Anthropic launched Claude Sonnet 4.5 on Monday, positioning the artificial intelligence model as “the best coding model in the world” in a direct challenge to OpenAI’s recently released GPT-5, as the two AI giants battle for dominance in the lucrative enterprise software development market.
The San Francisco-based startup claims its newest model achieves state-of-the-art performance on critical coding benchmarks, scoring 77.2% on SWE-bench Verified — a rigorous software engineering evaluation — compared to GPT-5’s performance. More remarkably, Anthropic says Claude Sonnet 4.5 can maintain focus on complex, multi-step tasks for more than 30 hours, a dramatic leap in AI’s ability to handle sustained work.
“Sonnet 4.5 achieves 77.2% on SWE-bench Verified (82% with parallel test-time compute). It is SOTA,” an Anthropic spokesperson told VentureBeat, using industry shorthand for “state of the art.” The company also highlighted the model’s 50% score on Terminal-bench, another coding benchmark where it claims leadership.
The announcement follows mounting pressure from OpenAI’s recent advances and pointed criticism from high-profile figures like Elon Musk, who recently posted on X.com that “winning was never in the set of possible outcomes for Anthropic.” When asked about Musk’s statement, Anthropic declined to comment.
The release arrives just seven weeks after OpenAI’s GPT-5 launch in August, underscoring the breakneck pace of competition in artificial intelligence as companies race to capture enterprise customers increasingly relying on AI for software development. The timing is particularly noteworthy as Anthropic grapples with questions about its heavy dependence on just two major customers.
Anthropic dominates coding market despite customer concentration risks
The competition centers on a market that has emerged as AI’s first major profitable use case beyond chatbots. Anthropic commands 42% of the code generation market — more than double OpenAI’s 21% share — according to a Menlo Ventures survey of 150 enterprise technical leaders. That dominance has translated into remarkable financial performance, with the company reaching a $5 billion revenue run rate earlier this year.
However, industry analysis reveals that coding applications Cursor and GitHub Copilot drive approximately $1.4 billion of Anthropic’s revenue, creating a potentially dangerous customer concentration that could leave the company vulnerable if either relationship falters.
“Our run-rate revenue has grown significantly, even when you exclude these two customers,” the Anthropic spokesperson said, pushing back on concerns about customer concentration. The company provided supportive quotes from both Cursor CEO Michael Truell and GitHub Chief Product Officer Mario Rodriguez praising Claude Sonnet 4.5’s performance.
The new model achieves significant advances in computer use capabilities, scoring 61.4% on OSWorld, a benchmark that tests AI models on real-world computer tasks. Just four months ago, Claude Sonnet 4 held the lead at 42.2%, demonstrating rapid improvement in AI’s ability to interact with software interfaces.
OpenAI’s aggressive pricing strategy threatens Anthropic’s premium positioning
Anthropic’s announcement comes as the company grapples with competitive pressure from GPT-5’s aggressive pricing strategy. Early analysis shows Claude Opus 4 costing roughly seven times more per million tokens than GPT-5 for certain tasks, creating immediate pressure on Anthropic’s premium positioning.
The pricing disparity signals a fundamental shift in competitive dynamics that could force enterprise procurement teams to reconsider vendor relationships previously built on performance rather than price. Companies managing exponentially growing AI budgets now face comparable capability at a fraction of the cost.
Yet Anthropic is maintaining its pricing strategy with Claude Sonnet 4.5. “Sonnet 4.5’s cost remains the same as Sonnet 4,” the spokesperson confirmed, keeping prices at $3 per million input tokens and $15 per million output tokens.
Claude Sonnet 4.5 delivers 30-hour autonomous work sessions and enhanced security
Beyond performance improvements, Anthropic positions Claude Sonnet 4.5 as its “most aligned frontier model yet,” showing significant reductions in concerning behaviors like sycophancy, deception, and power-seeking tendencies. The company has made “considerable progress on defending against prompt injection attacks,” a critical security concern for enterprise deployments.
The model is being released under Anthropic’s AI Safety Level 3 (ASL-3) protections, which include classifiers designed to detect potentially dangerous inputs and outputs related to chemical, biological, radiological, and nuclear weapons. While these safeguards sometimes flag normal content, Anthropic says it has reduced false positives by a factor of ten since initially describing them.
Perhaps most significantly for developers, Anthropic is releasing the Claude Agent SDK — the same infrastructure that powers its Claude Code product. “We built Claude Code because the tool we needed didn’t exist yet,” the company said in its announcement. “The Agent SDK gives you the same foundation to build something just as capable for whatever problem you’re solving.”
International expansion accelerates as $1.5 billion copyright settlement finalizes
The model launch coincides with Anthropic’s aggressive international expansion, as the company seeks to diversify beyond its U.S.-concentrated customer base. The startup recently announced plans to triple its international workforce and expand its applied AI team fivefold in 2025, driven by data showing that nearly 80% of Claude usage now comes from outside the United States.
However, the expansion comes amid significant legal costs. Anthropic recently agreed to pay $1.5 billion in a copyright settlement with authors and publishers over allegations the company illegally used their books to train AI models without permission. The settlement, approved by a federal judge last week, requires payments of $3,000 for each publication listed in the case.
Enterprise AI spending doubles as companies prioritize performance over cost
The rapid-fire model releases from both companies reflect the high stakes in enterprise AI adoption. Model API spending has more than doubled to $8.4 billion in just six months, according to Menlo Ventures, as enterprises shift from experimental projects to production deployments.
