Four ways to cut AI costs in 2026

Claim

Both labs give the same cost advice: stable prefix first, append-only turns, never switch models or tools mid-session, batch anything that can wait. Routing is done inside the harness by giving each subagent its own model, not by an external router. The only first-party agent cost figures are Claude Code's.

Evidence

• Anthropic list prices (October 2026, per MTok in and out): Opus 5.5 $4 and $20, Sonnet 5.5 $2 and $10, Haiku 5.5 $0.10 and $0.50 up to 100k prompt tokens then 5x, Fable 5.1 $10 and $50. Cache reads 0.1x (0.05x on Opus and Sonnet 5.5, 0.025x on Fable 5.1), writes 1.25x for 5 minutes or 2x for 1 hour; caching pays after one read at 5 minutes or two at 1 hour. Batch is 50% off and stacks with caching. 1M context at standard price for 4.6-plus except Haiku 5.5.
• OpenAI (2026-10-20): prompt_cache_options.ttl of 30 minutes or 12 hours (writes 1.25x or 2.5x, reads 0.1x, 0.05x on GPT-6.1 Sol), up to four explicit breakpoints, prewarm, cache diagnostics; minimum cacheable prefix 1,024 tokens; keep tool definitions and order stable and restrict with allowed_tools rather than removing tools.
• Claude Code costs page: about $13 per developer per active day, $150 to $250 per month, 90% under $30 per day; agent teams about 7x tokens; /usage shows cache share and the likely cause of the last miss; hooks that pre-filter tool output cut "tens of thousands of tokens to hundreds"; keep CLAUDE.md under 200 lines; --max-budget-usd; per-subagent model: haiku; compaction is a large request while /clear is free.
• Experiment numbers: harnessed app build $200 over 6 hours versus $9 solo; DAW build $124.70 with a $0.46 planner; C compiler just under $20,000. Third-party "$ per SWE-bench task" and "60% to 80% routing savings" claims are unsupported. UNVERIFIED.

Open questions

• Whether Batch applies to Managed Agents sessions (secondary says no).

Sources

• https://platform.claude.com/docs/en/about-claude/pricing — docs, Wed Oct 21
• https://developers.openai.com/api/docs/guides/prompt-caching — docs, Mon Oct 19
• https://code.claude.com/docs/en/costs — docs, Wed Oct 21
• https://www.anthropic.com/engineering/harness-design-long-running-apps — blog, Mon Mar 23

AI-assisted research: gathered and summarised with AI from the sources above, each claim checked against them.