Claude Opus 5 Closes the Gap With Fable 5, at a Fraction of the Cost
- David Borish
- Jul 25
- 5 min read
Anthropic shipped Claude Opus 5 on July 24, and the headline number is the price tag: $5 per million input tokens and $25 per million output tokens, unchanged from Opus 4.8, for a model the company says approaches the capability of its most advanced generally available model, Fable 5, on many tasks.
On CursorBench 3.2, run at maximum effort, Opus 5 landed within 0.5 percentage points of Fable 5's peak score while costing about half as much per task. On Frontier-Bench v0.1, it more than doubled Opus 4.8's performance at a lower cost per task than its predecessor. Those two data points frame the pitch for Opus 5: a model built to be used constantly, priced for daily workloads rather than reserved for occasional hard problems.
The model is now the default on Claude Max and the top-tier option on Claude Pro. It is the fourth Claude release in under two months, following Mythos 5, Fable 5, and Sonnet 5, all of which shipped in June. That pace, and the price point, respond directly to a complaint enterprise customers had raised about Fable 5's token consumption on extended tasks, which had pushed some users past their budgets. Anthropic is also not alone in leaning on cost efficiency as a selling point this month; OpenAI's GPT-5.6, released July 9, made similar claims about economical token use.
Where the Benchmark Gains Show Up
Anthropic's release notes credit Opus 5 with state-of-the-art results on Frontier-Bench and GDPval-AA, the company's coding and knowledge-work evaluations, while noting it still trails Mythos 5 specifically on cybersecurity tasks. The gains are not limited to coding. On ARC-AGI 3, a benchmark built around novel problem-solving rather than pattern recall, Opus 5 scored three times as high as the next-best model. On Zapier's AutomationBench, which measures whether a model can carry a business task through to completion rather than partway, Opus 5's pass rate ran about 1.5 times the next-best model at equivalent cost, and even at its lowest effort setting it passed more tasks than any competing model did at any setting.
OSWorld 2.0, a computer-use benchmark, produced one of the more striking cost comparisons in the release: Opus 5 matched Fable 5's best recorded result while using just over a third of the compute cost. Anthropic frames these results around its effort setting, a control that lets customers trade intelligence for speed and lower token spend, and the company says Opus 5 leads on cost-effectiveness at every effort level from low through max on several of these evaluations, including HLE and DeepSearchQA.
Life sciences performance also moved. Opus 5 beat Opus 4.8 on every internal life sciences evaluation Anthropic tracks, with the largest jump in organic chemistry, where the model scored 10.2 percentage points higher on inferring molecular structure from spectroscopy data. On protein-function prediction, tied to how sequence variations affect a protein's behavior, the gain was 7.7 percentage points.
What the Early Testing Reports Describe
Anthropic's release includes three specific examples meant to illustrate the model's persistence rather than its raw scores. In one Frontier-Bench task, Opus 5 was given a machine part drawing and asked to rebuild it as a 3D FreeCAD model, but was deliberately withheld direct image access. Rather than fail the task, it wrote its own computer vision pipeline to extract the geometry from raw pixel data and reconstructed the part, succeeding on repeated attempts where a competing model failed after five tries under the same constraints.
In a second example, given a real bug in a widely used open-source package manager, Opus 5 traced the issue to its root cause and fixed an edge case the community's own patch had missed, while a competing model addressed only the visible symptom and reported the bug closed. A third case, described by an engineer at a trading firm, involved building a market data feed for a new exchange in a single session, a task the engineer said prior models could not complete even with detailed plans supplied up front. With no live feed available to test against, Opus 5 built its own validation harness to confirm its code parsed the exchange's data correctly.
These are Anthropic's own selected examples rather than independently reproduced results, offered by the company as illustrations of behavior it chose to highlight rather than a representative sample of Opus 5's typical output.
Alignment Scores and the Cybersecurity Gap
Anthropic's pre-deployment behavioral audit rated Opus 5 as its most aligned model yet, with an overall misaligned-behavior score of 2.3, the lowest among its recent releases, and the company reports it as showing the lowest rates of deceptive behavior and the greatest resistance to misuse attempts among current models. Anthropic also describes it as the safest model to date in avoiding actions with hard-to-reverse consequences.
On cybersecurity specifically, the picture is mixed. On OSS-Fuzz, an internal evaluation of whether a model can find and then exploit software vulnerabilities without heavy human guidance, Opus 5 identifies vulnerabilities at a rate close to Mythos 5's, but its ability to develop working exploits from those vulnerabilities lags well behind. Anthropic attributes part of that gap to a deliberate choice: as with Opus 4.8, the company did not train Opus 5 on cyber tasks specifically, and the improvement it shows there is described as a byproduct of broader capability gains rather than targeted training.
The safeguards follow from that gap. Opus 5's cyber classifiers allow it to search source code for vulnerabilities but block binary-based vulnerability scanning, penetration testing, and exploit generation, and Anthropic says testing indicates the classifiers intervene roughly 85 percent less often than they do for Fable 5. Flagged requests in Claude.ai, Claude Code, and Claude Cowork fall back to Opus 4.8 by default, with that fallback also available as an API option.
Anthropic's Cyber Verification Program gives vetted enterprises and researchers access to a version of Opus 5 with fewer of these restrictions. On biology, the safeguards mirror those on Opus 4.8, and biology-related requests previously blocked on Fable 5 now route to Opus 5 rather than to Opus 4.8, though Anthropic says Mythos 5 remains the stronger model for long-running autonomous biological research, which the company identifies as the area of greatest concern for biology-related risk.
Availability and What Comes With It
Opus 5 is live today across Claude.ai, Claude Code, Claude Cowork, and the Claude API under the model ID claude-opus-5, at the same per-token pricing as Opus 4.8. A Fast mode runs the model at roughly 2.5 times default speed for twice the base price, matching how Opus 4.8's Fast mode was priced. Two related features are rolling out in beta alongside the model: mid-conversation tool switching on the Claude Platform, which lets developers change which tools are available to Claude without invalidating the prompt cache, and automatic fallbacks on the API, which route flagged requests on Opus 5 or Fable 5 to another available model instead of blocking them outright. As with prior Opus releases, Opus 5 carries no data retention requirement for general access.
For enterprises weighing where to run production workloads, the practical question raised by this release is less about peak capability and more about where Opus 5's cost-per-task advantage holds up under real usage patterns, particularly on the effort-setting tradeoffs Anthropic highlights across coding and automation benchmarks. Anthropic's own system card, referenced but not included in the release notes, will be the place to check for the fuller methodology behind these comparisons.
About the author: David Borish is an Enterprise AI Strategist and the author of the forthcoming book The Tony Hawk Paradox, which examines how capabilities that first appear in controlled or simulated environments tend to surface later in broader physical and economic systems. He writes long-form analysis on frontier AI, enterprise deployment economics, and open-source AI.
