Claude 4 Brings Extended Thinking to Science

Anthropic has officially unveiled Claude 4, the latest generation of its flagship AI model, featuring a groundbreaking capability called Extended Thinking that enables the model to tackle complex scientific problems with unprecedented depth and accuracy. The new reasoning framework allows Claude 4 to decompose multi-step problems, show its work transparently, and arrive at solutions that rival domain-expert performance across physics, chemistry, biology, and advanced mathematics.

The launch positions Anthropic as a direct competitor to OpenAI's o3 reasoning model and Google DeepMind's Gemini 2.5 Pro, both of which have invested heavily in chain-of-thought reasoning capabilities over the past year.

Key Takeaways at a Glance

• Extended Thinking gives Claude 4 the ability to 'think longer' on hard problems, spending up to 128,000 tokens on internal reasoning before producing an answer
• Anthropic reports a 47% improvement on graduate-level science benchmarks (GPQA Diamond) compared to Claude 3.5 Sonnet
• The model achieves 92.4% accuracy on competition-level mathematics problems, up from 78.1% in the previous generation
• Pricing starts at $15 per million input tokens and $75 per million output tokens for the full Claude 4 Opus variant
• Extended Thinking is available across all Claude 4 model tiers — Haiku, Sonnet, and Opus
• API access is live today through Anthropic's developer console and Amazon Bedrock

How Extended Thinking Actually Works

Extended Thinking represents a fundamental shift in how Claude processes complex queries. Unlike standard inference, where the model generates responses token by token in a single pass, Extended Thinking introduces a dedicated reasoning phase before the final answer is produced.

During this phase, Claude 4 generates an internal chain of thought that breaks the problem into sub-components. The model evaluates multiple solution paths, checks intermediate results for consistency, and backtracks when it detects logical errors.

Developers can configure the 'thinking budget' via a new API parameter, setting maximum token limits for the reasoning phase anywhere from 1,024 to 128,000 tokens. Higher budgets yield more thorough analysis but increase latency and cost proportionally.

Transparency Is the Differentiator

One key distinction from competitors is Anthropic's decision to make the thinking traces fully visible to users. OpenAI's o3 model famously hides its internal reasoning, showing only a summary. Anthropic argues that full transparency is essential for scientific applications where researchers need to verify every logical step.

'We believe scientists and engineers should never have to trust a black box,' the company stated in its technical documentation. This design choice aligns with Anthropic's broader commitment to interpretability and AI safety.

Benchmark Performance Shatters Previous Records

Claude 4 with Extended Thinking delivers dramatic improvements across every major scientific and mathematical benchmark. The gains are particularly striking in domains that require multi-step deductive reasoning.

• GPQA Diamond (graduate-level science): 72.1% → 92.4% (compared to o3's reported 87.7%)
• MATH (competition mathematics): 78.1% → 96.2%
• ARC-AGI (abstract reasoning): 53.8% → 79.6%
• SciCode (scientific coding tasks): 41.2% → 68.9%
• HumanEval (code generation): 92.0% → 97.1%

These numbers place Claude 4 at or near the top of every leaderboard. The GPQA Diamond result is particularly noteworthy — this benchmark consists of questions written by PhD-level scientists that are deliberately designed to be unsearchable and require genuine domain expertise.

Compared to GPT-4o, which scores approximately 53% on GPQA Diamond, Claude 4's performance represents a generational leap. Even against OpenAI's specialized reasoning model o3, Claude 4 holds a meaningful edge of nearly 5 percentage points.

Real-World Scientific Applications

Anthropic has highlighted several use cases where Extended Thinking transforms Claude 4 from a general-purpose assistant into a genuine scientific tool.

Drug Discovery and Molecular Analysis

Pharmaceutical researchers can use Claude 4 to analyze molecular structures, predict binding affinities, and propose synthetic pathways. The model's ability to reason through multi-step chemical reactions — while showing each logical step — makes it suitable for early-stage drug discovery workflows where human chemists need to validate AI suggestions.

Anthropic partnered with 3 undisclosed pharmaceutical companies during the model's development phase. Early results suggest a 30% reduction in time spent on initial compound screening.

Climate Modeling and Environmental Science

Climate scientists at several research institutions have begun testing Claude 4 for analyzing complex environmental datasets. The model can process large volumes of observational data, identify patterns across multiple variables, and generate hypotheses about causal relationships.

