Claude Opus 5: Near-Frontier Intelligence at Half the Cost

Anthropic shipped Claude Opus 5 on July 24, 2026, and the pitch is unusual for a flagship launch: it is not the smartest model the company makes. Instead, Opus 5 aims to deliver near-frontier intelligence at roughly half the cost of Anthropic's top model, Fable 5, with a new "effort" dial that lets you trade compute for savings on every request. For anyone running AI agents at scale, that trade-off matters more than another benchmark record. Here is what shipped, what it costs, and where it fits.

What is Claude Opus 5?

Claude Opus 5 is an update to Anthropic's hybrid reasoning model line, tuned for coding, knowledge work, and agentic tasks. It became available across all of Anthropic's platforms on launch day: the new default model on Claude Max, the strongest option on Claude Pro, and accessible to developers through the Claude API.

The headline framing from Anthropic is efficiency, not peak capability. The company positions Opus 5 as a "thoughtful and proactive" model that comes close to the frontier intelligence of Claude Fable 5 at half the price. In other words, this is a release aimed at the middle band of everyday work — the coding tickets, document analysis, and multi-step agent runs that make up most real usage — rather than the hardest research-grade problems.

Claude Opus 5 pricing

Pricing is unchanged from the previous generation. According to InfoWorld's report, Opus 5 costs:

• $5 per million input tokens
• $25 per million output tokens

That is the same as Opus 4.8. The efficiency gain does not come from a lower sticker price — it comes from Opus 5 doing more per token, plus the effort dial letting you spend fewer tokens when you do not need maximum reasoning depth.

The effort dial: cost vs. capability

The most practically important feature is the adjustable effort setting. It lets you tell the model how much computing power to devote to a task, so you can balance intelligence against speed and token cost on a per-request basis.

At lower effort levels, Anthropic says Opus 5 can preserve much of its performance while using fewer tokens and costing less to run. At the top of the range — high, xhigh, and max effort — it pushes toward frontier quality. This is a direct response to a real pain point: enterprise customers have complained about unpredictable AI bills, and giving users a low/medium/high knob puts cost control back in their hands.

If you have ever watched an agent burn through a token budget on a task that did not need deep reasoning, the effort dial is the feature to pay attention to.

Claude Opus 5 benchmarks

Anthropic leads with cost-adjusted performance rather than peak scores alone. A few concrete numbers from the launch:

• On Frontier-Bench v0.1, Anthropic says Opus 5 surpasses all other models and more than doubles Opus 4.8's performance at a lower cost per task.
• On CursorBench 3.2 at max effort, Opus 5 reportedly lands within 0.5% of Fable 5's peak score — at half the cost per task.
• Anthropic also claims better performance at a given cost than any other model at high, xhigh, and max effort settings.

The through-line: charts emphasize performance per dollar, not just the top of the leaderboard. Independent benchmarks will take time to confirm these claims, so treat vendor numbers as a starting point rather than a verdict.

Where Opus 5 sits in Anthropic's lineup

Opus 5 is the fourth model Anthropic has released in a short span, following Mythos 5, Fable 5, and Sonnet 5. It is important to be clear about the positioning:

• Fable 5 remains the smartest model, and Anthropic still recommends it for the most advanced projects — including long-horizon tasks a model might run autonomously for days.
• Opus 5 is the efficient default: near-frontier quality for the bulk of everyday coding, agent, and knowledge work at half the cost.
• Mythos 5 still leads Opus 5 on certain dual-use domains like biological research and offensive cybersecurity — an area where Anthropic deliberately held Opus 5 back.

Anthropic also describes Opus 5 as its most safety-aligned Opus model, reporting lower rates of deceptive behavior and stronger adherence to its model constitution than prior releases. On the developer side, the launch shipped alongside two beta updates: mid-conversation tool changes and automatic API fallbacks.

Why this matters for AI agents

For agent builders, the interesting story is not the benchmark table — it is the economics. Multi-agent systems multiply token usage: every planner, worker, and reviewer agent spends tokens, and a single complex run can chain dozens of model calls. At that scale, "near-frontier at half the cost, with a dial to spend even less on easy steps" is a bigger deal than a two-point benchmark gain.

Practically, that means you can:

• Route by difficulty. Run low or medium effort on routine sub-tasks (formatting, extraction, simple edits) and reserve high effort for the steps that actually need reasoning.
• Cap burn without capping quality. The self-verification and error-recovery Anthropic highlights mean fewer wasted retry loops, which is where agent budgets quietly disappear.
• Reserve Fable 5 for the hard 10%. Keep the expensive model for genuinely long-horizon or research-grade work, and let Opus 5 handle the rest.

This is the same logic behind model routing in coding tools — spend frontier-level compute only where it changes the outcome. We covered that pattern in Cursor Router Explained, and Opus 5's effort dial brings the same cost-quality control down to the individual request.

Should you switch to Opus 5?

A quick decision guide:

• Using Claude Max or Pro? You already have it — Opus 5 is the new default on Max and the top model on Pro.
• Building agents or coding workflows on the API? Opus 5 is likely your new workhorse; test the effort dial against your own tasks to find the cheapest setting that still passes your evals.
• Running long, autonomous, high-stakes jobs? Keep Fable 5 in the mix for that slice, per Anthropic's own guidance.

The smart move is not "Opus 5 for everything" — it is matching effort level and model to the difficulty of each task.

Do this with your own AI workforce

Opus 5's real lesson is about orchestration: point frontier-level compute at the steps that need it, and spend less everywhere else. That is exactly how a multi-agent workforce should run. Eigent is an open-source, local-first Cowork app that coordinates a team of AI agents across real workflows — and because it is bring-your-own-key, you can plug Opus 5 in and control effort and cost per task instead of paying frontier rates for routine work. Explore how it handles reviewing GitHub PRs, or download Eigent and put Opus 5 to work across your own agent pipeline.