📖 8 min read
Updated July 3, 2026. Most pricing pages are lies dressed up as tables. Anthropic’s is technically correct and practically useless. The numbers on it are real, but the model names shift every quarter, the introductory rates quietly sunset on August 31, 2026, and nobody tells you what a “real” Claude workload looks like on the bill.
I run Claude in three production apps. Last month the bills ranged from $48 to $11,420 for the same model. Same rate card, wildly different numbers. Here is the real Claude API rate card for 2026, the worked examples behind the bills, and the trap I almost walked into with the August 31 cliff.
If you are shopping for an LLM and comparing stacks, start with the Claude rate card below — and then run the same numbers through our AI API cost calculator against GPT-5 and Gemini. The cross-vendor comparison is in GPT-5 vs Claude vs Gemini: the 2026 developer bill I ran across 4 real workloads.
What I thought I’d pay vs what I actually paid
In June 2026 I spun up a side-by-side test: one Sonnet 4.6 workload at the advertised rate ($3 input / $15 output per 1M tokens), one at the introductory rate ($2 / $10), and one in production for a client where I’d never sat down and modeled the cost.
The numbers, on identical workloads:
| Scenario | Rate applied | Tokens used | Bill |
|---|---|---|---|
| Calc’d from pricing page | $3 in / $15 out | 12.4M in / 3.1M out | $83.70 |
| Calc’d with intro rate | $2 in / $10 out | 12.4M in / 3.1M out | $55.80 |
| Production client (no planning) | $3 in / $15 out + retries | 61.2M in / 18.7M out | $464.10 |
The first two are identical workloads with different advertised prices. The third is the same model, same calendar month, but with retries, a long system prompt, and two parallel agent calls. The rate card told me I’d pay $83. I actually paid $464. That’s a 5.5x multiplier on the headline number — and that’s before caching or batching.
The full Claude rate card (July 2026)
Three current models, plus the four pricing knobs that actually move your bill. Memorize these.
Model rates per 1M tokens
| Model | Input | Output | Notes |
|---|---|---|---|
| Claude Opus 4.7 | $5.00 | $25.00 | Flagship. Deepest reasoning, 1M context. |
| Claude Sonnet 4.6 | $3.00 | $15.00 | Default for production. Intro $2 / $10 until Aug 31, 2026. |
| Claude Haiku 4.5 | $1.00 | $5.00 | Cheap & fast. Bulk classification, routing. |
| Claude Sonnet 4.5 (legacy) | $3.00 | $15.00 | Still billed, soon-to-deprecate. |
The four pricing knobs
| Feature | Discount | When it actually helps |
|---|---|---|
| Prompt caching (5-min TTL) | ~90% off cached input tokens | Same system prompt or document re-sent every call |
| Prompt caching (1-hour TTL) | ~85% off cached input tokens | Long-running assistants, repeated RAG chunks |
| Batch API | ~50% off both directions | Async jobs: nightly summaries, bulk classification |
| Vision input (image tokens) | Billed as input tokens | Each 1024×1024 image ≈ 1,600 tokens |
Read that caching row twice. If your app sends the same 50K-token system prompt on every call — and most do — prompt caching is the single largest cost lever you have. It is not a marketing bullet. On my Sonnet 4.6 workloads it dropped the bill by 71%.
Three real developer workloads with the actual bill
Rate cards are fantasy until you put a workload on them. Here are three I ran, with the input/output token counts I measured and the bills I paid.
Workload 1: 5,000-line code review
A repo review agent. Each call: ~62K input tokens (code + system prompt), ~2.4K output tokens (review comments). Ran 800 reviews in June.
| Approach | Total tokens | Cost |
|---|---|---|
| Sonnet 4.6, no caching, no batching | 49.6M in / 1.92M out | $177.60 |
| Sonnet 4.6 + 5-min prompt cache | 49.6M in (3.6M fresh / 46M cached) + 1.92M out | $39.60 |
| Opus 4.7, no caching | 49.6M in / 1.92M out | $296.00 |
| Haiku 4.5, no caching | 49.6M in / 1.92M out | $59.20 |
The cache is doing 78% of the work here. Same review quality, $138/month saved.
Workload 2: 200K legal-doc Q&A
Long-context RAG over a contract corpus. Each call: ~210K input tokens (top-K chunks + system prompt), ~1.8K output tokens. 1,200 questions in June.
| Approach | Total tokens | Cost |
|---|---|---|
| Sonnet 4.6, no cache | 252M in / 2.16M out | $788.40 |
| Sonnet 4.6 + 1-hour cache | 252M in (12M fresh / 240M cached) + 2.16M out | $68.40 |
| Opus 4.7, no cache | 252M in / 2.16M out | $1,314.00 |
This is the workload where caching is non-negotiable. The system prompt and the top-K retrieved chunks barely change between calls. Without caching you pay ~12x more for the same answer quality.
Workload 3: Image + text pipeline
A product-listing generator: input is a product photo (~1,600 image tokens) + a 4K caption prompt, output is 1.2K tokens of structured JSON. 14,000 listings processed in June.
| Approach | Total tokens | Cost |
|---|---|---|
| Sonnet 4.6, no cache, no batching | 78.4M in / 16.8M out | $487.20 |
| Sonnet 4.6 + Batch API | 78.4M in / 16.8M out | $243.60 |
| Haiku 4.5 + Batch API | 78.4M in / 16.8M out | $121.80 |
Haiku 4.5 was good enough for this. The structured JSON came back parseable 99.4% of the time. We moved the whole pipeline to Haiku + Batch and the bill dropped from $487 to $121.
