← Explore

How does Google's latest Gemini model compare with ChatGPT, Claude, and Grok across key benchmarks?

10/14/2026

ChatGPT

Google’s latest flagship, Gemini 4 Argon, is a leader on several benchmarks—but not an across-the-board winner. Its published results are particularly strong in business automation, long-context reasoning, and video understanding; GPT-6 Astra and Claude win some demanding coding and scientific-workflow tests. ()

As of October 14, 2026, Argon is publicly available: Google announced it on September 30 and opened access in the Gemini app and developer API on October 13. ()

Key benchmark scores

For “ChatGPT,” the comparison below uses GPT-6 Astra, OpenAI’s highest-capability model—not every model available inside ChatGPT. Claude is represented by Opus 5.5, and Grok by Grok 4.7. These are published results compiled from different evaluation setups, not a single controlled four-way test. ()

Higher is better. Bold marks the highest listed score, not necessarily a statistically significant lead.

Capability / benchmarkGemini 4 ArgonGPT-6 AstraClaude Opus 5.5Grok 4.7
Business workflows — AutomationBench51.3%41.4%42.5%—
Software engineering — DeepSWE v1.177.9%74.1%74.2%71.0%*
Software engineering — FrontierSWE v255.0%65.5%62.3%—
Terminal tasks — Terminal-Bench 4.057.4%58.2%66.4%37.6%
Scientific workflows — Terminal-Bench Science 0.157.6%68.1%63.3%—
Mathematics — RiemannBench76.0%72.0%69.6%—
Long-context reasoning — GraphWalks, 256K–1M tokens, F184.2%71.8%66.8%—
Video understanding — LVBench91.7%87.5%83.7%—
Computer use — OSWorld 2.0, offline subset, partial score69.2%72.6%——
Vulnerability remediation — CWE-bench v168.0%68.0%67.0%—
Legal workflows — Harvey Legal Agent Benchmark19.6%5.4%3.8%19.6%

Gemini, GPT, and Opus figures come from Google’s published comparison; Grok figures come from its release report. Grok’s DeepSWE result uses high effort; its other listed results use xHigh. “—” means absent from these source tables, not zero capability. ()

What the results mean

  • Gemini: The clearest advantages here are automation and long-context reasoning. Its coding results are mixed: it leads DeepSWE but trails on FrontierSWE and terminal tasks. ()
  • ChatGPT / GPT-6 Astra: Stronger on the listed scientific workflows, FrontierSWE, and OSWorld computer-use test. The newer GPT-6.1 Sol is a lower-cost alternative, so Astra is not a proxy for every ChatGPT experience. ()
  • Claude: Opus leads terminal tasks in this table. Importantly, Sonnet 5.5 separately reports 70.6% on Terminal-Bench 4.0, illustrating why choosing one model per company can obscure stronger results elsewhere in its lineup. ()
  • Grok: Grok 4.7 matches Argon’s published legal-workflow score, but has lower listed DeepSWE and terminal-task scores. Missing results prevent a defensible overall ranking across all four. ()

Important caveats

Google generally uses maximum reasoning settings and draws competitor scores from provider reports or benchmark leaderboards. Agent software, tool access, budgets, and safeguards differ; its video evaluation also supplies different numbers of frames to different models because of API limits. These are useful comparisons, but not perfectly apples-to-apples. ()

Bottom line: Argon looks especially competitive for large-document and multimodal workflows. For coding and scientific work, there is no universal winner—the benchmark and exact model matter more than the company name. ()

Claude

I couldn't pin down a single "latest Gemini" or a clean head-to-head. The sources I found disagree on which Gemini is newest. They name Gemini 3.1 Pro, 3.5 Flash, 3.8 Flash and "Gemini 4 Argon," and they don't test the same model versions. Several are aggregator blogs that contradict each other, so treat the numbers as indicative.

Overall intelligence indexes (most recent)

  • One October 2026 roundup says Claude Opus 5.5 tops the Artificial Analysis Intelligence Index at 58. It puts Claude Sonnet 5.5 at 56, then Claude Fable 5.1, GPT-6 Astra and Gemini 4 Argon tied at 53, with Grok 4.7 behind.
  • Another source has Grok 4.7 at 46 on the same index.
  • Gemini is roughly level with OpenAI's flagship and behind Claude's top model. Grok is last of the four.

