Claude Opus 5: When the Expensive Model Really Pays Off

On July 24, 2026, Anthropic released Claude Opus 5, the next generation of its mid-tier flagship model - at the unchanged Opus price of $5 per million input tokens and $25 per million output tokens, identical to its predecessor Opus 4.8. For companies that already run AI agents or are currently building them, that's more than a footnote. It shifts the calculation you use to decide which model should handle which task in your agent pipeline - and when reaching for the most expensive available model actually pays off.

In this article, we break down the figures Anthropic and independent benchmarks have published around Claude Opus 5, and translate them into a practical decision framework for your business - including a concrete cost calculation for a typical agent workflow, a migration checklist, and answers to the questions we've been asked most often in recent weeks.

Claude Opus 5: New model, same price

Claude Opus 5 succeeds Opus 4.8 without any change to the price tag: $5 per million input tokens, $25 per million output tokens. The context window stays at 1 million tokens - both as the default and the maximum - and maximum output length is 128,000 tokens. That's relevant for agent workflows that need to process long documents, codebases, or multi-step conversations: you're not paying more than before for more context capacity.

Three technical details matter especially for day-to-day operations. First, Opus 5 supports five so-called "effort" levels - low, medium, high, xhigh, and max - which control how deeply the model thinks and how intensively it uses tools. The per-token price stays identical at every level, but actual token consumption per task varies considerably. More on that below, because it's the lever most companies overlook first and discover last.

Second, Opus 5 has its own rate-limit quota pool, separate from the previous Opus 4.x group. If you shift traffic from older Opus models to Opus 5, you don't free up capacity on the old models by doing so - a point you should factor into capacity planning before migrating production workflows. In practice: check your API quotas for Opus 5 separately, rather than automatically assuming you can rely on existing Opus 4.x headroom.

Third, there's also a "Fast Mode" available as a research preview, exclusively via the Claude API: up to 2.5x higher output speed at a premium price of $10 / $50 per million tokens - in other words, at Fable 5's price point, but with the focus on speed rather than depth of thinking. For latency-critical agent use cases, such as live customer support or chat widgets with a tight response-time window, this can be a lever worth evaluating separately.

How close does Opus 5 get to the most expensive model?

The most interesting figure for your budget planning comes from the CursorBench 3.2 leaderboard, a benchmark for agentic coding and tool-use behavior. There, Claude Fable 5 - Anthropic's most expensive and most capable model - leads with 70.5 percent. Claude Opus 5 follows at 70.0 percent, a gap of just 0.5 percentage points. At the same time, Fable 5, at $10 / $50 per million tokens, costs twice as much as Opus 5.

On CursorBench 3.2, Opus 5 reaches roughly 99 percent of Fable 5's performance - at half the token cost. At maximum effort setting, the cost per agent task is even lower for Opus 5, at $2.03 versus $2.75 for Fable 5 - about 26 percent cheaper.

Other benchmarks show a similar picture: on SWE-bench Verified, Opus 5 reaches 96.0 percent; on the significantly harder SWE-bench Pro, 79.2 percent; and on ARC-AGI-3, the model sets a new record at 30.2 percent. For your practice, that means: on tasks that demand deep code understanding, multi-step problem-solving, or abstract reasoning, Opus 5 delivers results close to what the most expensive available model achieves - just at a noticeably lower cost per completed task.

Important for perspective: a 0.5 percentage point difference on a leaderboard usually isn't what decides success or failure of an agent workflow in practice. What's more relevant is the cost calculation per task - and at near-equal performance, that clearly favors Opus 5. This is especially true if you're running a lot of tasks per day through your agents: the difference between $2.03 and $2.75 per task adds up to a noticeable amount across hundreds or thousands of operations a month.

The four model tiers at a glance - and when each one fits

Anthropic currently offers four model tiers with clearly staggered prices per million tokens (input/output): Claude Haiku 4.5 at $1 / $5, Claude Sonnet 5 at $3 / $15 (introductory price $2 / $10 until August 31, 2026), Claude Opus 5 at $5 / $25, and Claude Fable 5 at $10 / $50. This tiering is the actual lever for your cost control - not the question of which single model is "the best."

For routine tasks like email sorting, simple data extraction, or classification, Claude Haiku 4.5 is generally more than sufficient - the model is five times cheaper than Opus 5, on both input and output. Claude Sonnet 5 suits medium-complexity work: summaries, structured text generation, simpler agent steps where speed and cost matter more than maximum depth. For multi-step planning, complex contract review, or code and document analysis with high error-cost risk, Opus 5 pays off, because the minimal gap to Fable 5 on CursorBench 3.2 is usually irrelevant in practice, while Opus 5 costs only half as much. Fable 5 remains reserved for exceptional cases: very high demands on autonomy and depth over long periods, such as complex research tasks or particularly demanding, long-running coding projects.

The relevant question for your Mittelstand workflow, then, isn't "Opus 5 or Fable 5?" but "which sub-task in my agent pipeline actually needs the most expensive model at all?" Anthropic itself now recommends testing the effort level separately per route rather than defaulting to the highest level across the board - more on that next.

Effort levels: the second lever many people overlook

Alongside model choice, there's a second, often underrated lever for cost control: the effort level within the same model. Opus 5 supports five levels - low, medium, high, xhigh, and max - which control how deeply the model thinks and how intensively it uses tools before delivering an answer. Crucial to understand: the price per million tokens stays exactly the same at every level. What changes is the actual token consumption per task - and with it, your bill at the end of the month.