The hidden 2026 trap: intro rates ending August 31
This is the part I almost missed. Anthropic launched Sonnet 4.6 in May 2026 with an introductory rate of $2 input / $10 output per 1M tokens — about 33% off the standard rate. It is documented on the pricing page. It is also scheduled to sunset on August 31, 2026, at which point the price jumps to $3 / $15.
If you built a budget assuming the intro rate, that budget is wrong by 50% in three months. On my mid-sized Sonnet workload ($800/month at intro pricing), this is a $400/month increase I now have to model for. On a heavier workload I am watching, it is a $4,200/month swing.
What to do:
- Pull your last 90 days of usage from the Anthropic console now.
- Model the bill at the standard rate before Aug 31 hits.
- Use the time to either (a) introduce caching, (b) move down to Haiku where possible, or (c) renegotiate via enterprise commit — Anthropic does discount standard rates for $20K+/month commits.
When to use which model
There is no honest answer to “which Claude model should I use.” There is only “which model for this workload at this volume.” A few rules I run:
- Haiku 4.5 for routing, classification, extraction, and any task with a clear rubric. 5x cheaper than Sonnet, latency is half. Quality is fine for 80% of “LLM in the loop” calls.
- Sonnet 4.6 for customer-facing reasoning, code review, document Q&A, structured generation. The default. Always pair with prompt caching on the system prompt.
- Opus 4.7 only when reasoning is the product. Legal analysis, multi-document synthesis, deep research. Treat it as a senior consultant, billed by the hour.
One pattern I have standardized across all three of my apps: a Haiku router in front of a Sonnet worker. Haiku reads the request, classifies intent and difficulty in ~200 output tokens, and only forwards to Sonnet when the task warrants it. On a typical SaaS workload this cuts Sonnet usage by 55–70% with no measurable quality drop on the tasks that matter.
The real-world model: how my June 2026 bill actually broke down
Across all three production apps, here is the model-tier split of my actual June 2026 bill:
| Model | Calls | Share of bill |
|---|---|---|
| Haiku 4.5 (router + classification) | 284,000 | 11% |
| Sonnet 4.6 (worker + cache hits) | 96,400 | 71% |
| Opus 4.7 (deep research only) | 1,820 | 18% |
That is a healthy distribution. Opus is 18% of the bill on 1.3% of calls — exactly the shape you want. If Opus is showing up as 60% of your bill, you are using it where Sonnet would do.
FAQ
How much does Claude Sonnet 4.6 cost per 1M tokens in 2026?
The standard rate is $3 per 1M input tokens and $15 per 1M output tokens. An introductory rate of $2 / $10 runs until August 31, 2026.
Is Claude prompt caching actually 90% off?
Yes — on cache hits. The first write of the cached prefix is billed at full price; every subsequent read within the TTL window (5-min or 1-hour) is billed at 10% of the input rate. On a 50K-token system prompt hit on every call, this is a 60–80% bill reduction in practice.
When does the Sonnet 4.6 intro rate end?
August 31, 2026. After that date the standard $3 / $15 rate applies. Model your September bill at the higher rate now.
Is Claude Opus worth it over Sonnet?
Only if the task is reasoning-heavy and the cost of a wrong answer exceeds the 1.67x premium. For most production apps, Sonnet 4.6 + caching + a Haiku router beats Opus on both cost and quality.
What’s the cheapest way to run Claude in 2026?
Three levers, in order: (1) prompt caching on every system prompt and reusable context, (2) Batch API for any async workload, (3) Haiku 4.5 for any task that does not require Sonnet-tier reasoning. Most apps can hit 60–80% bill reduction with these three alone.
The verdict
Claude’s pricing is honest but the rate card alone will mislead you. The headline numbers are accurate; the bills are 2–6x higher than the math suggests once retries, long context, and uncached system prompts hit. The August 31 intro-rate sunset is the single biggest modeling risk for any team running Sonnet 4.6 today.
Build your cost model with caching, batching, and the post-Aug-31 standard rate. Then run it through the AI API cost calculator and see where you land. If your bill surprises you, it is almost always one of three things: a long uncached system prompt, a retry loop you didn’t know about, or a model tier that does not match the task.
What I’d build today if I were starting from zero
If I were spinning up a new Claude-powered product in July 2026, here is what I would do on day one — not after the first $5K surprise bill.
- Default to Sonnet 4.6 + prompt caching. Cache the system prompt and any stable reference material. Expect 60–80% savings versus the naive setup.
- Put Haiku 4.5 in front of every Sonnet call. Even a 200-token classification step pays for itself by stopping expensive reasoning where it is not needed.
- Move async workloads to the Batch API. Anything that can tolerate a 24-hour turnaround — nightly reports, bulk tagging, document summarization — belongs on Batch at half price.
- Instrument per-call token counts. Most teams find their model is generating 2–4x more tokens than necessary. A 1,200-token cap on outputs usually pays for the instrumentation in a week.
- Re-price your September 2026 forecast at the post-intro rate. If the number breaks the budget, redesign the workload before the rate hike, not after.
Claude is genuinely the most pleasant model family to build with in 2026. The pricing is fair, the caching model is the best in the business, and the per-token quality is excellent. Just do not pretend the headline rate card is your bill. It never is.