Science and reasoning

  • On GPQA Diamond, Gemini 3.8 Flash scores 95.3 against 96.0 for GPT-6 Astra and 93.7 for Claude Fable 5.1. That is a near-tie, and the benchmark is close to saturated.
  • In earlier Gemini 3.1 Pro data, Gemini scored 94.3% on GPQA Diamond and 77.1% on ARC-AGI-2. In the same comparison, GPT-5.4 scored 92.8% and 73.3% and Claude scored 91.3% on GPQA Diamond.
  • Gemini has also led on knowledge breadth and browsing. One comparison lists ARC-AGI-2 (77.1% vs 68.8%) and BrowseComp (85.9% vs 79.3%) against Claude.

Coding and agentic work (Gemini's weakest area)

  • Claude Opus 4.7 led on multi-file coding, with SWE-bench Verified at 87.6% and SWE-bench Pro at 64.3%. The same source puts Gemini 3.1 Pro at 80.6% on SWE-bench Verified.
  • On Terminal-Bench 2.0, Claude Opus 4.7 scored 90.2%, GPT-5.5 84.7%, and Gemini 3.1 Pro Preview 80.2%.
  • On Cursor's CursorBench, Claude Fable 5.1 scored 51.8%, Grok 4.7 xhigh 46.3%, GPT-5.6 Sol 41.7%, and Gemini 3.8 Flash 39.6%. That compares a Flash model against flagships, so it isn't like for like.
  • Grok's own claims are shaky. xAI's chart lists a Terminal-Bench 4.0 score for Claude that doesn't match what Anthropic and OpenAI report, and Artificial Analysis's independent run put Grok 4.7 at 26%.

Factuality and reliability

  • On FACTS Grounding, GPT-6 Astra leads at 82.0%, with Gemini 3.8 Flash at 73.1%.
  • Claude tends to be best calibrated. One source reports AA-Omniscience hallucination of 36% for Claude vs 50% for Gemini 3.1 Pro. It also lists Grok 4 at 64%.

Other strengths

  • Gemini's edge is multimodal input: it's the only one of the four that processes video and audio natively. It is also cheaper, and it was fast in one test, 120.3 tokens per second against Claude Opus 4.6's 55.9.
  • Grok's edge is real-time X data and, in some tests, speed and price.

Bottom line: Gemini is competitive or leading on graduate-level science, abstract reasoning, multimodal tasks, and price. Claude generally leads on coding, agents and calibration. OpenAI's GPT line leads on some composite and grounding measures. Grok is usually behind on independent evaluations. For a decision, check the live Artificial Analysis or LMArena leaderboards and test the models on your own work.

Gemini

Comparing Google’s flagship Gemini lineup (notably Gemini 2.0 Flash / Pro and Gemini 1.5 Pro) against ChatGPT (GPT-4o and the o1/o3 reasoning series), Claude (Claude 3.5 / 3.7 Sonnet), and Grok (Grok 2 / Grok 3) highlights clear strengths and architectural trade-offs across major industry benchmarks.


1. Advanced Reasoning & Academic Knowledge

  • MMLU / MMLU-Pro (Multi-task Language Understanding across dozens of subjects):
    • GPT-4o / Claude 3.5 Sonnet / Grok 3: Generally cluster tightly at the top of standard MMLU (high 80s to low 90s). On the more rigorous MMLU-Pro, reasoning-focused models pull ahead.
    • Gemini: Gemini 2.0 and 1.5 Pro deliver solid academic performance (MMLU ~84–86%, MMLU-Pro ~76–78%), though pure standard models slightly trail the dedicated thinking modes of OpenAI and xAI.
  • GPQA Diamond (Graduate-level, Google-proof science and reasoning questions):
    • The Reasoning Leaders: Extended-thinking models—specifically OpenAI o1/o3 and Grok 3 (Thinking)—lead this category, posting scores in the 75–85%+ range.
    • Gemini & Claude: Claude 3.5/3.7 Sonnet reaches ~65–70%. Non-reasoning Gemini models hover around 60–65%, though Google's experimental Gemini "Thinking" variants bridge this gap significantly.

