Google launches AI Co-Mathematician: the mathematician’s copilot has arrived

Google has built mathematicians’ "co-pilot"

On May 7, Google DeepMind quietly released a paper on arXiv, officially announcing AI Co-Mathematician—an agent system for mathematical research. Unlike those "exam-oriented" models that just ace MATH or AIME tests, Google’s target customers this time are clearly stated: professional mathematicians.

In plain terms, the progress of LLMs in math over the past few years has essentially been focused on solving “college entrance exam” questions—closed problems with unique answers and formulaic solutions. But real mathematical research isn’t like that: you face open problems, spend three months still unsure how to even formalize a conjecture, and spend most of your time iterating hypotheses, testing counterexamples, looking up papers, or verifying in SageMath. What AI Co-Mathematician aims to tackle is precisely this hard bone.

FrontierMath Tier-4: leaving GPT-5.5 Pro in the dust

The most direct impact comes from benchmarking. Epoch AI’s FrontierMath Tier-4 is currently recognized as the toughest math evaluation set, with questions authored by top mathematicians like Terence Tao—on par with graduate qualification exams or preliminary doctoral research. Tier-4 is the hardest level—this time last year, the combined solve rate of all frontier models on it was below 2%.

According to data disclosed in the paper:

• AI Co-Mathematician: achieved a significantly higher solve rate on Tier-4 than any previously published model
• GPT-5.5 Pro: left behind by nearly twice the margin
• Gemini 2.5 Deep Think: as one of its base models, performs noticeably worse when run standalone compared with the full agent version

The key difference doesn’t lie in the base model’s raw power, but in that it’s not a one-shot reasoning process. AI Co-Mathematician is an agent system configured as a workbench that repeatedly performs propose–verify–revise loops, calling on proof assistants, symbolic computation tools, and literature searches. The whole process resembles a PhD student reasoning at a whiteboard more than a test-taker solving a problem.

What exactly is it

In the paper, Google defines it as “an interactive workbench,” with several design points worth developers’ attention:

1. Not aiming for one-shot completion, but for iterability

Traditional math LLMs interact via the “I ask, you answer” mode. AI Co-Mathematician’s interaction is more akin to Cursor—mathematicians throw in an unfinished idea, the model helps formalize it, outlines possible attack paths, points out where steps are flawed, then you refine based on its feedback and iterate. It doesn’t pretend to crack the problem at once, but embraces the reality that most of the time is spent “exploring.”

This product philosophy aligns closely with what Terence Tao has repeatedly emphasized in recent years: the future of math is human–AI collaboration. AI won’t soon replace top mathematicians—but it will multiply a mathematician’s productivity severalfold.

2. Tool invocation as a first-class citizen

The agent has built-in access to engines like Lean, SageMath, and Mathematica, as well as arXiv search capabilities. Unlike ordinary ReAct-style frameworks, its verifier performs actual formal proofs, rather than letting the LLM “self-check”—the latter has been repeatedly shown unreliable.

3. Long context + stateful memory

A research problem can last days or weeks. The paper highlights that the agent maintains a cross-session “research notebook” state, recording tried paths, excluded counterexamples, and unresolved lemmas. This is crucial for real workflows—mathematicians’ biggest fear is “I tried this path last week, but forgot why it didn’t work.”

Its relationship with AlphaEvolve

Those familiar with Google’s roadmap will immediately recall last year’s AlphaEvolve—the agent that claimed to crack 300-year-old problems and auto-discover algorithms. The two have distinct focuses:

| | AlphaEvolve | AI Co-Mathematician | |---|---|---| | Goal | Automated algorithm discovery, combinatorial optimization | Open-ended mathematical research collaboration | | Interaction | Largely unattended | Emphasizes collaboration with mathematicians | | Output | Executable code / constructive solutions | Proof sketches, counterexamples, formalization fragments | | Base | Gemini + evolutionary search | Gemini 2.5 Deep Think + agent framework |

AlphaEvolve is “AI runs by itself,” while Co-Mathematician is “AI runs with you.” The latter fits the actual working context of most mathematicians—fully automated research remains distant in most domains, while collaborative tools can yield immediate productivity.

A few issues worth discussing

First, can the scores be trusted? FrontierMath’s questions are closed, and Epoch emphasizes safeguarding against data leakage. But Tier‑4 has a limited number of problems, so variance is inherently large. Google’s paper presents mean and variance over multiple runs, fairly transparently, but third-party replication is needed for confirmation.

Second, what does this mean for developers? In the short run, this tool likely won’t offer direct API access—it’s more like an internal tool used by DeepMind and partner mathematicians. However, the paradigm it defines—long-horizon agents + real tool invocation + stateful research memory—is a blueprint for all teams working on "research-oriented agents." Whether you’re building for law, biology, or code review, the recipe is similar.

Third, how strong is Gemini 2.5 Deep Think on its own? This is the signal worth watching. Co‑Mathematician’s base model is from the Deep Think series, and ablation studies in the paper show that when the agent layer is removed, the bare model still outperforms the standard GPT‑5.5 Pro on Tier‑4. This suggests Gemini’s pure reasoning capacity has caught up with or even surpassed OpenAI’s current flagship.

# If you want to test Gemini 2.5 Deep Think's reasoning ability
# OpenAI Hub now supports it, compatible with the OpenAI format
from openai import OpenAI

client = OpenAI(
    base_url="https://api.openai-hub.com/v1",
    api_key="your-key"
)

resp = client.chat.completions.create(
    model="gemini-2.5-deep-think",
    messages=[
        {"role": "user", "content": "Prove: For any prime p>3, p^2 - 1 is divisible by 24."}
    ],
)
print(resp.choices[0].message.content)

Co‑Mathematician itself, as a full agent system, currently lacks a public interface; so far only its base model can be accessed directly.

A broader trend

Viewed in the context of the past six months, the direction is clear:

• OpenAI is betting on GPT‑5.5 Pro for general reasoning
• Anthropic is pushing on agentic coding with Claude
• Google is focusing on vertical research agents—AlphaFold, AlphaEvolve, and Co‑Mathematician form a coherent line

DeepMind’s route excels because research scenarios allow reliable verifiers (Lean for math, wet labs for biology, benchmarks for algorithms), providing clean reinforcement-learning signals—far more dependable than the “LLM grading another LLM” approach of general agents. That’s why they keep widening the gap on hard metrics like FrontierMath.

Terence Tao remarked last year that the real turning point will come when AI can independently propose a meaningful mathematical conjecture. Co‑Mathematician isn’t there yet—it still needs humans to set research directions. But the step from “answering problems” to “collaborative exploration”—Google has now taken that leap.

References

• linux.do community discussion: AI Co‑Mathematician crushes GPT‑5.5 Pro on FrontierMath Tier‑4 – first-hand discussion and benchmark screenshots from the domestic developer community