Claude Code costs Seahawk Media roughly 1,200 to 2,400 USD per engineer per month at production usage in 2026, on Anthropic API pricing. That is the actual line-item cost, not the published per-token figure multiplied by an estimate. This post is the honest breakdown of what we spend, what drives the variance, and how the cost structure shapes how we deploy Claude Code across the team.
I run an agency that has been on Claude Code daily for over a year. If you are weighing whether agentic AI tools are economically viable at agency scale or planning the budget conversation with your CFO, this is the operator answer.
The cost structure, line by line
API token costs
The dominant line item. Claude Sonnet on the API runs at roughly 3 USD per million input tokens and 15 USD per million output tokens. A senior engineer using Claude Code intensively burns roughly 8 to 16 million tokens per day across all sessions. That is 25 to 75 USD per day in API costs alone, or 500 to 1,500 USD per month per engineer at 20 working days.
Claude Pro / Max subscriptions for non-API use
We pay Anthropic 200 USD per month per Pro subscription for engineers who use the chat interface alongside Claude Code. About a third of the team has it; the rest work API-only. This adds roughly 200 to 400 USD per month per Pro user.
MCP infrastructure
The MCP servers themselves are mostly free or per-use cheap (Brave Search at a few dollars a month, the rest unmetered). Aggregate MCP cost across the agency runs under 100 USD per month total.
Subagent and orchestration overhead
The agentic features (Tasks, swarms, multi-agent flows) burn tokens faster than single-session use. A single complex task that Claude executes autonomously can cost 5 to 15 USD on its own. We use this sparingly.
Mistakes and retries
Claude occasionally goes off the rails and burns thousands of tokens on a wrong path. We estimate 10 to 15 percent of total token spend is wasted on retries and corrections. Worth budgeting for.
Real monthly numbers from our engineering team
Anonymised but real numbers from our last billing cycle:
Senior engineer A (heavy Claude Code user, 8-hour daily sessions): 2,140 USD in API + 200 USD Pro subscription = 2,340 USD.
Senior engineer B (moderate use, mostly code review and refactors): 1,180 USD API + 200 USD Pro = 1,380 USD.
Mid-level engineer C (Claude Code primary, heavy on research and discovery): 980 USD API only = 980 USD.
Senior engineer D (light Claude Code, mostly Cursor with occasional Claude sessions): 410 USD API + 200 USD Pro = 610 USD.
Average across the engineering team: roughly 1,300 USD per engineer per month all-in. Range 500 to 2,400 USD depending on usage intensity.
What drives variance up
Long-running agentic tasks
Tasks that run autonomously for hours can burn thousands of tokens. The output is sometimes worth it; the cost line item is the most volatile.
Large codebases loaded into context
Working in a 200K-line monorepo against Claude burns more input tokens per turn than working in a 5K-line module. Project size scales token spend roughly linearly.
Heavy discovery and research
Engineers using Claude for codebase exploration and architecture research burn more tokens than engineers using Claude only for targeted code generation. The discovery work is often worth it; just be aware of the cost.
What drives cost down
Tight CLAUDE.md and prompts
A 400-line agency CLAUDE.md plus a clear prompt produces better output in fewer turns. Engineers who write disciplined prompts spend dramatically less than engineers who chat their way to the answer.
Caching where applicable
Anthropic prompt caching (when used correctly) cuts repeat input-token costs by 80 to 90 percent on multi-turn sessions. Worth setting up; we run it on most of our long-running workflows.
Right-sized model selection
Sonnet 4.6 for most work; Haiku for simple completions; Opus only when the task genuinely warrants it. Engineers who default to Opus burn 5x what engineers who default to Sonnet do, often without commensurate output quality.
How the cost compares to alternatives
Three reference points:
A senior engineer's loaded cost (salary plus benefits plus overhead) at Seahawk runs roughly 8,000 to 14,000 USD per month. Adding 1,300 USD of Claude Code on top is a 10 to 15 percent cost increase for a productivity gain we measure at 40 to 60 percent on greenfield work and 25 to 40 percent on maintenance work. The math is good.
GitHub Copilot Pro is 19 USD per month per user. It does dramatically less than Claude Code does. The 1,300 USD is for the agentic, multi-step, repository-wide work Copilot cannot do; the comparison is not really fair.
Cursor at the Pro tier is 20 USD per month per user with bundled API costs. We use Cursor alongside Claude Code; they are complements, not substitutes. Combined cost is roughly 1,400 USD per engineer per month.
Will the cost come down?
Anthropic API pricing has come down roughly 40 percent over the last 18 months on Sonnet-tier models. The trend is clearly downward. By Q4 2026 I expect the same usage to cost 700 to 1,400 USD per engineer per month rather than 1,200 to 2,400. The productivity gain stays roughly constant; the cost amortisation improves.
The agencies that adopt the workflow now (rather than waiting for cheaper pricing) lock in the productivity advantage and price their services accordingly. Waiting is a real cost too.
Bottom line
Claude Code at agency scale costs 1,200 to 2,400 USD per engineer per month all-in on Anthropic API plus subscriptions in 2026. The productivity gain is 25 to 60 percent. The math works comfortably.
Budget for this line item explicitly rather than burying it under "tools" expense. Track per-engineer spend monthly so you spot the engineers whose Claude usage is bringing genuine velocity versus the ones whose token spend is mostly retry-cost. The transparency is the management lever.