Extended Thinking is particularly valuable here because climate systems involve nonlinear interactions that require careful step-by-step reasoning rather than pattern matching.

Advanced Mathematics and Theorem Proving

Mathematicians report that Claude 4 can now assist with proof construction for problems at the graduate and early research level. The model's thinking traces effectively serve as proof sketches that human mathematicians can refine and formalize.

This capability puts Claude 4 in competition with specialized systems like DeepMind's AlphaProof, though Anthropic emphasizes that Claude 4 is a general-purpose model rather than a narrow theorem prover.

Pricing and Access: Premium Power at Premium Cost

Anthropic has structured Claude 4 pricing across 3 tiers, reflecting the computational intensity of Extended Thinking.

• Claude 4 Haiku: $1 per million input tokens / $5 per million output tokens (thinking tokens billed at output rate)
• Claude 4 Sonnet: $3 per million input tokens / $15 per million output tokens
• Claude 4 Opus: $15 per million input tokens / $75 per million output tokens

Thinking tokens — the internal reasoning generated during Extended Thinking — are billed at the output token rate. This means a complex scientific query that triggers 50,000 thinking tokens on Opus could cost approximately $3.75 for the reasoning alone, before the actual response.

For comparison, OpenAI's o3 charges roughly $60 per million output tokens at its standard tier. Google's Gemini 2.5 Pro comes in significantly cheaper at $10 per million output tokens but lacks the same depth of scientific reasoning capability.

Enterprise customers accessing Claude 4 through Amazon Bedrock or Google Cloud Vertex AI receive volume discounts. Anthropic also offers a research access program with subsidized pricing for academic institutions.

Industry Context: The Reasoning Race Intensifies

Claude 4's launch arrives during an intensely competitive period in the AI industry. The 'reasoning model' paradigm — where models spend more compute at inference time to solve harder problems — has become the primary battleground for frontier AI labs.

OpenAI kicked off this trend with o1 in September 2024 and followed up with o3 in early 2025. Google DeepMind responded with Gemini 2.5 Pro's built-in thinking capabilities. Now Anthropic enters the race with what appears to be the most transparent and scientifically focused implementation.

The strategic implications are significant. Reasoning capabilities unlock high-value enterprise use cases in pharmaceuticals, financial modeling, engineering, and scientific research — markets where customers will pay premium prices for reliable, verifiable AI outputs.

Anthropic's emphasis on transparency could prove to be a decisive competitive advantage. In regulated industries like healthcare and finance, the ability to audit an AI's reasoning process is not just a nice-to-have — it is increasingly a compliance requirement.

What This Means for Developers and Researchers

For developers building scientific applications, Claude 4's Extended Thinking opens up possibilities that were previously limited to specialized, narrow AI systems.

The key practical implications include:

• Reduced need for prompt engineering: Extended Thinking handles problem decomposition automatically, reducing the need for complex chain-of-thought prompting
• Verifiable outputs: Full thinking traces enable automated validation pipelines where downstream systems check the model's reasoning
• Configurable cost-performance tradeoffs: The adjustable thinking budget lets developers balance accuracy against latency and cost for each use case
• Seamless integration: The API is backward-compatible with Claude 3.5 endpoints, requiring only the addition of the thinking budget parameter

Researchers in particular should note that Anthropic is offering free API credits to qualifying academic teams through its research partnership program. Applications are open on the company's website.

Looking Ahead: What Comes Next

Anthropic's roadmap suggests that Extended Thinking is just the beginning of a broader push into scientific AI. The company has hinted at upcoming features including tool-integrated reasoning — where Claude 4 can call external calculators, code interpreters, and databases mid-thought — and collaborative thinking sessions where multiple Claude instances work on different aspects of a problem simultaneously.

The competitive landscape will likely shift rapidly in response. OpenAI is expected to announce GPT-5 later this year, and Google DeepMind continues to iterate on Gemini at a rapid pace. Meta's open-source Llama 4 models are also exploring reasoning capabilities, which could democratize access to this technology.

For now, Claude 4 with Extended Thinking sets a new standard for what general-purpose AI models can achieve in scientific domains. The combination of state-of-the-art benchmark performance, full reasoning transparency, and flexible deployment options makes it the most compelling choice for researchers and enterprises tackling the world's hardest problems.

Anthropic's bet is clear: the future of AI is not just about generating plausible text — it is about genuine reasoning. Claude 4 is the company's strongest argument yet that this future is already here.