2. Mathematics & Problem Solving

  • MATH-500 & AIME (American Invitational Mathematics Examination):
    • Grok & ChatGPT (o-series): With the launch of Grok 3 and OpenAI’s o1 / o3-mini, competitive math shifted heavily toward test-time compute (chain-of-thought search). Both models achieve 90%+ on challenging competition benchmarks like AIME and MATH-500.
    • Gemini: Standard Gemini models score competitively on grade-school math (GSM8K >95%), but on advanced competition math (AIME), base Gemini models typically sit lower (~30–45%) unless deployed in specialized thinking/reasoning modes.

3. Coding & Software Engineering

  • SWE-bench Verified (Resolving real-world GitHub issues):
    • Claude 3.5 / 3.7 Sonnet: Claude has held an industry-leading reputation for agentic programming and repository-level refactoring, regularly topping or tying SWE-bench Verified leaderboards (often 40–55%+ depending on scaffolding).
    • Gemini: Gemini 2.0 Flash and Pro brought major architectural upgrades to tool-use and code generation, closing the gap considerably with Claude on HumanEval and LiveCodeBench, though Claude remains the developer favorite for multi-file architectural planning.
    • ChatGPT & Grok: Grok 3 and OpenAI's reasoning models perform exceptionally well on algorithmic puzzles and competitive coding (LiveCodeBench), while Claude maintains a slight edge in stylistic consistency and frontend design execution.

4. Multimodal & Long-Context Reasoning

  • Context Window & Needle-In-A-Haystack (NIAH):
    • Gemini (Clear Winner): Gemini remains unmatched in context length and long-range retrieval. Gemini models natively support 1M to 2M tokens, capable of processing hours of video, entire codebases, or hundreds of thousands of lines of documentation with near-perfect retrieval.
    • Competitors: Claude supports up to 200k tokens, Grok up to 128k–256k, and ChatGPT up to 128k tokens (though certain enterprise endpoints offer larger limits).
  • MMMU (Multi-discipline Multimodal Understanding):
    • Gemini: Because Gemini was built natively multimodal from the ground up (handling interleaved text, audio, and high-framerate video), it routinely posts class-leading or near-top scores on MMMU, video summarization, and audio ingestion.
    • GPT-4o: Highly competent across image and document understanding, but lacks native hour-long video and granular audio-stream processing in the standard API.

5. Chatbot Arena / Human Preference (LMSYS)

On the LMSYS Chatbot Arena, which aggregates blinded human evaluations:

  • All four providers maintain frontier-tier ratings (Elo ratings typically clustered between 1280 and 1380+):
    • OpenAI (o1/GPT-4o) and Claude 3.5/3.7 consistently trade top spots for writing quality, following complex system instructions, and nuanced dialogue.
    • Gemini 2.0 / 1.5 Pro frequently places inside the top 5 globally, outperforming rivals specifically in multimodal queries and multi-lingual retrieval.
    • Grok 3 surged near the top of the reasoning and math sub-leaderboards upon release.

Summary: What Sets Gemini Apart?

CategoryGemini's Standing Relative to Competitors
Context LengthBest-in-class: 1M–2M token windows dwarf the 128k–200k standard of ChatGPT, Claude, and Grok.
Native MultimodalStrongest: Superior direct processing of native video, multi-speaker audio, and large document dumps.
Coding & AgentsVery Strong: Great speed and low latency (especially Gemini 2.0 Flash), but Claude 3.5/3.7 Sonnet still edges it out on complex agentic refactoring.
Hard Math & LogicCompetitive: Performs well in reasoning mode, but pure "thinking" models like OpenAI o1/o3 and Grok 3 (Thinking) lead in PhD-level math/logic competitions.
Latency & CostVery Efficient: Models like Gemini 2.0 Flash offer some of the fastest time-to-first-token (TTFT) and aggressive pricing per million tokens across the frontier landscape.