A higher effort level means more internal "thinking" - reasoning steps, tool calls, self-checking - before the final answer. That raises quality on complex, ambiguous, or error-critical tasks, but it also drives up token consumption - sometimes substantially. According to a single source that has not been independently verified, the "medium" effort level already reaches near the accuracy of the highest levels for many task types, at roughly 1.5x the compute cost of baseline - compared to roughly 3x for "high" effort. Even though this specific figure should be treated with caution, it matches the general observation: not every task needs the highest depth of thinking, and the difference between levels is rarely linear.

In practice, that means: test per workflow step which effort level is sufficient. An agent that checks an invoice for completeness probably doesn't need "max" effort depth. An agent that assesses a contract clause for legal risk, where a wrong decision would be costly, genuinely benefits from "high" or "xhigh," by contrast. The lever thus works in two dimensions: first, you choose the right model for the task, and second, the right effort level within that model. Both decisions together determine your actual cost per completed operation - not the list price per million tokens alone.

Without instrumentation, though, this cost advantage stays theoretical. If you don't log which model and which effort level were actually used per workflow step, you can't determine the cost-per-task metric for your own processes at all - let alone optimize it. Simple logging of model, effort level, and token consumption per step is therefore the basic prerequisite for even knowing where a change is worthwhile.

Practical example: what model choice means for a customer case

Take a typical Mittelstand agent workflow: automated invoice checking or a quote comparison, where a single customer case triggers around 20 LLM calls - from field recognition through format checking to the final approval decision. A pattern observed in the industry - not officially from Anthropic, but an industry observation - is a split of roughly 80 percent routine calls going to cheaper models and around 20 percent to the expensive flagship model.

Applied to our example with 20 calls, that means: around 17 routine steps - field extraction, format checking, routing, simple plausibility checks - run on Claude Haiku 4.5, which is five times cheaper than Opus 5. Only the two or three critical decision points in the process - such as assessing an unusual contract clause or the final approval decision on a borderline case - run on Opus 5, where accuracy and contextual understanding make the difference. The result: overall costs for the case drop significantly, without sacrificing quality at the critical points.

To put that in perspective: according to an industry blog - not an official Anthropic source - token prices in 2026 in the German Mittelstand are often quoted at €0.03 to €0.12 per 1,000 tokens, with a single user request in agent workflows easily triggering 10 to 50 LLM calls, depending on process complexity. Even though these figures come from a single source and haven't been independently verified, they match the experience that the number of LLM calls per case in multi-step agent pipelines quickly reaches double digits - and that's exactly why deliberate model and effort assignment per step pays off, instead of deploying the most expensive available model everywhere by default.

A second example illustrates the lever from a different angle: picture a research agent that produces competitive analyses for your sales team. The research and extraction steps - evaluating websites, extracting figures, gathering sources - are repetitive and not error-critical; Sonnet 5, or even Haiku 4.5 at a low effort level, is enough here. Only the final synthesis - the actual strategic assessment that ends up in a management presentation - benefits from Opus 5 at a higher effort level, because misjudgments there would be costly. Here too: it's not the most expensive component at the start of the chain that matters, but targeted use at the point where quality has the biggest lever.

Switching to Opus 5: checklist and frequently asked questions

If you currently run Opus 4.8 or are thinking about switching to Opus 5, these points help with the practical rollout:

Recheck rate-limit quotas. Because Opus 5 has its own rate-limit pool separate from older Opus models, you should recheck your API quotas with the provider - directly or via a wrapper service - rather than automatically assuming you have existing Opus 4.x headroom. Switching without this check can lead to unexpected throttling in production workflows.

Test the effort level per route, don't set it across the board. Don't automatically start with "max" or "high" for every use case. Test per workflow step which level is sufficient, and document the result - that way you build your own reference table for your processes over time.

Model mix instead of model monoculture. Check for every step in your agent pipeline whether it really needs Opus 5, or whether Haiku 4.5 or Sonnet 5 is enough. That's the biggest lever for cost control - bigger than the choice between Opus 5 and Fable 5.

No need to rush to Fable 5. For most Mittelstand use cases, even demanding contract review or financial analysis, Opus 5 is likely to offer the better price-performance ratio than the pricier Fable 5 flagship, at half the cost and near-identical benchmark performance. Fable 5 is better suited to exceptional cases with very high demands on depth and autonomy over very long periods.

Why doesn't the price change? A question we get often: how can Anthropic offer a more capable model at the same price as its predecessor? Part of the answer lies in infrastructure: Anthropic has secured, under a contract running through May 2029, with a volume of around $1.25 billion a month and more than $40 billion in total, the entire compute capacity of SpaceX/xAI's Colossus 1 data center in Memphis, Tennessee: over 300 megawatts from more than 220,000 Nvidia GPUs. This scale of proprietary compute infrastructure shifts what's possible in model performance at a stable price. For your decision-making, this is mainly relevant as context: the current price stability alongside rising performance isn't a coincidence, but part of a longer-term capacity strategy - a sign that while you shouldn't expect constant drastic price jumps when choosing a model, you should expect further performance leaps at stable prices.

What about "Fast Mode"? Fast Mode is worthwhile mainly where users are actively waiting for an answer - such as in a live chat widget. For asynchronous agent workflows, where nobody is sitting in front of a screen waiting, the premium usually isn't justified; here, the cost-per-task metric matters more than response speed.

Claude Opus 5 doesn't change the calculation through a lower entry price, but through a better ratio of performance to cost compared with the most expensive available model. For most Mittelstand applications, that means: Opus 5 is the new standard for everything that needs real depth of thinking, while Haiku 4.5 and Sonnet 5 handle the routine steps. The real competitive advantage isn't using the most expensive model - it's finding the right combination of model and effort level for every step of your agent pipeline, and measuring it consistently.