Claude Prompt Caching Pricing: 75% Cheaper Cache Reads (2026)

Claude prompt caching pricing just dropped by 75% for cache reads: as of the Claude Fable 5.1 launch on 1 September 2026, Anthropic charges $0.25 per million cached input tokens instead of $1.00, which the company says cuts total costs by roughly 25% on typical workloads and up to 45% on highly agentic ones. If you run AI agents, coding assistants or automation pipelines on Claude, this is the most important pricing change of the year — and you get it without changing a single line of code beyond pointing at the new model.

📺 Watch: Claude Fable 5.1 Just Set a New AI Performance Record

🔥 Get the Agent OS as a free bonus: AI Profit Boardroom members get the full Agent OS zip, prompt libraries, daily tutorials and weekly live coaching calls. → Get inside · Want AI SEO help 1-on-1? Book a free SEO strategy session →

All figures in this guide come from Anthropic's official Claude Fable 5.1 and Mythos 5.1 announcement, published 1 September 2026. Here is what changed, what stayed the same, and how to work out what your own bill will look like.

Claude Prompt Caching Pricing: What Changed on 1 September 2026

Prompt caching lets Claude store the parts of your prompt that repeat between requests — system prompts, tool definitions, long documents, conversation history — so subsequent requests re-read them at a discount instead of paying full input price every time. Until now, cache reads on Anthropic's frontier tier cost $1.00 per million tokens. With Claude Fable 5.1, that drops to $0.25 per million tokens, a 75% reduction, per the official announcement.

The rest of the price card is unchanged from Claude Fable 5:

• Input tokens: $10 per million
• Output tokens: $50 per million
• Cache reads: $0.25 per million (previously $1.00)

That asymmetry is the whole story. Standard input is now 40 times more expensive than a cache read. The more of your prompt that hits the cache, the closer your effective input cost gets to pennies — which is why Anthropic quotes different savings for different workload shapes: around 25% lower total cost for typical workloads, and up to roughly 45% for highly agentic workloads that replay huge contexts on every step.

Why Agent Workloads Benefit Most From the New Caching Price

An agent loop is almost perfectly built for prompt caching. Every iteration re-sends the same system prompt, the same tool definitions and the entire conversation so far, then adds one new tool result on the end. On a long run, well over 90% of the tokens Claude reads are tokens it has already seen — all billable as cache reads.

Run the arithmetic on a mid-sized example. Say your agent burns through 20 million input tokens a day and 90% of them are cache hits:

• Old pricing: 18M cached at $1.00 = $18.00, plus 2M fresh at $10 = $20.00 — about $38 a day on input.
• New pricing: 18M cached at $0.25 = $4.50, plus 2M fresh at $10 = $20.00 — about $24.50 a day on input.

That is a 35% cut on the input side of the bill from the cache change alone, and the saving scales with how context-heavy your agent is. Anthropic's own "up to ~45%" figure for total cost assumes the kind of long-context, many-step agent runs that tools like scheduled Hermes agent cron jobs produce all day long — the exact workloads that used to make frontier models feel too expensive to leave running.

It also changes the mental maths on model choice. Cheaper rivals have competed hard on raw token price — DeepSeek even discounts by time of day, which we broke down in the DeepSeek V4 off-peak pricing guide — but a 75% cache-read cut narrows the gap significantly for cache-friendly workloads while keeping frontier-level capability.

If you want to build AI agents that actually pay for themselves instead of burning your budget, check out the AI Profit Boardroom → get the full agent playbooks inside. Prefer 1-on-1 help with your AI SEO and automation stack? Book a free SEO strategy session and map it out with Julian's team.

What You Get for the Money: Claude Fable 5.1 in Numbers

The pricing cut landed alongside a genuine capability jump, so you are paying less per token for a stronger model. According to the benchmark table in Anthropic's 1 September 2026 announcement, Claude Fable 5.1 scores:

Two customer quotes from the announcement are directly about cost efficiency: Every reports Fable 5.1 running "twice as fast as Opus 5, half the tokens", and Rogo says it "matches Fable 5 accuracy with 20% fewer tokens". Fewer output tokens matter because output, at $50 per million, is the one line the caching change does not touch. For how the Claude tiers stack up against each other in practice, the Claude Opus 5 vs Fable 5 comparison covers the trade-offs the benchmarks hide, and the Goldie Bench write-up covers how these frontier brains compare in hands-on agent tests.

