Claude API Pricing (2026): Opus 5, Sonnet 5, Haiku 4.5 Costs

Searching "claude api pricing" in September 2026 surfaces an official subscription page, a Reddit thread arguing whether Pro or the API is cheaper, and cost-tracking sites that mostly describe models from the previous generation. None of that tells you what you actually came for: the current per-token rates for the Claude models you can call today, and how a real invoice is assembled from them. That is what this page covers, with every figure captured on September 3, 2026.

Three Claude models are available for API use right now — Opus 5, Sonnet 5, and Haiku 4.5 — and all three are flat-rate: one input price and one output price per million tokens, with no per-tier surcharge for long contexts. This article gives you the full table, explains the flat-rate and caching mechanics that decide your real bill, and closes with an honest route comparison for buying direct from Anthropic versus calling the same models through an aggregator. We run Modellix, an API gateway that lists all three models (and routes the Anthropic Messages protocol), so we have a commercial interest when Modellix appears below; prices here were read from the Modellix LLM price page and Anthropic's own pricing documentation on September 3, 2026, and we say so wherever a number could not be verified.

Claude API pricing at a glance (September 3, 2026)

All rates are USD per 1 million tokens, which is the only unit Anthropic bills in. "Successful responses are billed from token usage" — a failed request is not a surprise line item. The table below is the current rate card for the three in-sale models, as displayed on the Modellix LLM price page on September 3, 2026; Anthropic's own model pricing documentation listed identical numbers on the same day, which we note because the two pages are maintained independently.

Sources: Modellix LLM price page and Anthropic model-pricing documentation, both accessed September 3, 2026. The cached-write column shows the 5-minute cache-write rate (1.25× input); Anthropic also sells 1-hour cache writes at 2× input, which the Modellix table does not display.

The source of the table above: Modellix's "Models and pricing" list as captured September 3, 2026. Note the Discount column reads "—" for every Anthropic row — see the route comparison section.

Three things in this table are worth flagging before you compare it against an older article you bookmarked:

1. Sonnet 5 does not cost $3/$15. It launched at $2/$10 as "introductory" pricing, and several guides and forum posts still say it steps up to $3/$15 on September 1, 2026. That increase was cancelled: Anthropic's Sonnet 5 announcement records (August 10, 2026 edit) that the $2/$10 rate is now permanent, and the pricing docs state the scheduled increase "will not occur." If a 2026 article tells you Sonnet 5 is about to get more expensive, it is stale.
2. Opus 5 carried its predecessor's price. Anthropic's Opus 5 announcement confirms $5/$25 — the same rate card Opus 4.8 carried, and one that makes the previous-generation model no cheaper than the current one. What changed with Opus 5 is not the sticker price.
3. The context window is part of the price. Opus 5 and Sonnet 5 accept up to 1M tokens and bill every token at the same flat rate — a 900K-token request bills like a 9K one, no long-context surcharge. Haiku 4.5 tops out at 200K tokens. The window decides how much you can send in one request; it does not change the per-token rate.

One newer model now sits above this table: Claude Fable 5.1 (plus its limited-availability sibling Mythos 5.1) tops Anthropic's rate card at $10 input / $50 output per 1M tokens, with cache reads priced at a special $0.25 instead of the usual 0.1× rate. It is a new premium tier above Opus, not a replacement for these three, and this pricing guide keeps it as a boundary note rather than a fourth column.

Flat-rate billing: what "per million tokens" actually bills

Anthropic's Claude models are flat-rate, which is a pricing shape worth understanding precisely because most other vendors are not. On a flat-rate model there is exactly one input price and one output price, regardless of how many tokens you send. On a tiered model (several OpenAI and xAI models on the market use tiers), the input price doubles past a threshold like 200K or 272K tokens. Claude does none of that: the Modellix price table's "Input Context" column shows "—" for all three models, and Modellix's LLM pricing FAQ describes the models as flat-rate, with no per-tier surcharge for long contexts.

Flat-rate is not the same thing as a fixed context window. The two are independent: the context window (200K on Haiku 4.5, 1M on Opus 5 and Sonnet 5) is the supported input size, while the pricing tier — had there been one — would be the rate you pay inside it. Claude gives you the whole window at the standard rate. That is the main reason long-context agent workloads are viable on Sonnet 5 or Opus 5: a 700K-token codebase scan bills at $2 or $5 per million input tokens, not at a penalty rate for crossing a boundary.

