Claude API Pricing Comparison: How Fable 5, Opus 5, Sonnet 5, and Haiku 4.5 Are Billed — and How to Save with Caching and Batch

Claude API pricing comparison isn't about "how much is the plan" — it's about what per-model token rates look like, how much caching and Batch can save, and whether direct API or a cloud platform is cheaper. The current lineup is Fable 5 ($10/$50), Opus 5 ($5/$25), Sonnet 5 (intro $2/$10), and Haiku 4.5 ($1/$5) — the flagship input rate is 10x the entry tier; prompt cache hits bill at just 0.1x input, and Batch API knocks off roughly 50%. This article tallies the four models, caching and Batch, plus the direct-vs-cloud math. Prices below are as of 2026-08-13; check Claude API pricing for live rates.

How Claude API bills: per-token, separate from subscriptions

Claude API is a pay-as-you-go service on the Anthropic Console with no monthly plan: input and output tokens bill separately, each of the four models has its own rate, and caching, Batch, Web Search, and other line items stack on top (source). It is a separate system from Claude Pro/Max web subscriptions: Claude Code inside a subscription draws on plan quota, while self-hosted backends and production agents require a dedicated API key in the Console billed per token. Keys are created at console.anthropic.com, and the same models are also callable through Amazon Bedrock, Vertex AI, and Microsoft Foundry.

Model-by-model comparison: from $1 to $10, a 10x gap

Claude API currently runs four mainline models — Haiku 4.5 ($1/$5), Sonnet 5 (intro $2/$10), Opus 5 ($5/$25), and Fable 5 ($10/$50) — and the flagship input rate is 10x the entry tier. All four support prompt caching, and the three flagship models carry 1M context (Haiku has 200K); which one you pick depends on task complexity and budget (source). Full price table below, per 1M tokens:

Claude Haiku 4.5: $1/$5, fast and low-cost

Haiku 4.5 is the cheapest of the four at $1/M input and $5/M output, with 200K context, 64K max output, and the lowest latency. It targets fast Q&A, routing/classification, formatting, sub-agents, and first-pass steps in high-concurrency pipelines, supporting extended thinking (no adaptive thinking). Cache writes run $1.25/M and reads $0.10/M, with further halving via Batch API — a natural fit for "Haiku first pass, escalate to Sonnet/Opus for hard tasks" tiered architectures.

Claude Sonnet 5: intro $2/$10, the production workhorse until late August

Sonnet 5 carries introductory pricing through 2026-08-31 at $2/M input and $10/M output, then reverts to standard $3/$15. With 1M context, 128K output, and adaptive thinking, it is positioned for high-performance coding and agents — a fit for high-frequency API calls and shared team backends that want 1M context without Fable/Opus everywhere. During the intro window, cache writes are $2.50/M and reads $0.20/M.

Claude Opus 5: $5/$25, complex enterprise work

Opus 5 runs $5/M input and $25/M output with 1M context and 128K output, positioned for complex agentic coding and enterprise work. It suits complex code, enterprise workflows, and high-autonomy tasks; when Fable 5 is too costly but flagship capability is still needed, Opus 5 is the lower-cost pick. Cache writes are $6.25/M and reads $0.50/M; Fast Mode bills at 2x standard with roughly 2.5x speed.

Claude Fable 5: $10/$50, the long-running agent flagship

Fable 5 is the top API tier at $10/M input and $50/M output, with 1M context, 128K output, and adaptive thinking always on. Positioned as next-generation intelligence for long-running agents, it fits very long task chains, complex autonomous agents, and high-value knowledge work. Cache writes run $12.50/M and reads $1/M — for cost-sensitive workloads, tier with Haiku/Sonnet first and decide before committing to Fable.

Prompt caching and Batch API: the two most practical savings levers

The two levers worth using on Claude API are prompt caching and Batch API: cache-hit reads cost 0.1x input (5-minute window), Batch API discounts input/output by roughly 50% with 24-hour async completion, and stacked together repeated-context workloads can drop below a tenth of list price. Caching suits long fixed context like system prompts and tool definitions; Batch suits jobs that do not need a real-time response (source).

