Claude Opus 5 Pricing and the Effort Toggle: How to Cut Your AI Bill (2026)

Claude Opus 5 launched on July 24, 2026 as Anthropic's everyday workhorse for coding, agentic tasks, and knowledge work — positioned to come close to the frontier intelligence of Claude Fable 5 at roughly half the price. For engineering teams running these models in production, the headline benchmark numbers matter less than one practical question: what will this actually cost, and how do you keep the bill under control?

This is a cost guide, not a launch recap. If you want the full capabilities breakdown, benchmarks, and availability details, read the Claude Opus 5 launch guide. Here we focus on pricing mechanics, the new effort toggle, and a concrete strategy to cut your spend without cutting quality.

What does Claude Opus 5 actually cost?

Opus 5 keeps the same list price as Opus 4.8 and undercuts Fable 5 on input by 2x. There are two speed tiers: a standard mode and a "fast mode" that costs 2x but runs about 2.5x faster. Here is the full picture in dollars per million tokens.

The context window is 1M tokens by default and maximum — there's no smaller, cheaper variant to size down to. Max output is 128K tokens standard, or up to 300K via the Message Batches API with a beta header. Extended thinking is on by default. Those defaults are why the effort toggle, not the sticker price, is where cost is really decided.

How does the effort toggle change your real cost?

The effort toggle is a per-request parameter with three settings — low, medium, and high — that controls how much reasoning effort and compute the model spends before answering. Low is faster and cheaper for routine tasks; high runs longer reasoning for genuinely hard problems. Anthropic frames it plainly as the lever to "balance cost and capability."

Why does that lever move the bill so much? Because extended thinking is billed. The tokens the model generates while reasoning through a problem count as output tokens at the $25/M rate. A high-effort request doesn't just take longer — it generates far more billable output before it produces the final answer. Two requests with an identical prompt can differ 3x or more in cost purely on effort setting.

Why do thinking tokens matter for your bill?

Here's the mechanism that catches teams off guard. On a 40K-token input prompt, the input side costs the same regardless of effort — 0.04M × $5 = $0.20. The variable is the output. On low or medium effort a task might emit 6K output tokens; on high effort the same task might emit 20K because most of the extra is reasoning. That output difference is where your money goes.

Same prompt, 2x the cost, decided entirely by one parameter. The takeaway isn't "always use low" — some tasks genuinely need the deeper reasoning to get right on the first pass. The takeaway is that effort is a dial you should be setting deliberately per workload, not leaving on a single global default.

What does a real cost example look like?

Say your team runs 1,000 agentic coding tasks a day through Opus 5 — code review passes, small refactors, test generation, doc updates. Most of that work is routine and gets the right answer on low or medium effort. A minority is genuinely hard: gnarly debugging, architectural changes, multi-file reasoning that benefits from high effort.

Using the per-task figures above, compare running everything on high effort against a routing strategy that sends 80% of traffic through low/medium and escalates only the hard 20% to high.

Forty percent off the bill, with no drop in quality on the work that actually needed high effort — because that 20% still gets it. The engineering task is building the router: a cheap classifier or a few heuristics that decide which bucket each request belongs in before it hits the model. Opus 5 makes that easier than it used to be, because it self-verifies and recovers from its own errors without user intervention, so low/medium effort clears more of your workload than it would on a weaker model.

When should you pay for fast mode?

Fast mode doubles the token price ($10 input / $50 output) in exchange for roughly 2.5x throughput. That trade only makes sense when latency is on the critical path of something that earns or saves more than the premium costs — an interactive coding assistant where developers are waiting, a customer-facing agent, or a batch job with a hard deadline. For background jobs, overnight batches, and anything a queue can absorb, standard mode is the right default and fast mode is pure waste. Note the two dials are independent: you can run fast mode on low effort, or standard mode on high effort, and pick the combination each workload actually needs.

Which cost-adjacent wins should you factor in?