What does vary is the mix of line items a single request generates. A plain request bills two things — input tokens and output tokens. Once you use prompt caching, batch, or both, a request can touch four different rates, which is where per-token comparisons that only quote "input/output" understate reality. The next section walks the arithmetic.

The cache columns are where the real price swings live

Prompt caching is the single largest lever on a Claude bill, and it is priced as two separate line items on top of base input:

Multipliers from Anthropic's prompt-caching pricing section, accessed September 3, 2026. The 5-minute write is the rate most pricing tables — including Modellix's — display in a single "cached write" column.

The mechanics: the first time a request sends a prefix (a system prompt, a tool schema, a large document), those tokens are billed at the cache-write rate — which replaces the base input rate for the tokens being cached — and stored. Every later request that sends the identical prefix bills those tokens at the cache-read rate instead of the base input rate. The read rate is one-tenth of base input, which is why repeated context collapses in price after the first write.

How one request maps onto the four billing streams — illustrative Modellix artwork, not a product screenshot. Read and write channels show their relative rates as bar weight only, without numbers.

Concrete math on Opus 5, for a 100,000-token stable prefix (system prompt + schema + context) sent 100 times:

• No caching: 100 requests × 100,000 × $5 / 1M = $50.00 in input alone.
• With caching: one write of 100,000 tokens at $6.25 / 1M = $0.625, then 99 reads at $0.50 / 1M = $4.95 → ≈ $5.58.

That is a ~89% reduction on the repeated portion, and it is why agent harnesses — which resend the same system prompt and tool definitions on every turn — are effectively priced by their unique tokens, not their total tokens. The same multiplier math applies on Sonnet 5 ($0.20 read) and Haiku 4.5 ($0.10 read).

Two further knobs exist on Anthropic's own API that are worth knowing even though they are not part of Modellix's display: the Batch API applies a 50% discount to both input and output for asynchronous jobs (Opus 5 drops to $2.50 in / $12.50 out), and caching discounts stack with batch. Reasoning-heavy requests also deserve a budget line of their own: a model that spends a long time thinking before answering can turn a small prompt into a large bill — size your cost estimate around worst-case thinking, not the happy path.

Opus 5, Sonnet 5, Haiku 4.5: three price points, three jobs

The three models share the same billing machinery, so the price difference is a capability choice:

• Claude Haiku 4.5 — $1 / $5. The fast, cost-efficient tier. At half of Sonnet's input price it is the default for high-volume, low-complexity traffic: classification, extraction, support triage, summarization. Its 200K window is plenty for single-document work; its price is why most high-volume teams put their simplest calls here.
• Claude Sonnet 5 — $2 / $10. The balanced production default — Anthropic positions it as the best combination of speed and intelligence. Coding assistance, agent loops, and most product features run here; at this price (permanently, per the August 10 change) it undercuts its predecessor Sonnet 4.6's $3/$15 while adding the 1M window.
• Claude Opus 5 — $5 / $25. The flagship for complex agentic coding and enterprise workloads where accuracy justifies 2.5× Sonnet's input rate and the 1M window matters more than unit cost.

The practical selector for most teams: start on Sonnet 5, measure, and push anything that does not need its ceiling down to Haiku 4.5; reserve Opus 5 for the hardest reasoning and longest-horizon agents. A capability comparison between Opus 5 and Sonnet 5 is beyond this article's pricing scope — the numbers above are what each tier costs, and Anthropic's model pages own the benchmarks.

Direct from Anthropic or through an aggregator?

Per-token rates are only half the buying decision; the other half is where you call the model from. Because Modellix lists these three models at exactly the rates shown in the table above — its Discount column currently reads "—", meaning the price you pay equals the official list price — this is not a "cheaper per token" comparison, and we will not pretend otherwise. A gateway earns its place on operations, not on a price delta: one API key that also reaches OpenAI, Google, DeepSeek, Qwen, and xAI models; per-request cost logs that show exactly what each call billed; and stable ~ model aliases (for example ~anthropic/claude-opus-latest) that follow new releases without a code change. Anthropic direct, by contrast, gives you first-party billing, usage tiers, and the Batch API — and a separate vendor account to manage for every other model family you use.

