Anthropic drops Opus 5: cheaper, looser, and oddly better at checking its own work

Anthropic just released Opus 5 and Sonnet 5. The Opus tier gets a step change improvement for long running agents with better coding and professional work performance. Sonnet 5 delivers frontier performance at scale. Both models launched July 24, 2026.

What Opus 5 actually improves

Opus has always been the heavy lifter in Anthropic lineup. The model you reach for when the task needs sustained reasoning across hundreds of steps. Opus 5 pushes that further. The announcement calls out long running agents specifically. This matters because agents that run for hours not minutes hit different failure modes. Context drift. Accumulated errors. Tool call chains that go sideways three steps in.

The coding improvement is the concrete signal. Opus 4 was already strong at coding. Opus 5 moves the needle on professional work which means the model handles the messy parts of real software projects. Reading legacy codebases. Understanding implicit conventions. Navigating dependency graphs. Writing tests that actually catch regressions.

Checking its own work

The oddly better at checking its own work phrase in the announcement is doing heavy lifting. Self verification is the holy grail for agent reliability. A model that can catch its own mistakes before they cascade removes entire classes of failure. This is not the same as chain of thought reasoning. This is the model running a verification pass on its own output and correcting errors it finds.

If this works reliably it changes the economics of agent deployment. You spend less on oversight. You trust the agent with higher stakes tasks. You run fewer human in the loop checkpoints.

Sonnet 5 at scale

Sonnet 5 delivers frontier performance across coding agents and professional work at scale. The at scale part matters. Sonnet has always been the workhorse tier. Fast enough for production. Smart enough for real tasks. Cheap enough to run everywhere. Sonnet 5 keeps that positioning while pushing the capability ceiling.

For teams building products on Anthropic API this is the model you default to. Opus 5 for the hardest problems. Sonnet 5 for everything else. Haiku for the high volume simple stuff. The tiered approach means you match model to task value.

The competitive context

OpenAI GPT-5 family is in the wild. Google Gemini 2.5 Pro dropped recently. The frontier is crowded. Anthropic differentiation has been safety and reliability. The Opus 5 announcement leans into that. Long running agents that do not go off the rails. Self verification that reduces human oversight. Coding that handles professional complexity.

The pricing signal matters too. Cheaper and looser in the title suggests Opus 5 costs less to run than Opus 4. Looser likely means fewer refusals on legitimate tasks. The refusal problem has plagued all frontier models. A model that says no to valid requests because it cannot distinguish them from dangerous ones is a product problem not a safety feature.

What this means for agent architecture

If Opus 5 genuinely handles long running agents better the architecture pattern shifts. You can build agents that plan over days not hours. You can delegate entire workflows instead of single tasks. The verification capability means the agent can run a test suite, see failures, fix them, and verify the fix without human intervention.

This is the difference between an assistant and an autonomous system. Assistants wait for direction. Autonomous systems pursue objectives. Opus 5 appears to be targeting the autonomous system tier.

The open questions

Benchmarks are not production. The announcement cites coding and professional work improvements but does not publish the eval suite. Real world agent performance depends on tool integration quality, context management, error recovery, and cost predictability. None of those show up in a model card.

The self verification claim needs stress testing. Does it catch logic errors or just syntax errors? Does it work across multiple tool calls? Does it increase latency enough to matter for interactive use? These questions determine whether Opus 5 is a step change or an incremental bump.

Pricing details are not public yet. Cheaper than Opus 4 is the claim. The actual API rates will determine adoption curves for teams budgeting agent workloads.

Source: Anthropic newsroom, The Verge