Anthropic Claude API Pricing 2026: The Real Cost of Opus 4.7, Sonnet 4.6, and Haiku 4.5

What Anthropic charges in 2026

Anthropic runs six billable Claude families in July 2026, plus a free tier in the console for evaluation. The flagship claude-opus-4-7 sits at $5/$25 per million tokens — the same nominal rate as OpenAI’s gpt-5.5, but with materially better long-context quality past 200K. claude-sonnet-4-6 is the sweet spot at $3/$15 (introductory rate of $2/$10 until August 31, 2026), and claude-haiku-4-5 at $1/$5 is the cheapest Claude you can buy on a regular production key.

Two things move your Anthropic bill more than the model choice: prompt caching (up to 90% off input) and the Batch API (roughly 50% off for async work). I’ll get to both in a moment. First, the rate card.

The full Claude rate card (July 2026)

All rates per 1M tokens. Vision (image inputs) is included in the input rate at no extra charge on every Claude 4.x model. Prompt caching writes are billed at the cached-input rate; cache TTL is 5 minutes by default and extendable to 1 hour. “Long context” applies above 200K tokens for the 4.5+ family.

Batch and prompt-caching reference

The deadline you cannot miss: claude-sonnet-4-6 is on an introductory rate of $2 input / $10 output per 1M tokens until August 31, 2026. On September 1, 2026 the rate goes to $3/$15. That’s a 50% input-cost increase and a 50% output-cost increase if you don’t migrate. Most teams should pin their bills forward at the post-Aug 31 rate and treat the intro rate as a free option.

Introductory pricing ending August 31, 2026

Three things to do before the deadline:

• Lock your bill forecast at $3/$15 — don’t model business cases on $2/$10.
• Audit caching — if you’re not caching Sonnet 4.6 system prompts today, the 90% saving will be the difference between a profitable and an unprofitable agent in September.
• Re-evaluate Opus vs Sonnet — the post-Aug 31 Sonnet-to-Opus premium is 67% on input ($3 vs $5) and 67% on output. Worth the bump only on quality-sensitive tasks.

Three real workloads with actual bills

Workload A — 5,000-line codebase review

You’re running a repo-context-aware reviewer on a TypeScript monorepo. Average review is 50K input tokens (full file + imports + tests) and 2K output tokens (inline comments + summary). 20 reviews per day, six days per week.

• On claude-sonnet-4-5 (uncached): $3 × 0.05M × 120 + $15 × 2K × 120 = $18 + $3,600 = $3,618/month.
• On claude-sonnet-4-6 at introductory rate with 5-min cache: ~ 80% of inputs cached at $0.30/M. Effective input cost = $0.30 × 0.04 × 120 + $3 × 0.01 × 120 = $1.44 + $3.60 = $5.04 input; output $3,600. Total $3,605/month. The intro price saves you nothing on this workload — caching is the lever.
• On claude-haiku-4-5 with cache: input $0.10 × 0.04 × 120 + $1 × 0.01 × 120 = $0.48 + $1.20 = $1.68; output $2 × 2K × 120 = $480. Total $482/month. 7.5x cheaper than Sonnet, but quality on nuanced refactor suggestions drops noticeably.

Workload B — 200K legal-doc analysis

You’re summarising contracts for a small law firm SaaS. Each contract is ~190K input tokens (PDF → text + boilerplate context); output is a 4K-token risk report. 50 contracts per week.

• On claude-opus-4-7 (uncached, no long-context tier used since 190K is below 200K threshold): $5 × 0.19 × 250 + $25 × 0.004 × 250 = $237.50 + $250 = $487.50/week, $1,950/month.
• On claude-sonnet-4-6 (uncached): $3 × 0.19 × 250 + $15 × 0.004 × 250 = $142.50 + $150 = $292.50/week, $1,170/month.
• Adding 1-hour prompt caching for the legal-KB prefix (~80K tokens): drops to roughly $700/month on Sonnet. If the contracts share 70%+ boilerplate, the cache hits are essentially free.
• If a single contract exceeds 200K and crosses into long-context: rate jumps to $6/$22.50 on Sonnet, $10/$37.50 on Opus. Bill roughly 2x for that document.

Workload C — 1M-token research task (deep research agent)

An agent that ingests 30 sources (50K tokens each, fetched + cached) and produces a 10K-token research memo per request. 20 requests per week.

• On claude-sonnet-4-6 with caching (sources stay alive for 1 hour per project): input ≈ $12/week; output 0.20M × $15 = $3,000/week. Total $3,012/week.
• On claude-opus-4-7 for synthesis step only: Sonnet handles ingestion and routing at low cost; Opus runs only the final synthesis step (1M tokens in, 10K out). Opus leg = $5 + $250 = $255/request, $5,100/month. Worth it on enterprise research where accuracy is billable.
• On claude-haiku-4-5 for the routing layer: deduped at $1/$5, reduces ingestion cost by ~ 90%. Keep Opus only for the synthesis step.

The four levers that move your Claude bill

Beyond the model choice, four mechanics determine what you actually pay Anthropic each month. Get them right and a $20K monthly bill can become a $3K monthly bill on the same workload.