1. Enable prompt caching for repeated context: a fixed system prompt plus a long document pays full price on the first write, then 0.1x on hits within 5 minutes — Haiku 4.5 cache reads drop to $0.10/M.
2. Route async jobs through Batch API: ~50% off input/output with completion within 24 hours, for batch generation and offline scoring.
3. Choose US-only inference when data residency matters: input/output bill at 1.1x with data staying in the US; Bedrock/Vertex regional endpoints may also add 10%.
4. Move to Priority Tier when throughput is the bottleneck: pay official Priority multipliers for higher concurrency and lower latency in production.

How add-ons are priced: Web Search, Managed Agents, Code Execution

Beyond token rates, Claude API bills several add-ons separately: Web Search at $10/1K searches, Managed Agents at $0.08/session-hour, Code Execution at $0.05/hour (50 free hours daily), and Opus 5 Fast Mode at 2x standard. These line items are easy to overlook in agent toolchains, so budget them explicitly (source):

Direct Anthropic API vs Bedrock / Vertex: where the 10% premium comes from

The same models are available from the first-party Anthropic API or via Amazon Bedrock, Vertex AI, and Microsoft Foundry, but third-party cloud regional endpoints may add 10% — direct access carries no such premium. Direct means you manage keys, rate limits, and billing yourself in the Console; Bedrock/Vertex means you reuse your existing cloud account, IAM, and compliance chain. Token rates are the same at the source, so the real trade-off is "data residency + cloud ecosystem vs. 10% less premium + first-party direct" (source).

Is Claude expensive next to OpenAI, DeepSeek, and other APIs?

Claude API pricing sits in a delicate spot: flagship Fable 5 ($10/$50) is pricier than the OpenAI API flagship tier and an order of magnitude above DeepSeek API flagships — the value case rests on caching and Batch. Side-by-side mainstream API rates per 1M tokens (input/output), matching the data in our OpenRouter API pricing comparison article:

The takeaway: Claude's flagship tier leads the world in quality and in price. For first-party Claude quality, tier by model and lean on caching plus Batch; for many models under one key, look at OpenRouter API; for price-sensitive Chinese-language workloads, DeepSeek and Kimi API entry tiers cost an order of magnitude less.

Picking models and controlling cost

Claude API selection boils down to "tier by task and max out caching and Batch": light calls on Haiku, daily production on Sonnet, complex work on Opus, long-running agents on Fable — while making repeated context cacheable and async jobs batchable. Concretely:

1. Use Haiku 4.5 for high-frequency light calls: classification, routing, formatting, sub-agent first passes at $1/$5 with the lowest latency.
2. Run daily production API on Sonnet 5: lock in the $2/$10 intro rate before the end of August; it reverts to $3/$15 after.
3. Pick Opus 5 for complex coding and enterprise workflows: $5/$25, the middle ground when Fable 5 is too costly but flagship capability is still needed.
4. Reserve Fable 5 for long-running agents and high-value knowledge work: $10/$50 with 1M context and adaptive thinking — only when it is genuinely required.
5. Enable prompt caching for fixed long context: 1.25x writes and 0.1x reads, then halve again with Batch API.
6. Send async batch work through Batch API: ~50% off input/output, completed within 24 hours — don't let real-time requests carry that cost.

Caveats

1. Sonnet 5's promo price has a deadline: $2/$10 ends 2026-08-31 and reverts to $3/$15, so plan migrations before the window closes.
2. Legacy models remain in pricing tables: Sonnet 4.6, Opus 4.7, and older versions may still appear; new integrations should prefer Fable 5 / Opus 5 / Sonnet 5 / Haiku 4.5.
3. API and subscriptions bill separately: Claude Code inside a Pro/Max subscription draws on plan quota, not API balance — and API usage never counts toward plan quota either.
4. Add-on costs are easy to miss: Web Search $10/1K searches, Managed Agents $0.08/session-hour, Code Execution $0.05/hour (50 free hours daily), Opus 5 Fast Mode 2x.
5. Regional and cloud endpoints carry premiums: US-only inference bills at 1.1x and Bedrock/Vertex regional endpoints may add 10% — model costs for your actual deployment region.
6. Rates change: all prices follow the live claude.com/pricing page; this article's data is as of 2026-08-13.