Claude Opus 5: what Anthropic's new model changes for software development
Back to the blog
Artificial Intelligence

Claude Opus 5: what Anthropic's new model changes for software development

Jérémy
Jérémy
Founder & CTO
30 July 20266 min read

Anthropic launched Claude Opus 5 at the end of July 2026. We look at what this new model concretely brings to software development, from more reliable assisted coding to autonomous agents.

At the end of July 2026, Anthropic unveiled Claude Opus 5, its new flagship model. Behind the announcement lies a clear positioning: a model described as “thoughtful and proactive”, able to come close to the performance of Anthropic's most advanced model (Claude Fable 5) for roughly half the cost. For an agency like METEORA, which builds generative AI into the way it designs projects, this kind of leap is worth a closer look.

A model built for consistency, not just raw performance

On Frontier-Bench v0.1, Opus 5 outperforms every other model and scores more than double Opus 4.8, at a lower cost per task. On ARC-AGI 3, a problem-solving benchmark, its score is three times higher than the closest competing model. But the most interesting figure is not a raw score: it is the lower variance from one run to the next. A more consistent model is one on which you can build reliable processes, not just impressive demos.

The effort setting: intelligence on demand

A notable new feature: Opus 5 introduces an effort setting (low, medium, high) that lets you adjust the balance between the model's intelligence and the real cost of a task. At maximum effort, the model approaches the performance of Fable 5. At reduced effort, it remains very token-efficient for everyday tasks. The price has not changed from Opus 4.8: $5 per million input tokens and $25 per million output tokens, with a Fast variant that is 2.5 times faster at double the base price.

What this changes for code

On CursorBench 3.2, Opus 5 comes within 0.5% of Fable 5's best score, for half the cost per task. What strikes the developers who have tested it most is its ability to check its own work the way a real front-end developer would: opening the page it produced in a browser at different screen widths, spotting misaligned or off-screen elements, and fixing them before handing back. A trading firm reports having built a market data feed for a new stock exchange in a single session, a task considered beyond the reach of previous models.

Towards truly autonomous agents

This may be where the real change of scale lies. Opus 5 holds up better over long, multi-step workflows. One example documented by Anthropic: a monitoring agent checked one of its own hypotheses in production, found that a signal was benign, wrote the correction into its own memory, then removed on its own initiative the monitoring queries that had become unnecessary. This is no longer a model you prompt turn after turn, it is an agent able to manage part of its own context over time.

Alignment and safety

Anthropic highlights the best alignment score obtained to date on its automated behavioural audit. The model keeps a deliberately defensive stance in cybersecurity, well behind the offensive capabilities of other specialised models, a deliberate choice to limit malicious uses without restricting legitimate ones.

Our own experience

We are not only talking about benchmarks published by Anthropic. We have already used Opus 5 internally to develop some features of the e-palett software. The verdict is clear: very few corrections needed on the code it produced. The model is particularly good, and this concretely confirms what we described above about its consistency from one run to the next.

This result also comes from the fact that a developer reviews and approves every change. We look at what happens when that review disappears in our article on vibe coding for SMEs.

At METEORA, we follow this kind of development closely, and we now experience it first-hand every day on our own projects. A model that is more consistent from one run to the next and able to check its own work is not a gimmick: it is directly relevant to the way we design and evolve our clients' custom projects. We will keep writing about it as our own experience grows.

To build these models into your own tools, see our AI solutions for businesses.

Claude Opus 5AnthropicArtificial intelligenceAutonomous agentsAI-assisted coding

The project behind this article

The project this article draws on.

e-palett - Web, MVP, SaaS project for EPALIA
LiveOngoing development
EPALIA
2020

e-palett

A complete rebuild, from scratch, of EPALIA's e-palett platform, moving from Joomla to a Symfony/JS architecture, handling 16M pallets a year with multi-customer ERP integrations.

1 person
6 months
Symfony PostgreSQLTypescript+1 more
Web•MVP•SaaS
120,000
Monthly declarations
-25%
Processing time
View project

How much would your project cost?

Project type, features, timeline: a costed estimate in two minutes, with no commitment.

Estimate my budget

Looking for a partner who combines human expertise and cutting-edge AI?

We combine our know-how with the best AI tools available to deliver your project faster, without sacrificing code quality.