Google released Gemini 3.6 Flash on July 21, a model built for speed and efficiency rather than topping leaderboards. The new model uses 17% fewer tokens than its predecessor, a move that cuts costs for developers and end users. But on the Artificial Analysis index, Gemini 3.6 Flash ranks 10th — behind models from other major labs.
Why the token cut matters
Fewer tokens mean faster responses and lower bills. For companies running AI at scale, that trade-off can matter more than a high benchmark score. Google is betting that a leaner model will win over developers who need quick, cheap inference. The 17% reduction in token usage is the headline metric, but the model also delivers comparable performance on many tasks, according to internal tests.
Still, the ranking shows Google isn't leading the pack on pure capability. The company is training Gemini 4, a larger model, while Gemini 3.5 Pro remains in testing with select partners. That suggests Google is running a two-track strategy: one for speed and cost, another for raw power.
Personalized agents over leaderboard wins
CEO Sundar Pichai has made clear that Google's focus is on building personalized agents, not chasing benchmark scores. That vision extends to DeepMind, which is betting on world models — systems that understand the physical world — rather than recursive self-improvement. DeepMind's website lists Genie 3 and Gemini Robotics under its world models and embodied AI efforts. Project Genie was extended to Street View in May, and the lab released SIMA 2, an agent that learns in virtual 3D worlds.
Anthropic co-founder Jack Clark described DeepMind as “the most circumspect of the big three,” a nod to its cautious approach. Clark gave a 60% chance of AI conducting its own research by the end of 2028, and a 30% chance by 2027.
Competition heats up on speed and code
Anthropic reported that Claude wrote over 80% of code shipped by May 2026, up from near zero in February 2025. The company's models showed a 52-fold speed gain in April 2026, compared to a 2.9-fold gain a year earlier. Those numbers put pressure on Google to keep up.
On the MLE-Bench test for AI research skill, Google still leads: Gemini 3 scored 64.4% in February, and Gemini 3.6 Flash scored 63.9% in July. But the gap is narrow, and competitors are closing fast.
Google also skipped NVIDIA's open AI alliance, as did OpenAI and Anthropic. The decision keeps the three largest AI labs outside a consortium aimed at standardizing AI hardware and software.
Talent departures and market jitters
Alphabet shares fell 6% in June after two senior researchers left for rivals. The departures raised questions about Google's ability to retain top AI talent. Despite that, Alphabet's June quarter revenue hit $119.8 billion, up 24% year over year. The Gemini app now has 950 million monthly users, giving Google a massive distribution advantage.
Google is training Gemini 4, but Gemini 3.5 Pro is still in testing. Whether the next big model can close the gap on speed and capability — and whether Google can hold onto its researchers — are the open questions as the AI race accelerates.
