Anthropic's current Claude model lineup
Anthropic prices five Claude models for API access, spanning a wide range from budget to flagship. Claude Opus 5 is the new flagship, priced at $5 input / $25 output per million tokens, the same per-token rate as Opus 4.8, but it runs adaptive thinking by default, meaning it can spend tokens reasoning internally before producing its visible answer, and those thinking tokens bill as output. Claude Fable 5 is still priced above every other current Claude model on both input and output tokens ($10/$50), remaining the choice for work that needs Anthropic's single highest-quality output regardless of cost. Claude Opus 4.8 stays available at its existing rate, without Opus 5's adaptive-thinking behavior, for teams that want predictable output-token counts. Claude Sonnet 5 occupies the middle of the lineup and currently carries introductory pricing that runs through 2026-08-31, after which its rate is scheduled to step up to $3 input / $15 output per million tokens, worth building into any long-range budget rather than assuming today's rate holds. Claude Haiku 4.5 is the smallest and cheapest model in the current lineup, built for high-volume, latency-sensitive work rather than maximum capability.
Opus 5's adaptive thinking and what it does to cost
Opus 5's headline rate matches Opus 4.8's exactly, so on paper it looks like a free upgrade. The catch is adaptive thinking: Opus 5 decides for itself, per request, whether and how much to reason before answering, and every one of those thinking tokens is billed at the model's output rate, identically to the tokens in the visible response. A request that would cost a fixed amount on Opus 4.8 (predictable output length, no separate thinking phase) can cost meaningfully more on Opus 5 if it triggers a long internal reasoning pass, even though the visible answer is the same length. Budget off measured total output tokens for your actual traffic, not just the length of the answer text you see, and treat any Opus 5 estimate as a floor rather than a fixed number until you've measured it against real calls.
The Sonnet 5 introductory-pricing quirk
Claude Sonnet 5's current rate is explicitly introductory and time-boxed; it's set to increase after 2026-08-31 to $3 input / $15 output per million tokens. This is the single most important pricing detail to get right if you're modeling costs for a project with any real runway: a budget built on today's introductory rate will understate the true cost of running Sonnet 5 in production past that date. If your monthly volume is large enough that this increase materially changes your unit economics, it's worth re-running your projection with the post-introductory rate now, before you're locked into an architecture built around the cheaper number.
A worked example on Claude Opus 5
At Opus 5's rate ($5 input / $25 output per million tokens), a 1,000-token prompt with a 500-token visible answer and no thinking overhead costs 1,000 ÷ 1,000,000 × $5 = $0.005 for input, plus 500 ÷ 1,000,000 × $25 = $0.0125 for output, $0.0175 per call, or $1,750 a month at 100,000 calls. Now suppose the same request triggers 300 tokens of internal thinking before that 500-token answer: output becomes 800 tokens, 800 ÷ 1,000,000 × $25 = $0.02, for $0.025 per call and $2,500 a month, a 43% increase over the no-thinking estimate for an identical visible response. That gap is exactly why thinking-token behavior matters more to an Opus 5 budget than the sticker rate does.
Cutting your Claude API bill
The most effective lever is matching model tier to task: route simple, repetitive, or latency-sensitive calls to Claude Haiku 4.5 by default, and reserve Sonnet 5, Opus 4.8, Opus 5, or Fable 5 for requests where the extra capability is actually visible in the output. Beyond model choice, cap max output tokens so verbose generations don't inflate cost unnecessarily, keep system prompts and few-shot examples as lean as they can be while still holding quality, and avoid re-sending large blocks of static context on every call if your integration lets you persist it instead. On Opus 5, watch measured thinking-token counts specifically, a task that doesn't need deep reasoning may do better on Opus 4.8 or Sonnet 5, where output length is more predictable. Because Sonnet 5's rate is scheduled to change, revisit your model choice around the pricing transition date rather than assuming your current setup stays cost-optimal indefinitely.
Who the Claude API suits
Anthropic's five-tier lineup gives you a clear cost ladder from Haiku up to Opus 5 and Fable, which suits teams that want to reserve top-tier reasoning for a narrow slice of genuinely hard requests while running the bulk of traffic on a cheaper model, a pattern that plays well with Claude's known strengths in careful, instruction-following output. If you're deciding between Claude and another vendor rather than between Claude models, the all-provider hub calculator lets you compare rates across the whole market in one place.
See how Claude compares against OpenAI's GPT API pricing, Google's Gemini API pricing, or xAI's Grok API pricing.
