Anthropic holds 54% of the enterprise coding market, which is more than double the 21% share held by OpenAI. This dominance follows the release of Claude 3.5 Sonnet, a model that ranks first in coding for 18 consecutive months according to Menlo Ventures. The model improved its SWE-bench Verified score from 33.4% to 49.0% and achieved a jump from 62.6% to 69.2% in the TAU-bench retail domain. While it outperforms OpenAI’s o1-preview in reasoning tasks, the model costs $3.00 per million input tokens and $15.00 per million output tokens.
| Metric | Claude 3.5 Sonnet Value |
|---|---|
| Input price per 1M tokens | $3.00 |
| Output price per 1M tokens | $15.00 |
| Blended 3:1 rate per 1M tokens | $6.00 |
| Context window | 200,000 tokens |
Companies like GitLab report improved reasoning and lower latency using the upgraded model. Cognition and The Browser Company use the model for automation in web-based workflows. Anthropic also launched a "computer use" feature in public beta that lets developers instruct Claude to move a cursor, click buttons, and type text. Replit uses this feature to improve development workflows. The feature scored 14.9% on OSWorld’s screenshot-only tasks during early evaluations. Anthropic acknowledges that this technology remains in an experimental phase and advises developers to use it for low-risk tasks.
Coding and reasoning drive market share
Anthropic captured 40% of enterprise AI spending in 2025, rising from 24% in 2024 and 12% in 2023. This growth contradicts the earlier dominance of OpenAI, which held 50% of the enterprise market in 2023 but fell to 27% in 2025. Enterprise spending on large language models grew 180% in 2025 to an average of $7 million per company. Most enterprises prefer proprietary models over open-source alternatives, as open-source models held only 11% of the AI market in 2025. You should note that Claude 3.5 Sonnet remains more expensive than ChatGPT-4o Latest on a per-token basis, with ChatGPT-4o Latest costing $2.50 per million input tokens and $10.00 per million output tokens.
The Claude 3.5 Haiku model provides an alternative for users needing speed and lower costs. This model surpasses the previous Claude 3 Opus in coding tasks and works well for specialized sub-agent tasks. Anthropic will make Haiku available on its API, Amazon Bedrock, and Google Cloud’s Vertex AI later this month. Because Haiku handles classification and routing efficiently, it serves as a cheaper option for high-volume tasks. If a company sends 100 million input tokens and 25 million output tokens to Claude 3.5 Sonnet, the total cost reaches $675.
Security and governance in agentic workflows
Agentic SecOps uses AI to execute investigation steps inside the security stack, which shifts the control problem toward permission scope and auditability. LimaCharlie reports that Claude Code reduces SOC setup from days to approximately ten minutes by automating organization creation and sensor deployment. This automation uses conversational prompts to handle detection-rule translation. However, the agent acts using the authority of a human or service identity, which makes identity and authorization controls necessary. Organizations must treat prompt-driven SOC setup as a privileged workflow that requires scoped permissions and detailed logging.
Security teams face specific risks when using agentic interfaces for infrastructure. A researcher showed that a fake Sentry error could make AI coding agents run attacker code using developer credentials. Additionally, the Anthropic GTG-1002 AI espionage campaign involved a state-sponsored group using Claude Code agents to attack 30 organizations and harvest credentials. Effective governance requires that teams place detection content under the same review and change-control discipline as any other security policy. Can teams move fast enough to maintain these review gates without losing the speed gained from automation?
Managing AI risk for SOC 2 compliance requires constant monitoring because machine learning changes after launch. A model might drift because user behavior changes or data shifts, meaning a successful test in March does not guarantee the same performance in June. Auditors look for version history, approval logs, and prompt records to verify that AI systems maintain security. Access control must cover the entire AI lifecycle, including training data, prompt stores, and deployment pipelines. If an AI feature touches personal data, companies must follow the NIST Privacy Framework to keep that data clean.