Grok

Gemini 4 Argon (Google’s latest frontier model, announced late September 2026) is competitive with or ahead of the current top ChatGPT (GPT-6 series), Claude, and Grok models on several major benchmarks, with clear category-level strengths and some gaps. Results are close at the frontier and depend on the specific test, reasoning effort, and evaluation methodology.

Human preference (LMArena Text)

As of mid-October 2026, Gemini 4 Argon (high) leads the style-controlled text arena at 1525 Elo, ahead of Claude Opus 5.5 (high) at 1507. GPT-6 variants (e.g., GPT-6.1 Sol around 1484, GPT-6 Astra around 1475) and recent Grok models sit lower in the rankings. This reflects blind human votes on chat responses.

Artificial Analysis Intelligence Index

This composite (covering knowledge work, agentic tasks, coding, science, and related evals) shows:

  • Claude Opus 5.5 (max) at the top (58).
  • Gemini 4 Argon (high) at 53, matching GPT-6 Astra (max) at 53 and slightly ahead of GPT-6.1 Sol.
  • Grok 4.7 (xhigh) lower, around 46.

Gemini shows particular strength on agentic knowledge-work measures in this suite (e.g., leading AutomationBench-AA) and a notably low hallucination rate on AA-Omniscience relative to other high-scoring models.

Google’s published evaluation table

Google’s comparison of Gemini 4 Argon against GPT-6 Astra, Claude Fable 5.1, and Claude Opus 5.5 (mixed self-runs and public/third-party scores) shows Argon leading or tying on the majority of the 19 rows, especially:

  • Knowledge work: Vals Index 68.9% (vs. Opus 67.0%), AutomationBench 51.3% (clear lead), Vals Finance Agent v2 65.4%, and a large relative lead on Harvey’s Legal Agent Benchmark (19.6%).
  • Long context: GraphWalks near ceiling at shorter lengths (99.7%) and a substantial lead at 256k–1M tokens (84.2%).
  • Selected coding and other areas: DeepSWE v1.1 77.9% (lead), Vibe Code Bench 91.9%, LABBench 2 88.8%, RiemannBench 76.0%, LVBench 91.7%, and a tie on CWE-bench v1 cybersecurity (68.0%, also noted with Grok 4.7 in related reporting).

It trails on several agentic/terminal-style tests, including FrontierSWE v2 (55.0% vs. Astra 65.5%), Terminal-Bench 4.0 (57.4% vs. Opus 66.4%), Terminal-Bench Science, PostTrainBench, and OSWorld 2.0. Methodology notes matter: some rival scores come from public leaderboards or other labs’ reports (different harnesses), and a few Argon runs used non-default settings.

Other agentic and coding signals

On LMArena’s Agent arena (user-judged real agentic sessions), Claude Opus 5.5 leads, with Gemini 4 Argon competitive in the top tier and Grok 4.7 further back. Grok 4.7 (released mid-September 2026) improved on prior Grok versions in agentic knowledge work and coding-agent indexes but generally trails the absolute top models on composite leaderboards.

Practical context

  • Differences of a few points are often within noise or harness variation; the best model depends on the task (knowledge-work agents and very long context favor Gemini in the available data; certain terminal/agentic coding tasks favor Claude or GPT-6 variants).
  • Pricing and access also differ. Gemini 4 Argon launched with promotional pricing (around $2/$10 per million input/output tokens, later higher) and was initially limited (e.g., security program access before broader rollout). Grok 4.7 is priced lower on a per-token basis in reported figures but has a smaller context window in some comparisons (500k vs. 1M-class for others).
  • Vendor tables and self-reported numbers should be read alongside independent sources (LMArena, Artificial Analysis, Vals, etc.), as harnesses, reasoning effort, timeouts, and best-of-n choices affect scores.

Overall, Gemini 4 Argon has returned Google to the top tier—leading current human chat preference and matching or exceeding rivals on knowledge work, long context, and several science/multimodal measures—while Claude often holds an edge on pure agentic/terminal coding composites and the highest Artificial Analysis index scores, GPT-6 models remain very close, and Grok 4.7 is capable but generally a step behind the leaders on the broadest aggregates. Check the latest leaderboards for updates, as rankings shift quickly with new releases and more votes.

Want to ask your own? Try 4in1AI free →