Claude Code Session Costs and OpenAI's GPT-5.6 Price Drop

Two recent announcements approach AI cost from different directions.

Anthropic published a practical guide to maximizing the value of Claude Code sessions. OpenAI reduced the API prices of GPT-5.6 Luna and Terra.

The shared message is more useful than either price list: measure the cost of a successful task, not the price of one token.

Anthropic focuses on session hygiene

Anthropic's guidance explains how a coding session repeatedly sends its growing context back to the model. Files, command output, tool definitions, and conversation history all contribute.

The practical recommendations include:

• run /clear between unrelated tasks
• choose the model and effort level before starting
• mention files directly instead of making the agent search for them
• keep noisy command output out of the main context
• inspect loaded context in a fresh session
• compact before a long break while the prompt cache is still useful

The larger lesson is that context has carrying cost. A long session can feel convenient while making each later turn more expensive and less focused.

OpenAI cut model prices, not ChatGPT subscriptions

OpenAI's July 30 pricing announcement reduced GPT-5.6 Luna API pricing by 80 percent and Terra by 20 percent.

Luna is now listed at $0.20 per million input tokens and $1.20 per million output tokens. Terra is $2 per million input tokens and $12 per million output tokens.

This is sometimes described as a ChatGPT price drop. That is imprecise. ChatGPT and Codex subscription prices and quota budgets did not change. The cheaper models consume fewer credits in applicable paid products.

Cheap tokens can still produce expensive work

A lower-cost model can be the best option for clear, repetitive tasks. It can also become expensive if it requires several retries, extensive review, or a more capable model to repair the result.

Track:

• total model cost per accepted task
• time spent reviewing and correcting
• retry count
• latency to a verified result
• defects found after acceptance

This turns model routing into an engineering decision. Use a stronger model when uncertainty is the expensive part. Use a lower-cost model when the plan and verification are already clear.

A practical routing pattern

A team might use a frontier model to investigate an ambiguous bug and write a precise plan. A faster model can then implement bounded changes, run tests, and complete mechanical follow-up.

The split only works if the handoff contains enough context and the verification is real. Otherwise the apparent savings move into rework.

Cost discipline is not about asking agents to think less. It is about spending intelligence where it changes the outcome.