July 24, 2026 — Anthropic on Thursday released Opus 5, the newest model in its Claude family of artificial-intelligence systems, framing the launch as an efficiency upgrade that delivers near-flagship performance at a fraction of the cost of its most powerful models.
According to the company, Opus 5 is for regular office and computer programming, and “nears the capabilities of its more powerful sibling Fable 5 at half the cost. The model’s release comes only two months after the arrival of the new version of Opus, called Opus 4.8, on May 28, highlighting the acceleration of the release cycles between top AI providers.
Opus 5’s prices remain unchanged from Opus 4.8 and are lower than Fable 5’s $50/million output token fee, $0/million input tokens. This pricing is also a key focus of the Anthropic pitch: the same performance without the absolute leading edge of capabilities, but with a focus on cost at scale.
On performance, Anthropic said Opus 5 outperforms Fable 5 on several benchmarks and exceeds both Opus 4.8 and OpenAI’s GPT-5.6-Sol across most task categories, with coding singled out as a particular strength. The company also highlighted a shift in how the model works, describing Opus 5 as “much stronger at verifying its work and iterating carefully until it succeeds”, a self-correction ability aimed at reducing the errors that can derail automated software tasks.
Efficiency, however, is the headline theme. Anthropic said Opus 5’s safety classifiers engage roughly 85% less frequently than those on Fable 5, cutting down on the interruptions that developers often encounter when automated guardrails halt otherwise legitimate work. To smooth those remaining cases, the company introduced a beta feature called Automatic Fallbacks, which reroutes a request that trips a safety classifier to a less powerful model and returns a functional response rather than an error.
The launch also reflects Anthropic’s continued emphasis on security positioning. The company said Opus 5 is less capable of exploiting cyber vulnerabilities than Fable 5 and less susceptible to being manipulated into misuse than its other current models. Consistent with that stance, binary vulnerability scanning is prohibited on the model, while source-code scanning for defensive security purposes remains permitted. In a nod to enterprise privacy concerns, Anthropic said Opus 5 is exempt from the 30-day data-retention policy applied to its Fable and Mythos models.
Dianne Penn, a product leader at Anthropic, cast the release as part of a consistent strategy. Penn advised customers to reach for Opus 5 when value matters and to reserve Fable 5 for “days-long, very autonomous projects” that demand the company’s most capable system.
The release rounds out most of Anthropic’s move to its “5” generation. Mythos 5, Fable 5 and Sonnet 5 all launched in June 2026, leaving Haiku as the only member of the lineup still awaiting its upgrade. The rollout has not been entirely smooth: Fable 5, released in June, was temporarily pulled from availability amid U.S. security concerns over its potential military-intelligence applications, a reminder of the regulatory scrutiny now trailing the most powerful commercial models.
Competition is intensifying on multiple fronts. Beyond OpenAI and Google, Anthropic faces growing pressure from lower-cost, open-weight rivals, including China-based Moonshot’s Kimi K3. Asked about that threat, Penn was measured, saying “it remains to be seen” how open-weight models perform on complex, real-world tasks, an implicit argument that raw benchmark scores don’t always translate into dependable performance on the messy, multi-step work that businesses actually run.
That argument sits at the heart of Anthropic’s Opus 5 strategy. Rather than chasing a single, ever-larger flagship, the company is increasingly segmenting its lineup by use case and price, encouraging customers to match the model to the job: a cheaper, efficient workhorse for daily coding and office tasks, and a premium system for long-running autonomous projects. For enterprise buyers weighing per-token costs against capability, the release of a model that claims near-Fable 5 quality at half the price is likely to sharpen that calculus.
For developers, the practical draw may be less about benchmark leaderboards and more about friction. Fewer safety-classifier interruptions, automatic fallbacks in place of hard errors, and a model tuned to check and correct its own output all point to the same goal: keeping automated workflows moving. Whether Opus 5’s real-world reliability lives up to Anthropic’s benchmark claims will become clearer as customers put it to work in the weeks ahead, but the launch signals that, for now, the AI race is being fought as much on cost and dependability as on raw intelligence.