Claude Code Session Costs: Opus 5.5 Median $0.31, Sonnet 5.5 $0.16

What the median sessions cost

The figures come from real Claude Code sessions on OpenRouter in the 30 days to October 18, 2026. OpenRouter says Claude Code was the top app in its daily app ranking on October 21, so there is a lot of usage behind the numbers.

For long sessions of 10 to 49 turns, the median is $1.39 on Opus 5.5 and $0.60 on Sonnet 5.5. Claude Haiku 5.5 comes in under a cent for a 2-to-9-turn session. OpenRouter's table lists medians by session length:

• Claude Fable 5.1: $0.342 for 1 turn, $1.049 for 2 to 9 turns, $3.734 for 10 to 49 turns
• Claude Opus 5.5: $0.136, $0.305, $1.390
• Claude Sonnet 5.5: $0.071, $0.155, $0.596
• Claude Haiku 5.5: $0.002, $0.007, $0.053
• Kimi K3: $0.040, $0.117, $0.825
• GLM-5.3: $0.010, $0.044, $0.363
• DeepSeek V4.1 Flash: $0.003, $0.009, $0.047
• GPT-6 Luna: $0.002, $0.005, $0.024

According to OpenRouter, the model choice has the biggest effect on cost. A one-turn Fable 5.1 session costs more than a long Haiku 5.5 session. Within a model, length matters most: a 10-to-49-turn Opus session costs about ten times a one-turn session.

Why a single turn costs 14 cents

Most of the cost is Claude Code's own context, not the user's prompt. OpenRouter ran claude -p "Reply with exactly: ok" against Opus 5.5 in an empty folder on Claude Code 2.1.307. The reply was four tokens, yet the session wrote 17,047 tokens to the prompt cache. These are the system prompt and tool definitions. At Opus 5.5's cache-write price of $5 per million tokens, that came to $0.085.

A real project adds its own instructions and the files Claude Code reads, which explains the dataset's one-turn median of $0.136. Once the context is cached, later requests read it back at $0.20 per million tokens, 25 times cheaper than writing it. Costs rise again as the context grows, because each new file is written to the cache once and then read on every following turn.

OpenRouter's own bug-fix test

To compare models on a controlled task, OpenRouter gave Claude Code a small invoicing library with three bugs and five failing tests. It ran each model three times, and every run fixed all three bugs.

• Opus 5.5: $0.098 median per session, 5 turns
• Sonnet 5.5: $0.062, 5 turns
• Haiku 5.5: $0.0052, 6 to 8 turns

These costs sit below the dataset's 2-to-9-turn medians because the task was small and each model's three runs shared a warm cache. For a fresh session on a real project, OpenRouter says the dataset medians are the better guide. Sonnet cost about two thirds of Opus, and Haiku about a twelfth of Sonnet.

OpenRouter also warns against relying on Claude Code's own total_cost_usd. It uses Anthropic's built-in price table and guesses when it does not recognize an OpenRouter model name. In the company's runs it reported $0.16 to $0.31 for Haiku sessions that cost about half a cent, and overstated Opus sessions by about 50%. The Activity page on OpenRouter shows actual spend.

Monthly estimates and ways to cut the bill

Using the dataset medians and 22 workdays a month, OpenRouter estimates the following monthly costs on Opus 5.5 and Sonnet 5.5:

• Light (5 short sessions a workday): $34 and $17
• Regular (20 short sessions a workday): $134 and $68
• Heavy (10 long sessions a workday): $306 and $131

These are estimates, and real costs depend on codebase size and how much of it Claude Code reads. OpenRouter recommends making Sonnet the default, which it says roughly halves the bill for Opus-only users, and sending background jobs and sub-agents to Haiku 5.5. It also advises keeping sessions short: a 10-to-49-turn session costs about four to five times a 2-to-9-turn one. Starting a fresh session or running /compact when switching tasks avoids carrying old context along.

Claude Code can also call models from other labs through OpenRouter. According to the company, GLM-5.3 and Kimi K3 sessions cost less than Opus sessions at every length. It cautions that Claude Code is built for Anthropic models, so other models should be tested on real tasks first. Users can also set a credit limit on an API key so a runaway session stops at a chosen amount.

Running Claude Code through OpenRouter

OpenRouter's suggested route is Ori Harness, which starts the Claude Code already installed on a machine with OpenRouter settings and signs in through the browser. The commands are curl -fsSL https://openrouter.ai/labs/ori/install.sh | bash, then ori claude --model anthropic/claude-sonnet-5.5. Only sessions started with ori claude run on OpenRouter, while plain claude keeps using a subscription.

For developers paying per token, the data gives a rough budget by model and session length, and it shows that model choice and session length are the two levers that matter most.