Claude Opus 5 Effort Levels in Claude Code: Cut Your API Bill Without Losing Quality (2026)

Claude Opus 5 (launched July 24, 2026) costs the same $5 per million input tokens and $25 per million output tokens as Opus 4.8 — but its new effort setting (low/medium/high) can cut real token usage by 20-30% without touching that per-token rate, and stacking it with prompt caching and a cheaper relay endpoint can cut your effective Claude Code bill by 80%+.

Same Sticker Price, Different Bill

The headline from Anthropic's July 24 release was simple: Opus 5 is a drop-in upgrade over Opus 4.8 at identical rates ($5/$25 per million tokens), and it approaches Claude Fable 5's frontier intelligence at half Fable's price ($10/$25 vs $10/$50). What the headline number doesn't tell you is that Opus 5 runs adaptive thinking on by default, and thinking tokens bill at the output rate. If you just swap your model string from claude-opus-4-8 to claude-opus-5 inside Claude Code and change nothing else, your monthly invoice can quietly go up even though the per-token price didn't move.

The Effort Setting Is the Actual Cost Lever

Opus 5 introduces a low/medium/high effort parameter that doesn't touch the $5/$25 rate at all — it changes how many tokens the model burns to reach a given quality bar. Early adopters, including one legal-AI team, reported similar output quality using roughly 26% fewer tokens at a lower effort setting compared to Opus 4.8 running at max reasoning. For everyday Claude Code work — refactors, test writing, routine bug fixes — dropping from high to medium effort is usually invisible in output quality but very visible in the bill.

In Claude Code, you can set this per-session or via config:

// settings.json
{
  "model": "claude-opus-5",
  "effort": "medium"
}

# or per-invocation
claude --model claude-opus-5 --effort medium "refactor this module"

Reserve high effort for genuinely hard, multi-file architectural work, and default everything else to medium or even low. Treat effort the way you'd treat a linter setting: strict when it matters, relaxed for routine passes.

Stack Effort With Prompt Caching and Batch

Effort level is only one lever. Claude Code already auto-caches your CLAUDE.md, system prompt, and repeated file context, and that caching is where the biggest single savings live. Cache reads cost roughly 10% of the standard input rate — for Opus 5 that's $0.50/M against a $5/M base rate, a 90% discount on every repeated block. A long-running Claude Code session with an 8,000-token CLAUDE.md across 100 turns costs about $2.40 without caching and roughly $0.27 with it, an 89% reduction on that portion of the bill alone.

None of these levers are mutually exclusive. A team running medium-effort Opus 5 with caching enabled and occasional batch runs for eval work can realistically end up paying a fraction of the naive high-effort, no-cache sticker price — before even touching which endpoint serves the requests.

The Fourth Lever: Which Endpoint You Call

The three levers above assume you're calling Anthropic directly, which in turn assumes a US-billed credit card and no easy way to combine Claude with GPT-5.6 or Gemini under one key. That's the gap a relay like Safa API (aisafa.xyz) is built for: official, unthrottled routing to Claude, GPT and Gemini through a single endpoint, at lower per-token pricing than going direct, with prompt caching fully supported so the 90% cache-read discount above still applies. No credit card is needed — Alipay works directly — which matters if you're outside the US or just don't want to hand a foreign card to another SaaS. Point Claude Code's ANTHROPIC_BASE_URL at Safa API, keep your effort/caching setup exactly as described above, and the same cost math just runs at a cheaper base rate.

Frequently Asked Questions

Does Claude Opus 5 cost more than Opus 4.8 in Claude Code?

Per-token, no — both are $5/$25 per million tokens. In practice your bill can rise because Opus 5 runs adaptive thinking by default, which Opus 4.8 didn't. Setting effort to medium or low restores a spend profile closer to what you saw on Opus 4.8.

What effort level should I use for daily coding work?

Start at medium for most Claude Code sessions — routine refactors, tests, and bug fixes rarely need high effort. Reserve high for genuinely hard, multi-file architecture problems where you're willing to pay for the extra tokens.

Can I combine effort tuning with a cheaper API relay?

Yes. Effort level, prompt caching, and your choice of endpoint are independent levers. Running medium effort with caching through a lower-priced, Alipay-friendly relay like Safa API stacks all three savings instead of forcing a choice between them.