What Anthropic Shipped on July 24

Anthropic released Claude Opus 5 on July 24, 2026, and the headline is blunt: it costs $5 per million input tokens and $25 per million output tokens — exactly half of what Fable 5 charges — and it beats Fable 5 on most of the benchmarks that actually matter for production workloads. That is not a typical product launch story. Usually cheaper means worse. Here it mostly doesn’t.

Opus 5 is Anthropic’s fourth model release in under two months, following Mythos 5, Fable 5, and Sonnet 5. It has a 1M-token context window, up to 128K output tokens (rising to 300K on the Batch API with a beta header), a knowledge cutoff of May 2026, and adaptive thinking switched on by default. It is now the default model on Claude Max and the top-tier option on Claude Pro.

The Benchmark Numbers

Anthropic published four benchmark results at launch, and each one deserves a sentence rather than a bullet:

On Frontier-Bench v0.1, Opus 5 more than doubles Opus 4.8’s score and surpasses every other model in the field — including Fable 5. On CursorBench 3.2, it lands within 0.5% of Fable 5’s peak score while running at half the cost per task. On ARC-AGI 3, it scores three times higher than the next-best model. On OSWorld 2.0, which tests real computer-use tasks, it surpasses Fable 5’s best result at roughly one-third of the cost.

Independent comparisons have added detail. On agentic terminal coding, Opus 5 reaches 43% task completion against Fable 5’s 33%. It also leads on GDPval-AA and BrowseComp, two evaluations focused on real-world knowledge work and web research. The pattern is consistent: Opus 5 is particularly strong in agentic workflows — the tasks where a model has to plan, use tools, and recover from errors over many steps.

Adaptive Thinking: The Hidden Cost Variable

The biggest practical change for developers is that adaptive thinking runs by default on Opus 5, where it did not on Opus 4.8. Thinking tokens are billed as output at $25 per million. This means a prompt that looks cheap on input can generate a surprisingly large thinking trace before the model replies.

Anthropic offers five effort levels — low, medium, high, xhigh, and max — set per request. Low skips most thinking overhead and approximates the old Opus 4.8 spend profile. Max lets the model think as long as it needs on hard multi-step problems. For most production API use, setting effort to medium or low will recover the predictable cost structure developers are used to, while preserving the option to push higher for specific hard requests.

The practical implication: budget based on effort setting, not just token count. An Opus 5 run at max effort on a complex coding task will cost more than an Opus 4.8 run on the same task, even though the base price is identical. For batch inference at scale — summarization, classification, routine extraction — low effort is almost certainly the right default.

Opus 5 vs Fable 5: When to Use Which

Fable 5 launched in June at $10/$50 per million tokens as Anthropic’s top-of-the-line model, with positioning around ambitious, long-running agent work and vision-heavy document processing. Opus 5 undercuts it on price and beats it on most public benchmarks. So is Fable 5 obsolete?

Not entirely. Fable 5 retains advantages in a specific cluster of use cases: extended multi-day agent runs where the model’s judgment about when to pause and check matters, vision-heavy workflows processing complex diagrams or tables from PDFs, and enterprise deployments where Fable 5’s higher safety rating is a procurement requirement. Anthropic describes Fable 5 as gated by additional safeguards, which some regulated industries need on paper regardless of performance differences.

For everything else — coding assistance, research workflows, computer use, standard agentic automation — Opus 5 at half the price is the rational default. The Claude Sonnet 5 pattern from June (near-Opus performance at lower cost) has now repeated one tier up: Opus 5 is near-Fable performance at half the cost.

Safety and Alignment

Anthropic describes Opus 5 as its most aligned model to date, with the lowest measured rates of deceptive behavior across its model family. The company applies Constitutional AI training and reinforcement learning from human feedback at each generation, and each new model generation since Claude 3 has shown measurable improvement on internal alignment evaluations.

What that means concretely: Opus 5 is less likely to hallucinate confidently, less likely to agree with false premises in a user prompt, and more likely to flag uncertainty rather than paper over it. These properties are particularly relevant in agentic contexts where the model makes many sequential decisions without a human in the loop — a single confident error at step three can cascade through the rest of a workflow. The Claude Opus 4.8 release from May introduced dynamic workflows and parallel agent orchestration; Opus 5 builds on that foundation with better reasoning fidelity.

What Developers Should Do Now

If you’re running Opus 4.8 in production, the migration path is straightforward: swap the model string to claude-opus-5, set effort to medium or low, and run your existing eval suite. You should see quality equal to or above what you had, at the same or lower price depending on thinking overhead.

If you’re currently using Fable 5 for general coding or agentic tasks, the calculus is simple: run Opus 5 against your benchmarks. If it matches on your specific workload — and public evidence suggests it will for most — switching cuts your API bill in half. Reserve Fable 5 for the narrow set of use cases where its specific safeguard profile is a genuine requirement, not just a habit.

The broader trend here is real: Anthropic is pushing frontier-class capability down the price curve faster than the top-tier price itself is moving. A developer paying Fable 5 rates for Opus-class work a month ago now has a cheaper option that benchmarks better. That compression matters for anyone building agents at scale.

Further Reading

• Introducing Claude Opus 5 — Anthropic — The official release post with Anthropic’s full benchmark breakdown and safety evaluation methodology.
• Claude Opus 5 vs Fable 5: Benchmarks and Pricing — DataCamp — Side-by-side comparison across seven evaluation categories with cost-per-task analysis.
• Claude Opus 5 Developer Guide — Essa Mamdani — Practical API migration notes covering effort levels, billing implications, and the Batch API 300K output change.