Claude Opus 5 Review: The AI Upgrade That Doesn't Cost More

On July 24, 2026, Anthropic dropped a surprise that few in the AI community saw coming. Not because the release of Claude Opus 5 was unexpected – everyone knew it was in the pipeline. The surprise was the price tag. In an era where every new model seems to come with a bigger bill, Anthropic held the line. Opus 5 costs exactly what Opus 4.8 did: $5 per million input tokens and $25 per million output tokens. For a model the company describes as coming “close to the frontier intelligence of Claude Fable 5 at half the price,” that’s a statement.

The timing matters too. The past month has seen a relentless wave of releases – Fable 5, GPT-5.6, Sonnet 5 – each promising frontier intelligence and new benchmarks to study. Users and developers alike are beginning to feel what some are calling “intelligence fatigue.” The question is no longer just “how smart is it?” but rather “what can I actually do with all this intelligence, and what will it cost me?”

What Opus 5 Brings to the Table

Anthropic has positioned Opus 5 as the everyday workhorse. It’s not the trophy model – that title belongs to Fable 5. But it is now the strongest model available to regular users on Claude Pro, and the new default on Claude Max. Here’s what you get:

A 1-million-token context window, capable of processing entire codebases or lengthy documents in a single pass.
Up to 128,000 tokens of output, allowing for extended reports, full feature code, or multi-section documents without interruption.
Thinking turned on by default. The model no longer waits for permission to reason – it checks its work, verifies its steps, and plans before executing.
Five effort levels that give users granular control over how deeply the model thinks, directly impacting both quality and cost.
The same pricing as its predecessor, keeping the door open for individual developers and smaller teams, not just enterprise budgets.

Where Opus 5 Excels – and Where It Stumbles

Anthropic took an unusual approach with this release. Rather than publishing a single performance number per benchmark, they released full cost-versus-performance curves across all five effort levels. This transparency lets users decide not just whether Opus 5 is good, but whether it’s good at the budget they can afford.

Clear wins:

On Frontier-Bench v0.1, which tests agentic terminal coding, Opus 5 scored 43.3%. That’s a decisive lead over Fable 5 at 33.7% and GPT-5.6 Sol at 34.4%, and more than double the 21.1% managed by Opus 4.8.
On ARC-AGI 3, a benchmark specifically designed around novel problems the model has never encountered, Opus 5 hit 30.2%. GPT-5.6 Sol managed just 7.8%, while Opus 4.8 barely registered at 1.5%.
On AutomationBench, which measures end-to-end business workflow completion, Opus 5 reached 26%, leaving a field of competitors clustered between 17% and 18%.

Where it falls short:

On DeepSWE v1.1, a standard agentic coding evaluation, Opus 5 scored 68.8%, trailing GPT-5.6 Sol at 72.7%. This is the most notable loss in the official table, and it’s on the very territory Opus 5 is marketed for.
On HealthBench Professional, it landed at 59.8%, behind both Mythos 5 at 66% and GPT-5.6 Sol at 60.5%.
On the Legal Agent Benchmark, it managed just 11.7% – slightly behind Fable 5 at 13.3%, though both far outpaced GPT-5.6 Sol’s 2.5%.

The pattern is telling. Opus 5 dominates when tasks are long-horizon, multi-step, and require the model to maintain its own state across extended chains of reasoning. It draws or loses when the task is a single hard question answered in one pass. As one observer noted, this is a model built for marathons, not sprints.

What Changes When You Switch From Opus 4.8

If you’re migrating from the previous generation, there are two breaking changes you need to know about before changing a single line of code.

First, thinking is now on by default. On Opus 4.8, a request that didn’t specify a thinking setting ran without thinking. On Opus 5, that same request now thinks. Since the max_tokens parameter caps both thinking and the visible response combined, any workload that had max_tokens tightly tuned to its answers on 4.8 may now find its responses cut off mid-sentence. Revisit every route that never set a thinking field.

Second, disabling thinking is now blocked at higher effort levels. Sending thinking: disabled together with xhigh or max effort returns a 400 error. This is a deliberate design choice – Anthropic’s documentation advises against disabling thinking on this model entirely, and documents specific failure modes when you do, including phantom tool calls that never execute and leaked internal tags appearing in visible output.

Beyond these breaking changes, users will notice behavioral differences. Opus 5 writes longer responses by default. It verifies its own work without being asked. It delegates to sub-agents more readily in multi-agent setups. And it narrates its progress more frequently during long tasks – which can be either helpful or exhausting, depending on whether you’re watching or just waiting for the result.

What Early Users Are Saying

The reviews that poured in during the first week paint a surprisingly consistent picture: extraordinary output, frustrating process.

One tester captured the sentiment with a blunt admission: “I love it and I hate it.” Opus 5 produced front-end designs and application prototypes that earned perfect scores in blind evaluations, outperforming every other model tested. The same tester, however, found themselves repeating instructions multiple times, asking the model to stop apologizing and just execute.

The verbosity is a recurring complaint. Ask for a simple task, and the model returns paragraphs of explanation before it begins. Ask for an opinion, and the answer arrives wrapped in layers of caveats and disclaimers. One reviewer coined a term for it: “Claude Slop” – the tendency of Anthropic’s more intelligent models to produce prose that is technically correct but exhausting to read.

