Claude API Pricing (July 2026)

Before comparing tiers, turn on the discounts. Prompt caching bills cache reads at roughly 10% of the standard input rate — on Claude Fable 5 Anthropic documents a 90% input-token discount explicitly. If your application repeats a system prompt, tool definitions, or a document prefix across requests, the cached share of your input bill drops by an order of magnitude. Cache writes cost 1.25x standard input, so caching pays for itself after the first reuse.

Batch processing takes a flat 50% off both input and output for asynchronous workloads — nightly document runs, bulk classification, evals. The two levers stack: a cached, batched Sonnet 5 workload can land near Haiku list price while keeping Sonnet quality. Model both with the cost calculator.

Claude Code has no separate API price list — it consumes standard Claude API tokens. When you run Claude Code with an API key, every request bills at the per-token rate of the model it uses: Sonnet 5 at $3/$15, Opus 4.8 at $5/$25, Haiku 4.5 at $1/$5, or Fable 5 at $10/$50 per million input/output tokens.

Two things make agentic coding sessions cheaper than the headline rates suggest. First, Claude Code caches aggressively — long sessions reuse the same repository context, so most input tokens bill at the ~10% cache-read rate. Second, subscription plans (Pro/Max) include Claude Code usage without per-token billing; the API path is for CI, automation, and teams that want usage-based costs. For a typical active coding session, budget a few dollars per hour on Sonnet-class models, several times that on Opus or Fable — and check the calculator with your real token volumes.

Flagship-for-flagship, the two providers now mirror each other: Claude Opus 4.8 at $5/$25 sits right next to GPT-5.6 Sol and GPT-5.5 at $5/$30 — Anthropic is slightly cheaper on output. Mid-tier, Sonnet 5 ($3/$15, or $2/$10 intro) runs against GPT-5.6 Terra ($2.50/$15). Budget tier, Haiku 4.5 ($1/$5) faces GPT-5.6 Luna ($1/$6). Above them all, Claude Fable 5 ($10/$50) undercuts GPT-5.5 Pro ($30/$180) by 3–3.6x at the frontier tier.

Prices this close make benchmarks the tiebreaker. See the full OpenAI API pricing table, or the head-to-head Opus 4.8 vs GPT-5.6 Sol comparison on the compare hub.

Start at the bottom and escalate on evidence. Haiku 4.5 ($1/$5) clears the bar for classification, routing, extraction, and most summarization. Sonnet 5 ($3/$15) is the production default — coding, reasoning, customer-facing content — and its intro pricing makes the upgrade nearly free through August. Opus 4.8 ($5/$25) earns its 1.67x premium when instruction-following precision or long-context work measurably reduces rework. Fable 5 ($10/$50) is for the slice of work where you have measured that even Opus falls short — it currently holds the top overall score on BenchLM's leaderboard.

A routing strategy beats a single tier: send 60% of traffic to Haiku, 35% to Sonnet, and the last 5% to Opus or Fable, and the blended rate lands 30–40% below running everything on Sonnet.