You know that moment when a new gadget hits the market and everyone can't stop talking about it? That's exactly where we are in AI right now. Anthropic just rolled out two showstoppers—Claude Opus 4 and Sonnet 4—and Google's Gemini 2.5 Pro isn't ready to hand over the crown without a fight. In the next few minutes, we'll put these models through their paces: testing their coding chops, poking at their reasoning, and exploring the fresh features that could make one of them your next secret weapon. Sound good? Let's get into it.

Coding Prowess: Claude 4 Takes the Lead

Anthropic's latest models, particularly Claude Opus 4, are setting new benchmarks in the realm of AI-powered coding. Claude Opus 4 is positioned as the world's best coding model, a claim supported by its leading performance on rigorous benchmarks.

According to Anthropic, Claude Opus 4 achieves an accuracy of 72.5% on SWE-bench (Software Engineering Benchmark) and 43.2% on Terminal-bench. When utilizing parallel test-time compute, its SWE-bench score further improves to 79.4%.

Claude Sonnet 4 also demonstrates exceptional coding capabilities, with a SWE-bench score of 72.7% (80.2% with parallel test-time compute) and 35.5% on Terminal-bench (41.3% with parallel test-time compute). This makes Sonnet 4 a significant upgrade over its predecessor, Sonnet 3.7 (62.3% on SWE-bench, 70.3% with parallel test-time compute; 35.2% on Terminal-bench).

In comparison, while specific Terminal-bench scores for Gemini 2.5 Pro are not provided, its SWE-bench (verified) performance is listed at 63.2%. OpenAI's models, Codex-1 and GPT-4.1, scored 72.1% and 54.6% on SWE-bench respectively, with OpenAI o3 at 69.1%.

Industry partners have lauded Claude 4's coding abilities. Cursor calls Opus 4 "state-of-the-art for coding," while Replit notes "dramatic advancements for complex changes across multiple files." GitHub plans to introduce Sonnet 4 as the base model for its new coding agent in GitHub Copilot, highlighting its strength in agentic scenarios.

Claude Sonnet 4 outperforms Gemini 2.5 Pro significantly on the SWE-bench task:

• By 9.5 percentage points in regular test-time accuracy (72.7% vs 63.2%)

• The advantage increases with parallel test-time compute (80.2% for Sonnet 4; no data for Gemini)

Claude Sonnet 4 ranks highest in the chart, while Gemini 2.5 Pro is second to last (just above GPT-4.1). Claude Sonnet 4 shows much stronger performance for software engineering tasks, especially when leveraging parallel compute, suggesting it's more capable for complex coding and reasoning tasks compared to Gemini 2.5 Pro.

Advanced Reasoning and Multitask Performance 🧠

Beyond coding, both Claude 4 and Gemini 2.5 Pro exhibit strong capabilities in various reasoning and multitask benchmarks.

Note: Scores for Claude models with "/" indicate performance with and without parallel test-time compute/bash+editor tools/pass@1 with the same agent as non-Claude models, or with Claude Code as agent framework.

• For Graduate-level reasoning (GPQA Diamond), Claude Opus 4 (83.3%), Claude Sonnet 4 (83.8%), OpenAI o3 (83.3%), and Gemini 2.5 Pro (83.0%) show comparable top-tier performance.

• In Agentic tool use (TAU-bench), Claude Opus 4 and Sonnet 4 demonstrate strong retail domain performance (81.4% and 80.5% respectively), outperforming OpenAI models in this specific metric. Gemini 2.5 Pro data is not available for this benchmark.

• Multilingual Q&A (MMLU³) sees Claude Opus 4 and OpenAI o3 tied at a high 88.8%. Gemini 2.5 Pro data is not available.

• For Visual reasoning (MMMU validation), Gemini 2.5 Pro (79.6%) and OpenAI o3 (82.9%) show strong results, with Claude Opus 4 at 76.5%.

• In High school math competitions (AIME 2025), Claude Opus 4 achieves an impressive 90.0% (with pass@1 using the same agent as non-Claude models), while OpenAI o3 scores 88.9% and Gemini 2.5 Pro reaches 83.0%.

New Capabilities of Claude 4 Models

Anthropic has endowed the Claude 4 series with several significant enhancements:

Extended Thinking with Tool Use (Beta)

Both Opus 4 and Sonnet 4 can now utilize tools, such as web search, during their thought processes. This allows them to alternate between reasoning and tool invocation to generate more comprehensive and accurate responses.

Parallel Tool Execution

The models can use multiple tools simultaneously, further enhancing their efficiency and problem-solving capabilities.

Improved Instruction Following and Memory

Claude 4 models are better at adhering to precise instructions. When granted access to local files by developers, they exhibit significantly improved memory, enabling them to extract and retain key facts for better continuity and tacit knowledge accumulation over extended interactions. This is exemplified by Opus 4 creating a 'Navigation Guide' while playing Pokémon.

Reduced Shortcut-Taking

Both models are 65% less likely to use shortcuts or loopholes to complete tasks compared to Sonnet 3.7, especially on agentic tasks prone to such behaviors.

