Anthropic Lowers Cost, Tightens API Access
Anthropic is gaining enterprise adoption with Claude Sonnet 4.6, a model that offers the intelligence of its flagship Opus model at one-fifth the cost. Concurrently, the company has banned the use of its OAuth tokens in third-party tools, enforcing the policy with server-side blocks. The move highlights a trend of platform providers making AI more affordable while simultaneously increasing control over their ecosystems.
- On software engineering benchmarks, Claude Sonnet 4.6 scores 79.6% on SWE-bench, nearly matching the flagship Opus 4.6's score of 80.8% while being five times cheaper. For tasks involving expert-level scientific reasoning, however, Opus maintains a significant lead, scoring over 17 points higher on the GPQA Diamond benchmark. - In early testing, developers preferred the new, cheaper Sonnet 4.6 over the *previous flagship model*, Opus 4.5, in 59% of cases, citing that it was less prone to over-engineering and better at following instructions. - The ban on OAuth tokens directly impacts third-party tools like OpenClaw and Cline, which allowed users to leverage their flat-rate consumer subscriptions for high-volume tasks. Developers using these tools must now switch to more expensive per-token API keys, ending a popular cost-saving "loophole." - The move is part of a larger strategy to control the ecosystem, similar to moves by other platform providers. Anthropic is also pushing for industry standards like its Model Context Protocol (MCP), designed to standardize how AI agents connect to enterprise systems. - This enterprise focus has led to significant growth, with Anthropic's annualized revenue reportedly jumping from $1 billion in January 2025 to $14 billion in January 2026. - Major systems integrators are driving enterprise adoption; for instance, Cognizant is deploying Claude to its 350,000 associates to help clients scale their use of AI. - Anthropic recently updated its `claude-code` command-line tool to hide the specific file names being accessed by default, a move that drew criticism from developers who rely on that information for debugging and security.