Two properties of Opus 5 lower your total cost in ways that don't show up on the per-token line item:

• No 30-day data retention. Unlike Fable 5, Opus 5 is not subject to the 30-day data-retention policy. For teams whose compliance or legal review would otherwise gate a model on retention terms, that removes a real procurement cost and a source of delay.
• ~85% fewer spurious refusals. Anthropic expects the safety classifiers to trigger about 85% less often than on Fable 5. Every false refusal is a wasted round-trip: you pay for the tokens, get nothing usable, and burn a retry. Fewer refusals means fewer wasted calls, which is a quiet but real cost reduction on top of the list price. A new "Automatic Fallbacks" safeguard also routes a tripped request to a less-powerful model and returns a usable response instead of an outright error.

How does Opus 5 pricing compare to Fable 5 and Opus 4.8?

The clean summary: Opus 5 holds Opus 4.8's price while delivering a large capability jump, and it sits at half of Fable 5's input cost while landing close on most benchmarks. On Frontier-Bench v0.1 (agentic coding) Opus 5 scores 43.3% versus Fable 5's 33.7% and more than doubles Opus 4.8. On SWE-bench Pro it hits 79.2%, under a point behind Fable 5 (80.0%) and Mythos 5 (80.3%), and far ahead of Opus 4.8's 69.2%. For the large majority of coding and knowledge-work traffic, that makes Opus 5 the value pick — reserve Fable 5 for the long-horizon autonomous agents Anthropic still recommends it for, and Mythos 5 for frontier cybersecurity and the hardest biology research.

If you're deciding between these models for a real workload, the model-by-model tradeoffs are covered in the Claude Fable 5 launch guide and the Claude Opus 4.8 launch guide. For patterns on wiring effort routing into a production agent, our AI coding agents guide walks through the orchestration side.

Putting it together

Opus 5's pricing story is simple once you see past the sticker price: the $5/$25 rate is the floor, and the effort toggle is what you actually pay against. Set effort deliberately per workload, route the routine 80% through low/medium, escalate the hard 20% to high, and keep fast mode for latency-critical paths only. That single discipline is worth a ~40% cut on a realistic agentic workload — before you count the savings from fewer refusals and no retention overhead.

Building an AI-heavy product and want engineers who already know how to cost-tune model routing, effort budgets, and agent orchestration? Hire vetted remote developers through Codersera and extend your team with people who ship production LLM systems, not proofs of concept.

FAQ

How much does Claude Opus 5 cost per million tokens?

In standard mode, Opus 5 costs $5 per million input tokens and $25 per million output tokens — the same as Opus 4.8. Fast mode is $10 input / $50 output for roughly 2.5x the speed. Fable 5's input price is $10, twice Opus 5's.

Does the effort toggle actually change what I pay?

Yes, significantly. The low/medium/high effort setting controls how much reasoning the model does, and those reasoning ("thinking") tokens bill as output at $25/M. A high-effort request can cost 2x or more than the same prompt on low effort, purely from the extra generated tokens.

How do I cut my Claude Opus 5 bill without losing quality?

Route most traffic through low or medium effort and escalate only the genuinely hard requests to high. On a 1,000-task-per-day workload, an 80/20 split can cut the bill about 40% versus running everything on high, while the hard 20% still gets full reasoning.

When is fast mode worth the 2x price?

Only when latency is on the critical path — interactive coding assistants, customer-facing agents, or deadline-bound batches. For background and queued work, standard mode is the correct default and fast mode is wasted spend. Effort and speed are independent dials, so you can mix them per workload.

Is Opus 5 cheaper to run than Fable 5?

On input, yes — Opus 5 is half of Fable 5's price and lands close on most coding and knowledge-work benchmarks, making it the value pick for the bulk of traffic. Opus 5 also avoids Fable 5's 30-day data-retention policy and triggers roughly 85% fewer spurious refusals, both of which lower total cost. Reserve Fable 5 for long-horizon autonomous agents.