Claude Sonnet 4.6, Gemini 3.1 Pro, and GLM‑5: Deep Technical Comparison and Benchmark Review

Core Model Releases

In February 2026 Anthropic released Claude Sonnet 4.6 (Feb 18), Google released Gemini 3.1 Pro (Feb 20), and Zhipu AI released GLM‑5 (Feb 12). All three target higher task‑oriented performance.

Technical Highlights

Claude Sonnet 4.6

Beta support for a 1 000 000‑token context window, enabling full‑code‑base or multi‑document reasoning.

OSWorld‑Verified computer‑use benchmark: 72.5% (significant improvement over the predecessor).

Prompt‑injection resistance comparable to Opus 4.6.

User surveys: 70% prefer Sonnet 4.6 over Sonnet 4.5; 59% rate it better than Opus 4.5 in specific scenarios.

Gemini 3.1 Pro

Hybrid‑expert Transformer architecture.

Supports up to 1 M input tokens and 64 k output tokens; multimodal inputs (text, video).

ARC‑AGI‑2 abstract‑reasoning benchmark: 77.1% (more than twice Gemini 3 Pro and higher than Claude Sonnet 4.6 58.3%).

GPQA Diamond scientific‑knowledge benchmark: 94.3% (top among the three models).

LiveCodeBench Pro Elo: 2887; SWE‑Bench Verified coding: 80.6%.

Generates high‑resolution SVG and 3D interactive scenes.

GLM‑5

744 billion parameters, 28.5 trillion training tokens.

DeepSeek sparse attention reduces deployment cost.

SWE‑Bench Verified: 77.8%.

Terminal‑Bench 2.0: 56.2%.

Front‑end success rate 98.0% (close to Claude Opus 4.5).

Released under MIT license; deployable on Huawei and Cambricon chips.

GitHub repository: https://github.com/zai-org/GLM-5 Hugging Face model page: https://huggingface.co/zai-org/GLM-5 OpenRouter endpoint:

http://openrouter.ai/z-ai/glm-5

Benchmark Comparison (selected scores)

ARC‑AGI‑2: Claude 4.6 58.3%, Gemini 3.1 Pro 77.1% (best), GLM‑5 N/A.

GPQA Diamond: Claude 89.9%, Gemini 3.1 Pro 94.3% (best), GLM‑5 N/A.

SWE‑Bench Verified: Claude 79.6%, Gemini 3.1 Pro 80.6% (best), GLM‑5 77.8%.

Terminal‑Bench 2.0: Claude 59.1%, Gemini 3.1 Pro 68.5% (best), GLM‑5 56.2%.

OSWorld‑Verified: Claude 72.5% (only model reported).

BrowseComp intelligent search: Claude 74.7%, Gemini 3.1 Pro 85.9%, GLM‑5 87.4% (best).

Capability Breakdown

Long‑Context Processing – Claude Sonnet 4.6 leads with a 1 M token window and efficient full‑context inference; Gemini 3.1 Pro also supports 1 M input but shows 26.3% point‑wise performance at that scale; GLM‑5 offers a 200 K token window but demonstrates high efficiency for long texts.

Coding Ability – Gemini 3.1 Pro dominates competitive and scientific programming (LiveCodeBench Elo 2887, SWE‑Bench 80.6%); Claude Sonnet 4.6 excels in code understanding and context integration (user feedback on precise modifications); GLM‑5 provides strong cost‑effective coding for open‑source projects (front‑end success 98.0%).

Multimodal Capability – Gemini 3.1 Pro leads with text, video, SVG, and 3D generation; Claude Sonnet 4.6 scores 75.6% on MMMU‑Pro visual reasoning with tool assistance; GLM‑5 focuses on text and code, with limited multimodal support.

Safety & Reliability – Claude Sonnet 4.6 matches Opus 4.6 in prompt‑injection resistance; Gemini 3.1 Pro reduces hallucination rates, improving data‑driven reliability; GLM‑5 relies on user‑configured security frameworks.

Pricing (per million tokens)

Claude Sonnet 4.6: $3 for input, $15 for output.

Gemini 3.1 Pro: $2 input for ≤200 k tokens, $4 beyond; $12 output for ≤200 k tokens, $18 beyond; cache $0.2‑0.4 per million tokens, storage $4.5 per million token‑hour; first 5 000 search calls free, then $14 per 1 000.

GLM‑5: $0.8 input, $2.56 output (≈1/6 and 1/10 of Claude Opus 4.6); open‑source deployment free, subscription plan increased 30% for paying users.

Access & Deployment

Claude Sonnet 4.6 – available to all Claude plan and API users via Claude website, Cowork, API, and major cloud platforms; no special limits.

Gemini 3.1 Pro – available to developers, enterprises, and consumers via Google AI Studio, Vertex AI, Gemini App, NotebookLM; preview tier imposes quota limits.

GLM‑5 – accessible to code‑subscription users and the open‑source community via Zhipu website, GitHub, Hugging Face, OpenRouter; rollout limited by compute capacity and international market restrictions.

Application Fit

Claude Sonnet 4.6

Core scenarios: enterprise office automation, legal contract analysis, long‑document summarization, code audit.

Target users: SMBs, legal professionals, content creators, individual developers.

Advantages: 1 M token context, consistent performance, simple pricing, strong security for sensitive data.

Gemini 3.1 Pro

Core scenarios: scientific research, complex system development, multimodal content creation (SVG, 3D), agent development, creative coding.

Target users: tech enterprises, research institutes, professional developers, designers.

Advantages: top abstract reasoning, multimodal input, efficient token utilization for large tasks.

GLM‑5

Core scenarios: open‑source project development, SMB tech rollout, front‑end/back‑end coding, internal tool building, deployment on domestic chips.

Target users: startups, development teams, organizations with domestic‑chip compliance, student developers.

Advantages: very low usage cost, open‑source flexibility, support for Huawei/Cambricon chips, coding performance close to closed‑source flagships.

Industry Impact

Gemini 3.1 Pro shows that core reasoning breakthroughs can create more value than raw parameter scaling. Claude Sonnet 4.6 demonstrates market demand for stable, practical performance. GLM‑5’s open‑source pricing lowers entry barriers and promotes AI democratization.