Skip to main content

Google's Secret Math AI 'Mathematica' Leaks: 1M Token Context

Google's Secret Math AI 'Mathematica' Leaks: 1M Token Context

Google appears to be quietly working on a new math-focused AI model, and the details are starting to slip out. Dubbed Mathematica, this internal experiment is built on the DeepThink V3 model and is designed to tackle heavy computation and complex symbolic reasoning. If the leaks hold up, it could become Google's most powerful mathematical reasoning AI to date.

What the API data reveals

According to leaked API data, the model's internal identifier is models/deepthink-mathematica-tf-raw-thoughts. That's a mouthful, but the numbers behind it are what really catch the eye: a 1 million token context window and an output limit of 65,536 tokens. For context, the context window is the total amount of text the model can process at once, while the output limit is how much it can generate in a single response. In plain terms, this thing is built to handle enormous, intricate problems without losing track.

Image

'Unstable' and 'Teamfood': what the labels mean

The leaked configuration also includes the label UNSTABLE_EXPERIMENTAL, which suggests this version is still very much a work in progress. There's also a Teamfood tag — that's Google's internal shorthand for employee testing. In other words, the service is currently locked behind closed doors, being poked and prodded by Googlers before it ever sees the light of day.

Google's math ambitions are heating up

This isn't Google's first foray into math-heavy AI. TestingCatalog notes that the company previously launched the Gemini DeepThink IMO mode for advanced math competitions, and on September 15, it rolled out Gemini 3.8 Live and Live Extended Thinking. So Mathematica isn't a bolt from the blue — it's part of a broader push to make AI that can reason through tough mathematical problems with precision.

Why this matters

Mathematical reasoning is a tough nut for AI. It requires not just pattern recognition but also the ability to follow logical steps, handle abstract symbols, and maintain coherence over long stretches of text. A 1 million token context window could let Mathematica digest entire textbooks or massive datasets in one go, while the 65,536 token output limit means it could produce detailed, step-by-step solutions without cutting corners.

Of course, the 'unstable experimental' label is a reminder that this is early-stage stuff. But if Google gets it right, Mathematica could be a game-changer for researchers, engineers, and anyone who's ever stared at a thorny equation and wished for a smarter assistant.

Key Points

  • Leaked model: Google's internal 'Mathematica' is based on DeepThink V3 and optimized for heavy computation and symbolic reasoning.
  • Massive specs: 1 million token context window, 65,536 token output limit.
  • Internal testing: Tagged 'UNSTABLE_EXPERIMENTAL' and 'Teamfood' — currently limited to Google employees.
  • Part of a trend: Follows Gemini DeepThink IMO mode and Gemini 3.8 Live, signaling Google's focus on advanced math AI.