Two honest notes on routing Claude traffic through a gateway like Modellix. First, it is a distribution layer, not the model maker: the model is Anthropic's, the rate card is Anthropic's, and we do not operate our own Opus or Sonnet. Second, per-token parity is not a promise that survives forever — Modellix's page previously showed a 3% discount on these rows in late August 2026, and it no longer does; re-check Modellix's LLM price page before you budget, the same way you would re-check Anthropic's page.

If the per-token table is all you came for, the earlier section is your answer. If you are comparing Claude against the rest of the market, our cross-vendor LLM API pricing comparison puts these rows next to OpenAI, Google, DeepSeek, and others on one table. And if you have already decided on Claude and just want your existing Anthropic-protocol client (Claude Code, the Anthropic SDK, an agent harness) to call it through one key, the Anthropic base URL guide covers exactly that — what to set, where, and the no-/v1 trap.

Claude Models on the Modellix LLM Gateway

See the current model list, protocol routing, and pricing notes in the Modellix LLM documentation.

View Docs

Frequently Asked Questions

How much does the Claude API cost per token?

Anthropic bills per million tokens, not per token. Haiku 4.5 at $1 input / $5 output, Sonnet 5 at $2 / $10, and Opus 5 at $5 / $25 per 1M tokens. Cache reads cost one-tenth of input (Fable 5.1 excepted), cache writes 1.25× (5-minute) or 2× (1-hour) input, and Batch API jobs get 50% off input and output.

Is the Claude API free?

No. Every successful request is billed from token usage. Anthropic's pricing FAQ describes limited testing credits for new developer accounts; those do not make the API free. Modellix ended its automatic $1 signup credit on August 19, 2026. Its two $0-priced LLM rows, modellix-ai/free-llm and zai/glm-4.7-flash, are unrelated to Anthropic and are not a free way to call Claude.

Is the API cheaper than a Claude Pro or Max subscription?

They are different products with different cost structures, so the comparison only works per workload. Subscriptions (see claude.com/pricing) are flat monthly fees with usage caps on Anthropic's chat and coding surfaces; the API is pay-per-token with no cap and a separate meter. Heavy interactive use inside a subscription's limits is cheaper on a plan; programmatic, elastic, or capped-budget workloads are cheaper — or only possible — on the API. Which one minimizes your bill depends on whether your usage fits the plan's envelope; we are not going to tell you to route around a subscription to save money, because the two meters do not share tokens.

Which Claude model should I use for API calls?

Start with Sonnet 5 ($2/$10) for most production work, use Haiku 4.5 ($1/$5) for high-volume simple tasks, and reserve Opus 5 ($5/$25) for the hardest reasoning and longest agent runs. This is a pricing guide, so treat capability claims as Anthropic's, not ours.

Is Sonnet 5 really $2/$10, or is it going up to $3/$15?

It is $2/$10, permanently. Sonnet 5's introductory rate was scheduled to rise to $3/$15 on September 1, 2026, but Anthropic recorded on August 10, 2026 that the introductory pricing "has since been made permanent," and the pricing docs say the increase will not occur. Guides and forum posts still circulating the increase are outdated.

How much does prompt caching actually save?

Cache reads cost 0.1× the base input rate, so a repeated prefix that would cost $5.00 per 1M uncached on Opus 5 costs $0.50 per 1M once cached — a 90% cut on the cached portion, after paying a one-time write premium of 1.25× (5-minute TTL) or 2× (1-hour TTL). Workloads that resend a stable system prompt or large context benefit most; one-shot requests never touch the cache columns.

What about Opus 4.8 and Sonnet 4.6 — are those cheaper?

No — they are the previous generation. Opus 4.8 costs the same $5/$25 as Opus 5, and Sonnet 4.6 costs $3/$15, which is now more than Sonnet 5's permanent $2/$10. There is no price reason to stay on the older models, and Anthropic is retiring the 4.x line over time; Anthropic's model-deprecation list is the authority on exact retirement dates.

Should I switch clients to save money on Claude?

Changing the price you pay and changing the client you use are different decisions. This article covers the price side; whether to switch Claude Code or your harness at all — and what a cheaper client path looks like — is covered separately in our Claude Code alternatives guide.

Check Current Claude Model Pricing on Modellix

Log in to compare the live listed rates for Opus 5, Sonnet 5, and Haiku 4.5 and see per-call cost logs on one key.

Login