from a patient (if occasionally stubborn) graduate student.
Benchmark Reality Check
DeepThink rules some leaderboards and lags on others:
| Domain Benchmark | DeepThink Score | Closest Rival | Take-away |
|---|---|---|---|
| Humanity’s Last Exam (general reasoning) | 34.8% | Grok 4 Heavy 25.4% | Best at encyclopedic, tricky Q&A |
| LiveCodeBench v6 (algorithmic coding) | 87.6% | Grok 4 79% | Top for puzzle-style coding |
| SWE-bench (real-world codebases) | 63.8% | Claude 4 Opus 72.5% | Trails in large-scale software engineering |
| USAMO 2025 (proof-based math) | 49.4% | Grok 4 Heavy 61.9% | Not the math king everywhere |
The pattern is clear: DeepThink shines on deep, self-contained problems but stumbles where long-haul context or gigantic repos matter.
The $250 Question: Is It Worth It?
Access lives behind the Google AI Ultra plan—$250/month for a handful of DeepThink prompts per day. Early testers exhausted their quota after as few as five hard queries. For most developers and hobbyists, that math doesn’t add up. Yet a sliver of power users—quant researchers hunting edge-case bugs, lawyers sifting terabytes of discovery, mathematicians chasing conjectures—argue the price is trivial when one correct answer can move markets or careers.
Safety Disclosures Few Noticed
Buried in the model card: DeepThink hits Google’s “early-warning alert” threshold for CBRN and high-impact cyber guidance. Translation: the model knows enough chemistry, biology, and exploit craft to help the wrong person if guardrails slipped. Google pre-emptively admits the risk—an unusual dose of transparency meant to reassure regulators, but also a reminder that raw IQ comes with shadow sides.
Real-World Impressions
- A mathematician fed DeepThink an open conjecture; the model explored “20 or 100” proof paths before landing a novel solution .
- A software engineer asked DeepThink and OpenAI o3-pro to refactor a package; DeepThink proposed swapping libraries entirely, o3-pro dug in its heels and argued. The engineer sided with DeepThink’s simpler fix .
- Most Hacker News commenters? They balked at the price and daily cap, calling the launch “bizarrely uncompetitive” .
Should You Try It?
Consider DeepThink if you:
- wrestle with discrete, high-value puzzles—formal proofs, algorithm design, forensic audits.
- can expense $250/month without blinking.
- need an AI that occasionally refuses harmless requests but almost never hallucinates equations.
Skip it if you:
- spend your days in sprawling production code (Claude 4 Opus is still the workhorse here).
- want unlimited tinkering for side projects (o3-pro or the free Gemini Studio tier is saner).
- hate waiting: DeepThink is intentionally “slow AI.”
Bottom Line
DeepThink signals a fork in the AI road. One lane races toward faster, cheaper chatbots; the other, slower and pricier, chases verifiable reasoning. Google is betting that for some problems, quality beats latency. Whether that niche grows or stays boutique will depend on how many of us truly need Olympiad-level answers—and can afford the tab.