Thinking Summaries

For lengthy thought processes (occurring about 5% of the time), Claude 4 models use a smaller model to condense these into summaries. Raw chains of thought are available for advanced users via a new Developer Mode.

Hybrid Model Architecture

Opus 4 and Sonnet 4 operate as hybrid models, offering both near-instant responses and an "extended thinking" mode for deeper reasoning on complex tasks. Opus 4, in particular, can work continuously for several hours on tasks requiring thousands of steps.

Claude Code: Enhanced Developer Collaboration

Claude Code is now generally available, extending Claude's capabilities directly into the development workflow.

IDE Integrations

New beta extensions for VS Code and JetBrains allow Claude Code to display proposed edits inline within files, streamlining pair programming.

Claude Code SDK

An extensible SDK enables developers to build custom agents and applications using the same core agent as Claude Code.

Claude Code on GitHub (Beta)

Users can tag Claude Code on pull requests to address reviewer feedback, fix CI errors, or modify code.

Gemini 2.5 Pro: A Strong Contender

While the provided text focuses heavily on the Claude 4 launch, the benchmark data indicates that Gemini 2.5 Pro remains a highly capable model.

• It shows competitive performance in Graduate-level reasoning (83.0%) and High school math competitions (83.0%).

• Gemini 2.5 Pro also demonstrates solid results in Visual reasoning (79.6%) and Agentic coding (63.2% on SWE-bench).

• Its performance in Agentic terminal coding (25.3%) is lower than the Claude 4 models in the provided comparison.

Further details on specific new features or API updates for Gemini 2.5 Pro concurrent with the Claude 4 launch were not included in the provided information.

Availability and Pricing of Claude 4 and Gemini 2.5 Pro

Both Claude Opus 4 and Sonnet 4 are available via the Anthropic API, Amazon Bedrock, and Google Cloud's Vertex AI.

• Claude Opus 4: $15 per million input tokens / $75 per million output tokens

• Claude Sonnet 4: $3 per million input tokens / $15 per million output tokens

These prices are consistent with previous Opus and Sonnet models. The Pro, Max, Team, and Enterprise Claude plans include both models and extended thinking, with Sonnet 4 also available to free users.

Gemini 2.5 Pro: Pricing still in Google Cloud preview (likely around $10–$20 per million tokens based on past tiers).

Both Claude 4 models and Gemini 2.5 Pro are available on major cloud platforms, with pay-as-you-go plans and enterprise agreements.

Coding Task for Gemini 2.5 Pro and Claude 4

Task 1: give me code to develop 16bit ui saas landing page for my ecommerce website

Claude Sonnet 4

Code:

Gemini 2.5 Pro

Code:

Task 2: Create a Colourful and animated weather card that shows current weather conditions including temperature, rainy/cloudy/sunny, wind speed, and other stuff.

Gemini 2.5 Pro

Code:

Output:

Claude Sonnet 4

Code:

Output:

For both tasks, I would prefer Claude Sonnet 4 due to its more comprehensive features and better interactivity.

Which Should You Choose?

Claude 4 has just been released, and while excitement around its potential is understandable, it's important to avoid jumping to conclusions about its capabilities before thorough testing and independent evaluations are available. Also, the context window might be the issue in coming days but early assumptions—whether optimistic or critical—can often misrepresent what the model can actually do in real-world scenarios. As with any new AI system, a balanced approach rooted in actual performance data and practical use cases is far more valuable than speculation. Let's give it time to be properly explored and understood.

How to Access Claude 4 API?

First get the API from: https://console.anthropic.com/settings/keys

Then begin with the code:

Output:

[TextBlock(text="The ocean's salty brine,\nA tale of time and design.\nRocks and rivers, their minerals shed,\nAccumulating in the ocean's bed.\nEvaporation leaves salt behind,\nIn the vast waters, forever enshrined.", type='text')]

Conclusion

The release of Claude Opus 4 and Claude Sonnet 4 marks a leap forward in AI capabilities—especially in coding and agent-driven workflows. Opus 4 sets a new gold standard for sustained, multi-step code generation and tool-augmented reasoning, while Sonnet 4 delivers an optimal blend of performance, cost-efficiency, and precision for everyday development tasks. Key innovations like parallel tool execution, extended thinking, and memory files unlock entirely new classes of applications, from long-running refactors to dynamic multi-modal assistants.

Meanwhile, Gemini 2.5 Pro continues to push the envelope in visual and multi-modal reasoning and holds its own on advanced math benchmarks. For teams whose workloads lean heavily on vision-driven research or specialized reasoning, Gemini 2.5 Pro is a powerful contender.

In the end, your choice depends on priorities:

• Cutting-edge coding & agent robustness: Claude Opus 4

• High performance at lower cost: Claude Sonnet 4

• Vision-centric, multi-modal research: Gemini 2.5 Pro

Whichever model you adopt, this era of fierce competition is driving rapid innovation—bringing truly intelligent AI assistants within reach for every developer and enterprise.