How to Actually Capture the Caching Savings

The discount only applies to tokens that hit the cache, so cache-hostile prompt design still pays full price. A few rules of thumb for getting your effective Claude prompt caching pricing as low as possible:

1. Put stable content first. Caching works on prefixes. Keep your system prompt, tool definitions and reference documents at the top of the prompt, and append the changing parts — new messages, new tool results — at the end.
2. Do not churn your system prompt. Injecting a timestamp or a random ID near the top of the prompt invalidates the cached prefix on every request. Move anything volatile as late in the prompt as it can live.
3. Keep sessions alive. Cache entries expire after a period of inactivity. Agents that poll or batch work in bursts inside one session cache far better than jobs that cold-start hourly.
4. Watch your cache-hit rate, not your token count. The metric that predicts your bill is the share of input billed at $0.25 instead of $10. Most API dashboards and agent frameworks now surface it.

If you run open-source agent stacks, the same prefix-stability logic applies whether the backend is Claude or something like Grok's tooling — the Grok Build guide shows a similar pattern on the xAI side — and the Agent OS resource covers how to structure agent prompts so the stable layers sit where caching can see them.

Where Claude Fable 5.1 Is Available

Per the announcement, Claude Fable 5.1 is live from 1 September 2026 on the Claude API under the model ID claude-fable-5-1, plus claude.ai, Claude Code, Amazon Web Services, Google Cloud and Microsoft Azure. The new cache-read price applies through the API, so anything built on it — agents, SaaS tools, automation pipelines — inherits the saving as soon as it moves to the new model. Its restricted sibling, Claude Mythos 5.1, is the same underlying model with permissive safeguards, limited to vetted cyber and life-science professionals through Anthropic's verification programmes — not something a typical builder needs.

What the New Caching Price Means for AI Business Margins

If you sell anything built on Claude — an agent service, a content pipeline, a SaaS feature — the caching change lands directly on your gross margin. A product whose cost of goods was dominated by cached context just saw that line shrink by three-quarters, and Anthropic's blended estimates of 25% to 45% total savings are wide enough to cover most real products. That creates three options: pocket the margin, cut your price to undercut competitors still quoting on old economics, or spend the savings on more model calls per customer — more retries, deeper reasoning, longer agent runs — at the same unit cost. For agency-style work, it also changes what you can promise: always-on monitoring agents and daily automated deliverables that were marginal at $1.00 cache reads are comfortably profitable at $0.25.

Two adjacent details from the same announcement matter for business users. Anthropic's Enterprise Frontier Safeguards programme, rolling out from autumn 2026, lets eligible customers run Fable 5.1 with zero data retention — relevant if client contracts restrict where prompts can be stored. And the improved safeguards fire far less often on legitimate work: Anthropic reports 60% fewer false positives per session on cybersecurity tasks and 85% fewer on benign biology requests versus Fable 5, which means fewer interrupted runs burning tokens on refused work.

Is Claude Worth It Now Compared to Cheaper Subscriptions?

The honest answer depends on workload shape. If your usage is light and conversational, flat-rate coding plans from Chinese labs can still undercut everything — we ran that maths in is the GLM coding plan worth it. But for context-heavy agents that read far more than they write, the new cache economics turn Claude from "premium splurge" into a defensible default: a frontier model whose effective input price on a well-cached agent loop lands near budget-model territory, per the official figures above. The right move this week is simple — check your cache-hit rate, estimate your blended input price at $0.25 versus $10, and re-run the comparison you last did when cache reads cost four times more.

If you want the exact agent setups, prompt libraries and cost-optimised workflows Julian's community uses to turn AI into income, check out the AI Profit Boardroom → join the AI Profit Boardroom here. And if you would rather have a personal roadmap first, book a free SEO strategy session and get your questions answered live.

Real wins from inside the AI Profit Boardroom

See all 3,000+ members →

← Back to all posts