Customer behavior patterns suggest enterprises consistently prioritize performance over price, upgrading to the newest models within weeks of release regardless of cost. This behavior could work in Anthropic’s favor if Claude Sonnet 4.5’s performance advantages prove compelling enough to overcome GPT-5’s pricing advantage.
However, the dramatic price differential introduced by GPT-5 could overcome typical switching inertia, especially for cost-conscious enterprises facing budget pressures. Industry observers note that model switching costs remain relatively low, with 66% of enterprises upgrading within existing providers rather than switching vendors.
For enterprises, the intensifying competition delivers better performance and lower costs through continuously improving capabilities. The rapid pace of model improvements — with new versions launching monthly rather than annually — provides organizations with expanding AI capabilities while vendors compete aggressively for their business.
While the corporate rivalry between Anthropic and OpenAI dominates industry headlines, the real economic impact extends far beyond Silicon Valley boardrooms. The development of AI systems capable of sustained coding work for 30 hours represents a fundamental shift in how software gets built, with implications that extend across every industry relying on technology infrastructure.
These advancing capabilities signal broader workplace transformation ahead. As AI systems demonstrate increasing proficiency at complex, sustained intellectual work, the technology industry’s competition for coding supremacy foreshadows similar disruptions across fields requiring analytical thinking, problem-solving, and technical expertise.
Subscribe to get latest news!
Deep insights for enterprise AI, data, and security leaders
VB Daily
AI Weekly
AGI Weekly
Security Weekly
Data Infrastructure Weekly
VB Events
All of themGet updates
Partner Content
Why identity-first security is the first defense against sophisticated AI-powered social engineering
Enterprise security is having an identity crisis. Attackers aren’t going after zero-day exploits on a server or an operating system; instead, the vast majority of security breaches are happening in a surprisingly low-tech wave of identity compromise via social engineering.
“Con men, and social engineering, have been around for a long time,” says Matt Caulfied, VP of product, identity at Cisco. “The oldest trick in the book is sneaking in by putting on a construction vest and walking in the front door, and this is essentially the same thing. You trick someone into giving you access to their account, and use it to get all the access that they have, as far as you can go.”
Consider spearphishing which once meant laboriously researching a few high-value targets. With AI, attackers can generate target lists, identify those targets’ nearest relatives, and fire off convincing emails and texts at scale — multiplying their odds, even for non-native speakers without strong language skills.
However, there’s a clear disconnect between awareness and execution in the enterprise. Cisco Duo’s 2025 State of Identity Security report found that 51% of organizations have suffered financial losses from identity-related breaches. So why do 74% of IT leaders admit that identity security is an infrastructure-planning afterthought?
“It’s a fundamentally hard problem to solve,” Caulfield says. “Identity security is unique in that it combines social aspects, and a psychological aspect, with a technical aspect. Over time, just as their targets get better at defending themselves, attackers get better at attacking their targets. And while we know how to prevent identity breaches entirely, most of those mechanisms have been incredibly expensive and difficult to scale, from an operational perspective.”
But strong identity and access management (IAM) is no longer optional — it must actually be the foundation of enterprise security, rather than just one pillar, especially as AI agents gain a foothold in organizations as a third class of users, without any of the restraints or guardrails that humans presumably have.
A new definition of zero trust
Today you can’t trust users just because they’re on the network, or coming from a corporate device; you can only establish trust through strong cryptographic identity authentication. That shifts trust from the network over to identity systems that authenticate the user. And since a zero-trust system is just going to enforce what the identity system tells it to, identity has to be the foundation of an enterprise security process — keeping systems safe, humans from being hijacked, and AI agents performing only the actions they’re meant to take.
If that authentication and authorization step is wrong, then it doesn’t matter how good your network access control is. However, traditional second-factor and multi-factor authentication is no longer enough, since an SMS message, call-back number or even a verified push notification can all be hacked.
“Only one in three leaders trust their current identity providers to stop identity-based attacks. Just because you’re doing identity doesn’t mean you’re doing identity securely,” Caulfield explains. ” Phishing-resistant authentication is the new gold standard, where a user cannot be tricked into giving away the keys to the kingdom. They would need to literally be at your desk with you, while you’re using your laptop, in order to take over your account.”
However, until now, phishing-resistant MFA approaches have either been too complex or too expensive to implement. While 87% of leaders believe phishing-resistant MFA is critical to a security strategy, only 19% of companies have deployed FIDO2 tokens, which are a standard way to achieve phishing-resistant MFA. Hardware tokens are often reserved for privileged users, adoption often stalls out here due to token management complexity (what happens when a token is lost, for example?), the expense and complications of training, and just the cost of creating and distributing a hardware solution.
Security as an enabler
Awareness of identity security is growing, Caulfield adds, with 82% of financial decision-makers increasing budgets for identity security. But security can’t be treated as an add-on, because that results in tool sprawl, which adds additional costs, complexity, and misalignment, along with decreased visibility overall. To address that head-on, 79% of leaders are exploring identity vendor consolidation, which massively cuts down the operational drag of tool proliferation.
Integrated tools that offer interoperability in multi-cloud environments offer strategic simplification that not only reduces costs and increases security, but improves organizational efficiency for IT and end users.
“Identity management and security is not just a necessary evil, it’s an enabler for a workforce and for customers interacting with a business. It’s as much a security concern as it is a productivity and IT concern,” he says. “Phishing-resistant authentication is that easy button to get to the identity-first approach to security that makes it work.”
Learn how Duo and Cisco Identity Intelligence are helping global teams make sense of the complex identity landscape: Download Cisco Duo’s report, The 2025 State of Identity Security: Challenges and Strategies from IT and Security Leaders.
