Claude Opus 5 Pricing 2026: Complete Cost Guide & Comparison

Claude Opus 5 launched today, July 24, 2026, priced the same as the model it replaces. That's the one-line version. The more useful version is how that price behaves once you put it next to every other model you'd realistically consider for the same job, inside Anthropic's own lineup and outside it. This guide lays out Opus 5's exact rate card, compares it against Sonnet 5, Haiku 4.5, Fable 5, and Mythos 5, then against OpenAI's GPT-5 family and Perplexity's Sonar models, and closes with the cost levers (effort setting, caching, batch, Fast mode) that actually decide what you pay per finished task.

If you're also tracking what Claude Code costs your engineering org day to day, see our Claude Code pricing guide.

Claude Opus 5 Pricing: The Rate Card

The headline number that matters: Opus 5 costs exactly what Opus 4.8 cost. Anthropic didn't raise the price for the new model, it raised the performance at the same price. On Frontier-Bench v0.1, Anthropic reports Opus 5 more than doubles Opus 4.8's score at a lower cost per completed task, and on CursorBench 3.2 it lands within 0.5% of Fable 5's peak score at half the cost per task. That "cost per task" framing matters more than the per-token rate, because Opus 5 also introduces an effort setting, explained below, that changes how many tokens a given task actually burns.

Claude Opus 5 vs. Anthropic's Other Models

Opus 5 sits in the middle of Anthropic's current five-model lineup: Haiku 4.5 at the bottom, Sonnet 5 in the middle, Opus 5 above that, and Fable 5 and Mythos 5 at the frontier tier. Here's the full internal comparison, current as of this launch.

Two things stand out here. First, Opus 5 is priced at exactly half of Fable 5 on every line item, and Anthropic is explicit that this is by design: Opus 5 is meant to close most of the capability gap to Fable 5 while staying at the Opus price point, not to replace Fable 5 for the hardest frontier work. Second, Sonnet 5's price increase on September 1, 2026 (from $2/$10 to $3/$15) is worth flagging now if you're forecasting Q4 spend, since it narrows the gap between Sonnet 5 and Opus 5 from 60% to 40% cheaper on output tokens.

Mythos 5 is listed at limited availability and matches Fable 5's price. Anthropic's own positioning is that Mythos 5 stays ahead of Opus 5 specifically on cybersecurity and biology research tasks, areas Anthropic has intentionally not optimized Opus 5 for. For general coding, knowledge work, and agentic tasks, Opus 5 is the better cost-to-capability trade unless your workload specifically needs that frontier bio/cyber ceiling.

Claude Opus 5 vs. OpenAI Pricing

OpenAI's GPT-5 family spans a wider price range than Anthropic's lineup, which makes a fair comparison depend on picking the right tier. Here's where Opus 5 lands against each.

The honest read: Opus 5 versus GPT-5.5, OpenAI's actual flagship, is close to a wash on sticker price, with Opus 5 winning on output tokens, which is usually where agentic and coding workloads spend the most. Versus GPT-5.5-pro, OpenAI's extended-reasoning tier, Opus 5 is not close, it's dramatically cheaper. The lighter GPT-5.1 and GPT-5.2 models are cheaper still, but they're not the same weight class, comparing them to Opus 5 on price alone without accounting for the capability gap is the kind of comparison that looks good in a headline and falls apart in production.

Claude Opus 5 vs. Perplexity Pricing

Perplexity's Sonar models are priced for a different job than Opus 5. Sonar is built for search-grounded question answering with live web retrieval baked into the API; Opus 5 is built for agentic coding, long-horizon reasoning, and knowledge work. They show up in the same procurement conversation often enough that the comparison is worth making explicit, and worth being honest about where it doesn't hold.

On pure per-token price, every Sonar tier undercuts Opus 5. That's expected, and it's not the useful takeaway. If your workload is "answer this question using current web sources," Sonar is the right tool and the cheaper one. If your workload is "write and debug this code," "run this multi-step agentic task," or "do this analytical work end to end," Sonar isn't built to compete there and Opus 5's price should be judged against Fable 5, GPT-5.5, and GPT-5.5-pro instead. Buying on price without matching the model to the job is how teams end up re-platforming twice.

What Actually Changes Your Opus 5 Bill

The $5/$25 rate is fixed, but four levers determine what you actually pay per completed task.

Effort setting (new in Opus 5)

Opus 5 introduces a low/medium/high effort toggle. It doesn't change the per-token rate, it changes how many tokens a task consumes to reach a given quality bar. Anthropic and early-access customers report meaningful token savings at lower effort settings without proportional accuracy loss. One legal-AI partner reported similar performance using 26% fewer tokens at lower reasoning levels compared to Opus 4.8 at max reasoning. This is the single biggest new cost lever in this release, and it means two teams running identical workloads on Opus 5 can see materially different bills based on effort defaults alone.

Prompt caching

Cache hits cost $0.50 per million tokens, a 90% discount off the $5 base input rate. For any workload with repeated system prompts, long documents, or multi-turn context, caching is the highest-leverage optimization available before you touch the model or the task design.

Batch API

Batch processing cuts both input and output pricing in half, to $2.50/$12.50 per million tokens. This only applies to asynchronous, non-time-sensitive work, but for eval runs, bulk content generation, or offline analysis, it's a straightforward 50% cut.

Fast mode

Fast mode runs Opus 5 at roughly 2.5x default speed for double the price, $10/$50 per million tokens. It's a legitimate trade for latency-sensitive product surfaces, and a bad default for anything that doesn't need the speed.

Tracking Claude Opus 5 Spend at Scale With Finout

None of the levers above are visible from a single invoice line. A team running Opus 5 across coding agents, customer-facing features, and internal tooling is really running three or four different cost profiles under one model name, split across effort levels, cache hit rates, Fast mode usage, and Batch API jobs. Per-session tools show what one call cost. They don't show which team, which feature, or which effort-level default is driving the trend.

This is what Finout's Anthropic integration is built for. Opus 5, Sonnet 5, and every other Claude model land in MegaBill alongside your cloud spend, reconciled down to the token.

What that unlocks for an Opus 5 rollout:

Allocation by team, feature, or product surface through Virtual Tags, so "who's spending on Opus 5 and why" has an answer without a re-tagging project.

Unit economics: cost per resolved ticket, cost per completed agent run, cost per feature shipped, the questions that turn a token bill into a business metric.

Anomaly detection that catches an effort-level misconfiguration or a Fast mode default left on in production before it shows up as a surprise on next month's invoice.

One ledger across Anthropic, OpenAI, and every other AI and cloud vendor you run, so the Opus 5 vs. GPT-5.5 cost comparison in this article is something you can actually verify against your own workloads, not just Anthropic's benchmark numbers.

Frequently Asked Questions

Pricing current as of July 24, 2026, Opus 5's launch date. Anthropic pricing sourced from the Claude Platform pricing docs and the Claude Opus 5 announcement. Rates change; verify current pricing before budgeting.