Anthropic just released Claude Sonnet 4.5, and the performance numbers tell a story about what happens when you optimize relentlessly for one thing: building AI that developers actually want to use for real work. The model scores 77.2% on SWE-bench Verified—the most rigorous software engineering benchmark available—and can maintain coherent focus on complex, multi-step tasks for more than 30 hours straight. That's not a chatbot. That's a coworker who doesn't sleep, doesn't lose context, and doesn't need reminders about what you discussed six hours ago.
Anthropic commands 42% of the code generation market according to a Menlo Ventures survey of 150 enterprise technical leaders, more than double OpenAI's 21% share. That market dominance translated into a $5 billion revenue run rate earlier this year, making coding AI's first genuinely profitable use case beyond conversational interfaces. The company's success isn't accidental—it's the result of consistently shipping models developers prefer to work with, despite OpenAI's aggressive pricing and relentless product velocity.
Claude Sonnet 4.5 extends that lead with state-of-the-art performance across critical benchmarks. Beyond the 77.2% SWE-bench Verified score (reaching 82% with parallel test-time compute), the model achieves 50% on Terminal-bench and 61.4% on OSWorld, which tests real-world computer task performance. Just four months ago, Claude Sonnet 4 led OSWorld at 42.2%, demonstrating the rapid improvement curve Anthropic is sustaining. According to VentureBeat's analysis, these aren't marginal gains—they represent fundamental advances in AI's ability to understand code structure, maintain context across complex projects, and interact reliably with development environments.
The headline capability is Claude Sonnet 4.5's ability to maintain focus on complex, multi-step tasks for more than 30 hours. This isn't about token context windows or memory tricks—it's about genuine sustained reasoning across the kind of sprawling, iterative work that defines real software development. You can hand the model a refactoring task on Monday morning and check back Tuesday afternoon to find coherent, context-aware work that remembers decisions made 20 hours earlier.
This fundamentally changes how developers can delegate work to AI. Current generation models lose coherence after a few hours or require constant re-prompting to maintain context. Claude Sonnet 4.5 lets you structure work the way you'd assign it to a junior developer: here's the problem, here's the codebase, here are the constraints, go solve it and check back when you're done. For enterprise teams managing complex modernization projects, technical debt reduction, or large-scale refactors, this sustained focus eliminates the overhead of breaking work into AI-sized chunks and manually stitching results together.
Anthropic positions Claude Sonnet 4.5 as its "most aligned frontier model yet," with significant reductions in concerning behaviors like sycophancy, deception, and power-seeking tendencies. For procurement teams evaluating AI coding assistants, these aren't philosophical concerns—they're operational requirements. A model that agrees with bad ideas, hides errors to appear competent, or optimizes for metrics rather than actual code quality creates maintainability nightmares and security vulnerabilities.
The company reports "considerable progress on defending against prompt injection attacks," addressing one of the most critical security concerns for enterprise AI deployments. Claude Sonnet 4.5 ships with AI Safety Level 3 (ASL-3) protections, including classifiers designed to detect potentially dangerous inputs related to chemical, biological, radiological, and nuclear weapons. Anthropic has reduced false positives by a factor of ten since initially describing these safeguards, according to their announcement, making the security layer functional rather than obstructive.
Perhaps the most strategically significant release is the Claude Agent SDK—the same infrastructure powering Anthropic's Claude Code product. "We built Claude Code because the tool we needed didn't exist yet," the company stated in their announcement. "The Agent SDK gives you the same foundation to build something just as capable for whatever problem you're solving." This is Anthropic turning their internal tooling into developer infrastructure, letting enterprises build domain-specific coding assistants tuned for their particular technical stacks, workflows, and requirements.
The SDK addresses a fundamental limitation of general-purpose coding models: they optimize for average performance across broad tasks rather than specialized excellence in specific domains. With the Agent SDK, a fintech company can build a coding assistant that deeply understands their regulatory requirements, preferred architectural patterns, and internal libraries. A healthcare technology company can create tooling that navigates HIPAA compliance and medical data standards. The foundation is Claude Sonnet 4.5's reasoning and context capabilities; the specialization is whatever enterprises need for their particular problems.
Claude Opus 4 costs roughly seven times more per million tokens than GPT-5 for certain tasks according to VentureBeat's analysis, creating immediate pressure on Anthropic's premium positioning. Yet the company is maintaining pricing with Claude Sonnet 4.5 at $3 per million input tokens and $15 per million output tokens—the same as Sonnet 4. That's a bold move that only makes sense if performance advantages are compelling enough to overcome cost differentials.
Enterprise behavior suggests they are. Model API spending has more than doubled to $8.4 billion in just six months according to Menlo Ventures, and customer patterns show consistent prioritization of performance over price. Enterprises upgrade to the newest models within weeks of release regardless of cost, with 66% upgrading within existing providers rather than switching vendors. Winsome's research on enterprise AI procurement found that switching costs for coding assistants are higher than other AI tools because they integrate deeply into development workflows, IDE configurations, and team processes.
Anthropic's 42% market share reflects this dynamic. Developers and technical leaders choose Claude because it produces better code, maintains context more reliably, and integrates more smoothly into existing workflows—not because it's cheaper. In a market where model capabilities directly impact developer productivity and code quality, performance compounds over time. A coding assistant that's 15% more accurate or 20% better at maintaining context across complex tasks delivers exponentially more value than the cost differential suggests.
Industry analysis reveals that coding applications Cursor and GitHub Copilot drive approximately $1.4 billion of Anthropic's revenue, representing significant customer concentration. Critics flag this as a vulnerability—if either relationship falters, revenue takes a major hit. But this analysis misses the strategic picture. Cursor and Copilot chose Claude specifically because developers prefer working with it. That preference is upstream of the partnership deals. Anthropic doesn't win by locking in customers through contracts; they win by shipping models developers actively choose over alternatives.
"Our run-rate revenue has grown significantly, even when you exclude these two customers," an Anthropic spokesperson told VentureBeat, indicating diversification beyond the headline partnerships. The company is tripling its international workforce and expanding its applied AI team fivefold in 2025, driven by data showing nearly 80% of Claude usage now comes from outside the United States. That global expansion, combined with the Agent SDK enabling enterprises to build proprietary tooling on Claude infrastructure, suggests revenue diversification is accelerating.
The rapid-fire model releases—with Claude Sonnet 4.5 arriving just seven weeks after GPT-5's August launch—reflect intensifying competition that benefits everyone building software. Better performance, sustained context, improved security, and accessible infrastructure for custom tooling compound into fundamental workflow changes. The 30-hour work session capability isn't just longer context—it's the ability to structure development work around AI assistance rather than constantly working around AI limitations.
For enterprises, this translates directly into shipping velocity and code quality improvements. Technical debt reduction projects that would take quarters with manual refactoring become weeks with sustained AI assistance. Modernization initiatives that require understanding sprawling legacy codebases become tractable when you can hand context to a model that maintains coherence across the entire codebase. Security audits, documentation generation, test coverage expansion—all the work that developers know they should do but rarely have time for—become feasible when AI can sustain focus longer than humans can maintain attention.
We help engineering leaders integrate AI coding assistants into development workflows without creating new technical debt or security vulnerabilities—where to deploy autonomous capabilities, how to structure human oversight, and when sustained AI context actually improves outcomes versus creating maintenance burdens. If you're figuring out what Claude Sonnet 4.5's capabilities mean for your development velocity and code quality, let's talk.