1. Prompt caching with 1-hour TTL. The default 5-minute cache TTL suits rapid back-and-forth. For research, doc analysis, or any agent that ingests the same KB repeatedly, set the 1-hour TTL — it costs slightly more on cache writes but the hit rate justifies it for most production agents. On Sonnet 4.6 the 1-hour cache write is $3.75 per 1M tokens; the cache read is $0.30 per 1M. If you re-read 5x in an hour, you break even on the second read.

2. The Message Batches API. Roughly 50% off input and output for jobs that complete within 24 hours. Ideal for nightly re-summaries, weekly evals, and bulk classifier calls. Streamed responses are not allowed — the trade-off is throughput for discount. Use it for 30–60% of any non-realtime agent loop and watch your bill halve.

3. Vision handling. Images are billed as tokens based on resized dimensions. A 1024×1024 image is roughly 1,634 input tokens on Sonnet/Opus 4.x. A 2048×2048 image is ~6,550 tokens. If you routinely OCR scanned documents, pre-resize aggressively — billing only the model, not the original size, will save 30–50% on document pipelines.

4. Tool-use design. Anthropic charges input tokens for tool definitions on every call. Tool definitions that bloat past 500 tokens per agent add up. Keep tool schemas tight, pass dynamic args in the user message, and consider splitting mega-agents into smaller routable sub-agents — both for cost and for accuracy.

When Claude wins, when it loses (vs GPT and Gemini)

Use Claude when:

• You need best-in-class long-context quality past 200K. Sonnet 4.6 and Opus 4.7 lead on multi-document legal, financial, and academic synthesis.
• You write nuanced, multi-step prompts with many tools. Claude 4.x has the best tool-use adherence in the industry — fewer dropped arguments, fewer hallucinated parameters.
• Your workload is reasoning-heavy and worth paying more for fewer mistakes. Opus 4.7 vs gpt-5.5 is roughly a wash on price but Opus wins on hard reasoning benchmarks.
• You want free in-console evaluation. Anthropic’s console has a free tier for prompt design and tracing that OpenAI does not.

Avoid Claude when:

• You’re running ultra-high-volume classification. gpt-5.4-nano at $0.20/M or Gemini 2.5 Flash at $0.075/M will crush Haiku 4.5 on cost.
• You need single-call multimodal including image gen. OpenAI bundles gpt-image-1.5; Anthropic doesn’t have an image-gen API.
• You need real-time voice with low latency. OpenAI’s gpt-realtime-2 is the production-ready choice; Anthropic’s voice beta is still gated.
• You depend on a free production tier. Google AI Studio is still the only one with a free tier in 2026.

FAQ

How much does Claude Opus 4.7 cost per million tokens?

$5.00 input / $25.00 output per 1M tokens. Cached input is $0.50/M (90% off). Long-context past 200K is $10/M in, $37.50/M out. Vision (image) inputs are included in the input rate at no surcharge.

What’s the introductory pricing on Claude Sonnet 4.6?

$2 input / $10 output per 1M tokens until August 31, 2026. On September 1, 2026 the rate steps up to $3/$15. Plan your fall budget at the post-Aug 31 rate.

Is Haiku 4.5 the cheapest Claude in 2026?

Yes — claude-haiku-4-5 is $1/$5 per 1M tokens, with cached input at $0.10/M. The legacy claude-haiku-4 is $0.80/$4 but is on a deprecation path through Q4 2026.

How much does Claude prompt caching save?

Up to 90% on input — the cached-input rate is roughly 1/10 of the base input rate (e.g., $0.30 vs $3 on Sonnet 4.6). You pay a cache-write fee on first hit (1.25× base input rate), so break-even depends on hit count. Heavy system prompts + repeated calls hit break-even inside a single session.

Does Claude have a vision API surcharge?

No. Vision (image inputs) is included in the input token rate on every Claude 4.x model. There is no per-image surcharge in 2026.

Is there a free Claude tier?

The Anthropic console has a free evaluation tier for prompt design and tracing. Production API access requires a paid account with a minimum $5 credit. There is no free production inference tier.

How does Claude Batch pricing work?

The Message Batches API offers roughly 50% off input and output for jobs that complete within 24 hours. Use it for backfills, weekly reports, dataset scoring, and any offline agent leg.

Should I pick Opus 4.7 or Sonnet 4.6 for my agent?

Default to Sonnet 4.6 unless you have measured Opus beating it by enough margin on a quality benchmark to justify the 67% rate premium. Sonnet is genuinely good enough for 80%+ of production workloads in 2026; Opus is for the 20% that needs top-tier reasoning on long, structured inputs.

Verdict

Anthropic is the right pick in 2026 when quality on hard reasoning beats cost-per-million. Sonnet 4.6 at $3/$15 (or $2/$10 until August 31) is the best general-purpose agent model in the industry once you account for tool-use adherence and long-context quality. Opus 4.7 is the answer when Sonnet fails on your eval set and the failures cost real money. Haiku 4.5 at $1/$5 is genuinely competitive for routing and lightweight classification, but it loses to gpt-5.4-nano on raw price and to Gemini Flash on cost-per-classification.

The single biggest lever is still prompt caching — a 90% input discount is roughly 10x cheaper than running uncached on the same model. The second lever is the Aug 31 deadline: build your fall budget at $3/$15 on Sonnet, not at the intro rate. If you nail those two, Claude is a margin-positive vendor for almost any agent in 2026.

Updated July 3, 2026. Reviewed against the Anthropic pricing dashboard. Next review: August 1, 2026.

More from