Yet there’s also widespread admiration for the model’s autonomous initiative. In one documented case, Opus 5 was given a drawing of a machine part and asked to rebuild it as a 3D model, but was intentionally given no way to view the image directly. The model responded by building its own computer vision pipeline from scratch, extracting the geometry from raw pixels, and completing the task. No competing model with the same constraints could solve it.

The practical takeaway from those who’ve spent real hours with Opus 5: if you plan to read everything it writes, prepare for frustration. If you plan to let it work in the background and review only the final output, the odds are high you’ll be impressed.

The Five Effort Levels: Where the Real Value Lies

The headline numbers for Opus 5 are all recorded at the top of the effort ladder. But the ladder itself – five distinct settings from low to max – is the feature that changes how this model should be used in production.

Low: Minimal token consumption and maximum speed. The right setting for repetitive, high-volume tasks like text classification or summarizing short passages, where depth of reasoning adds cost without adding value.
Medium: The first real balance point. It adds meaningful reasoning depth without a steep cost increase, making it suitable for routine daily work that doesn’t demand rigorous verification.
High: The default recommended starting point for everything except complex coding. At this level, the model’s analytical and writing capabilities begin to show their true strength.
Xhigh: The sweet spot for agentic coding and long-horizon projects. This is where Opus 5 posted its headline-grabbing Frontier-Bench and AutomationBench results. Anthropic specifically recommends this level for programming and extended planning tasks.
Max: The deepest thinking available. But here’s the counterintuitive part – it’s not always better. On Frontier-Bench, Xhigh slightly outperformed Max while costing less. The official documentation warns that Max “can be prone to overthinking,” burning tokens on details that don’t improve the final output.

The practical advice from experts who’ve run extensive evaluations: don’t assume Max is the “pro mode.” Start at High or Xhigh, run your own tests, and adjust up or down based on what you observe. A substantial portion of production traffic can likely move down a rung without any measurable quality loss.

Under the Hood: New Features Worth Noting

Alongside the model itself, Anthropic shipped three additions that matter more to developers than casual users, but whose effects reach everyone:

Mid-conversation tool changes: You can now add or remove the tools the model can use between turns without invalidating the prompt cache. Previously, any change to the tool list meant re-billing the entire prefix. This translates directly to faster and cheaper conversations.
Automatic fallbacks: When a safety classifier blocks a request, it can now be routed automatically to another model instead of being rejected outright. For production systems, this means fewer hard failures and a smoother user experience.
Lower prompt cache minimum: The threshold dropped from 1,024 tokens to 512. Short prompts that were previously too small to cache now benefit automatically, with zero code changes required.

Safety: Finding Vulnerabilities Without Exploiting Them

Anthropic has drawn a deliberate line with Opus 5 on cybersecurity. The model is nearly as good as Mythos 5 – Anthropic’s most restricted model – at finding software vulnerabilities. On the OSS-Fuzz evaluation, the two models identify issues at comparable rates.

But when it comes to developing exploits – turning a discovered vulnerability into a working attack – Opus 5 lags far behind. This is by design. Anthropic explicitly avoids training its models on offensive cyber tasks. The ability to find vulnerabilities came along with general intelligence; the ability to exploit them did not.

For practical use, this means Opus 5 can scan source code for security flaws, but it will refuse requests involving binary-based vulnerability scanning, penetration testing, or exploit generation. The company also reports that these safety classifiers intervene roughly 85% less often than they do for Fable 5 – a direct response to complaints about overly aggressive filtering on the more restricted model.

Who Should Actually Use Opus 5?

Not everyone needs this model. Opus 5 is built for specific use cases:

Developers working on long-horizon projects: multi-file refactors, full feature builds, and tasks that require sustained reasoning across extended sessions.
Teams running autonomous AI agents that need to plan, execute, and verify their own work with minimal human intervention.
Anyone who wanted Fable 5’s intelligence but couldn’t accept its 30-day data retention requirement – Opus 5 carries no such obligation for general access.
Users who regularly process very long documents or need to turn lengthy reports into structured presentations.

Outside these categories, alternatives may be more practical. For maximum speed at minimum cost, Claude Sonnet 5 closes much of the capability gap at $3 per million input tokens. For single-pass coding evaluated the way DeepSWE measures it, GPT-5.6 Sol still holds a defensible lead. And for clinical or health-professional content, Mythos 5 remains the stronger option.

The Bottom Line

A week after launch, three messages stand out from the noise.

First, the price freeze is the real story. Keeping the same price as the previous generation while delivering a step-change improvement in performance signals that Anthropic is playing a volume game, not a margin game. The strategy is adoption, not extraction.

Second, transparency has become a product feature. Publishing cost curves instead of single scores, leaving losing rows in the comparison table, and openly discussing the overthinking problem – these are choices that build trust over time, even when they don’t paint the rosiest picture.

Third, users now have genuine granular control. The ability to choose the intelligence level appropriate for each task, rather than paying for maximum capability on every request, changes the economics of using AI in production. Those who invest time in tuning effort levels to their specific workloads will get better results than those who simply pay more.

Claude Opus 5 is not the best model at everything. But for the long, complex, multi-step tasks that form the backbone of serious professional work, it currently offers the strongest combination of capability and cost on the market. And that combination wasn’t available with this clarity before its release.