Fable-5-Grade Intelligence at Last Gen's Price: How Claude Opus 5 Rebuilt "the Opus for Long-Running Agents"
Anthropic released Claude Opus 5 on July 24. At the unchanged $5/$25 pricing, it delivers performance approaching the publicly top-tier Fable 5 and replaces the default Opus in Claude Code. Here's a rundown of the shifts in pricing, benchmarks, and agent operations—along with the caveats to keep in
On July 24, 2026, Anthropic released a new model in the Opus tier: Claude Opus 5. The company frames it as "intelligence approaching that of the top-tier Fable 5, at half the price"—and it becomes the default model on Claude Max and the highest-end option you can select on Claude Pro. It's also built into Claude Code v2.1.219 as the default Opus, meaning that for developers who keep coding agents running for long stretches, the "brain they run every day" has quietly been swapped out. Here we take stock of the pricing, the capabilities, and the operational changes from the perspective of a working development team.
Same Price Tag, Aiming for "Half the Cost of Fable"
The standard pricing for Opus 5 is the same as the Opus 4.8 it replaces: $5 per million input tokens and $25 per million output tokens. The pitch is that they raised performance without raising the price. The key point is that at this price band it comes close to the publicly top-tier Fable 5 ($10/$50), and Anthropic describes it as "near-frontier intelligence at half the cost." On top of that, there's a Fast mode that runs roughly 2.5x faster, priced at double the standard rate (equivalent to $10/$50).
| Model | Input / 1M tokens | Output / 1M tokens | Positioning |
|---|---|---|---|
| Claude Opus 5 (Standard) | $5 | $25 | New default Opus |
| Claude Opus 5 (Fast) | $10 | $50 | ~2.5x faster, double the price |
| Claude Opus 4.8 | $5 | $25 | Model being replaced |
| Claude Fable 5 | $10 | $50 | Publicly top-tier |
Availability spans Claude.ai, Claude Code, Claude Cowork, and the Claude API. The model name on the API is claude-opus-5, and Claude Code's claude-api skill has also updated its default to Opus 5, with migration guidance provided for moving from Opus 4.8. Note that Fast mode is limited to Opus 5 and 4.8; Opus 4.7 has been dropped from Fast mode support.
Where It Stands, as Seen in the Benchmarks
By Anthropic's published figures, Opus 5 substantially outperforms the previous-generation Opus 4.8 and, on several metrics, surpasses or nearly matches Fable 5. Pulling out the parts most relevant to developers, it looks like this.
| Benchmark | Opus 5 | Fable 5 | Reference |
|---|---|---|---|
| Frontier-Bench v0.1 | 43.3% | 33.7% | GPT-5.6 Sol 34.4% |
| OSWorld 2.0 (PC control) | 70.6% | 66.1% | — |
| ARC-AGI-3 (novel tasks) | 30.2% | — | Opus 4.8 at 1.5% |
| AutomationBench | 26.0% | 18.1% | — |
On coding, Anthropic says that on CursorBench 3.2's maximum-effort setting, Opus 5 closes to within 0.5% of Fable 5's top score while keeping costs at half. That said, it's not a clean sweep. On the more software-engineering-oriented DeepSWE v1.1, Opus 5 scores 68.8% against GPT-5.6 Sol's 72.7%, and on the hard-problem Humanity's Last Exam it falls just short of Fable 5; in cybersecurity and some specialized domains, Mythos 5 is said to still hold the lead. Rather than "the best at everything," the more accurate read is that it's a model with strong price-performance across the bandwidth of everyday coding and knowledge work.
Unshowy but Heavyweight Changes That Matter for Running Agents
Beyond the numbers, this release also brings operational improvements that matter when you run long-lived agents. Three of them land directly in the development trenches.
- Swapping tools out mid-conversation no longer discards the cached prompt. Even for agents that dynamically change their tool configuration, it becomes easier to keep the cache of accumulated context.
- A request that trips the safety filter is routed to a fallback model instead of failing. This makes it easier to avoid the situation where an unattended run "gets blocked and grinds to a halt."
- Claude Code's default Opus is now Opus 5. Any workflow that doesn't explicitly pin a model will run on this model from here on out.
The design of "Opus-grade intelligence at the same unit price as Opus 4.8" pairs well with usage patterns like running subagents in parallel or looping through verification and self-correction many times over—but the more you run, the more your token consumption piles up. Precisely because the default model has grown stronger, keeping an eye on Fast mode's double price, and on the output tokens that balloon over long sessions, becomes more of a design concern than ever.
Caveats Worth Keeping in Mind Behind "Close and Half-Price"
It's a promising announcement, but there are points to hold onto rather than take at face value. First, the pace of releases. Opus 5 follows Fable 5 (June) and Sonnet 5 (late June) in short order, and a situation where models keep replacing the default one after another forces teams to redo their verification and cost assumptions. Second, how to read the benchmarks. Some reporting notes that the Frontier-Bench score includes fallback runs handled by Opus 4.8 when requests were blocked by the safety filter, and that the proportion of these isn't disclosed. It's safest to treat the published figures with the understanding that they aren't necessarily "pure new-model performance under identical conditions." When adopting it, the basics don't change: try it on a small scale with your own tasks and evaluate it alongside actual measured costs.
Even so, this update—raising the Opus tier's real-world capability without raising the price, and swapping out the default for long-running agents—is reason enough to reconsider "which model you run your daily development on." A good place to start is checking, in your own workflow, the differences in speed, quality, and cost that come from the default model changing to Opus 5.
References: Anthropic, "Introducing Claude Opus 5" / CNBC / The Next Web / tech-ish (benchmark details and caveats) / Claude Code changelog v2.1.219
