Anthropic released Claude Opus 5 on July 24, 2026, and the headline number is the one creators care about most: it costs the same $5 per million input tokens and $25 per million output tokens as Opus 4.8, yet Anthropic says it more than doubles Opus 4.8's score on its internal Frontier-Bench v0.1 evaluation at a lower cost per task. In plainer terms, the same budget now buys roughly twice the useful work. Opus 5 is available immediately in the Anthropic API as claude-opus-5, inside Claude.ai, in Claude Code, in Claude Cowork, and on cloud platforms including Amazon Bedrock.
This is the release that resets the price-to-capability line for anyone building agents, coding pipelines, or research workflows on Claude. Below is a full breakdown of what shipped, how it compares to Opus 4.8 and the rival frontier model Fable 5, and exactly how to switch your projects over today.
What Anthropic actually shipped
Opus 5 is a frontier reasoning model positioned as Anthropic's most capable Opus release to date. The two changes that matter for day-to-day building are reliability and economics. On reliability, Anthropic and early testers highlight that the model is markedly stronger at verifying its own work and iterating until a task genuinely succeeds, rather than stopping at a plausible-looking first answer. TechCrunch reported that in one demo the model wrote its own computer vision pipeline from an incomplete prompt, filling gaps by checking and re-running its code.
On economics, Opus 5 holds Opus-tier pricing while pushing into territory that previously required paying for the pricier Fable 5. Anthropic states Opus 5 lands within 0.5% of Fable 5's peak score on CursorBench 3.2 at half the cost per task, and surpasses Fable 5's best OSWorld 2.0 result at just over a third of the cost. The TechCrunch launch coverage framed the takeaway bluntly: Opus 5 is both cheaper and less restrictive than Fable, which makes it the default choice in most use cases.
Opus 5 vs Opus 4.8 vs Fable 5
Here is how the three models line up on the dimensions Anthropic disclosed at launch. Pricing for Opus 5 and Opus 4.8 is identical; the relative benchmark positions come directly from Anthropic's announcement.
| Dimension | Claude Opus 5 | Claude Opus 4.8 | Fable 5 (rival) |
|---|---|---|---|
| Input price (per 1M tokens) | $5 | $5 | Higher (Anthropic: Opus 5 is cheaper) |
| Output price (per 1M tokens) | $25 | $25 | Higher |
| Fast mode | 2x price, 2.5x faster | Available | Not disclosed |
| Frontier-Bench v0.1 | More than 2x Opus 4.8, lower cost/task | Baseline | Below Opus 5 |
| CursorBench 3.2 (coding) | Within 0.5% of Fable 5 at half the cost | Lower | Peak score |
| OSWorld 2.0 (computer use) | Beats Fable 5 at ~1/3 the cost | Lower | Prior best |
| ARC-AGI 3 (reasoning) | 3x the next-best model | Lower | Next-best tier |
| Safety classifier hits | 85% fewer than Fable 5 | Not disclosed | Baseline |
The pattern is consistent across every benchmark Anthropic published: comparable or better output than the frontier competitor, delivered at a materially lower cost per completed task. For teams running high-volume agent loops, cost per task is the metric that actually shows up on the invoice, and it is where Opus 5 makes its case. If you want to understand why that number swings so much between models, our breakdown of why AI coding sessions cost what they do unpacks the mechanics.
The benchmarks, decoded
Benchmark names are easy to gloss over, so here is what each one is measuring and why it matters for creative and building work.
Frontier-Bench v0.1 is Anthropic's broad capability index. Doubling Opus 4.8 here is the single strongest signal that this is a genuine generational step rather than a point update. CursorBench 3.2 measures real coding agent performance inside an editor loop, so its near-parity-with-Fable-at-half-cost result is the one coding teams should weigh most heavily. OSWorld 2.0 tests computer use, meaning the model driving a real desktop to click, type, and navigate. ARC-AGI 3 probes abstract reasoning and generalization, and a 3x lead over the next-best model is the most eye-catching gap in the whole release. Zapier AutomationBench, where Opus 5 posts roughly 1.5x the pass rate of the next-best model at the same cost per task, is the closest proxy for real multi-step automation across connected apps.
Anthropic also called out a domain-specific gain: scientific reasoning on organic chemistry problems improved by 10.2 percentage points, a reminder that the reliability improvements extend into technical research, not just software.
What this enables for creators and builders
The practical unlock is running more ambitious agent workflows without the bill scaling out of reach. Because Opus 5 verifies its own output more aggressively, the failure mode where an agent confidently produces broken code or a half-finished asset and then stops becomes less common. That reliability is what makes longer autonomous chains, the kind that write, test, and fix in a loop, viable to leave running.
Three concrete workflows get noticeably better today. First, coding agents: near-Fable coding quality at half the cost per task means you can point Claude Code at larger refactors and multi-file features and trust the verification pass. Second, computer-use automation: the OSWorld and Zapier gains mean agents that operate real apps and stitch tools together fail less and cost less. Third, research and analysis: the reasoning and scientific gains make Opus 5 a stronger fit for deep, multi-step investigation where a wrong intermediate step derails the whole answer.
How to start using Opus 5 today
Switching is intentionally low-friction because the pricing and API surface match Opus 4.8. Here is the fastest path for each entry point.
- Anthropic API: change your model string to
claude-opus-5. Pricing is unchanged from Opus 4.8, so no budget re-modeling is needed. Confirm the current identifier on the Anthropic models documentation. - Claude Code: select Opus 5 as your model in the CLI or IDE extension and point it at a real task. The verification improvements are most visible on multi-file changes. See the Claude Code product page for setup.
- Claude.ai and Claude Cowork: Opus 5 appears in the model picker for eligible plans, no configuration required.
- Amazon Bedrock: use model ID
global.anthropic.claude-opus-5. Per the AWS launch post, it is live in US East (N. Virginia), Asia Pacific (Melbourne), Europe (Ireland), and Europe (Stockholm), runs with zero data retention by default, and supports mid-conversationtool_additionandtool_removalblocks through the Bedrock Converse and Messages APIs.
If you already run Opus 4.8 in production, the migration is effectively a one-line change plus a regression pass on your own evals. Because output quality goes up while price holds, most teams will see the switch pay for itself immediately.
Pricing and the cost-per-task story
The list price, $5 per million input tokens and $25 per million output tokens, is unchanged from Opus 4.8, and Fast mode remains available at double the base price for 2.5x faster responses. What changed is the denominator. When Anthropic quotes results at "half the cost per task" or "a third of the cost," it is folding in the fact that a more reliable model needs fewer retries and shorter correction loops to finish a job. Fewer wasted turns means the effective cost of a completed task drops even when the per-token rate is flat. Full rate details live on the Anthropic pricing page.
One safety-side change is worth flagging for builders: Anthropic reports 85% fewer safety classifier engagements than Fable 5 and introduced an Automatic Fallbacks feature that routes a request to a less powerful model when a safety trigger fires, rather than refusing outright. In practice that should mean fewer hard blocks interrupting a legitimate automation.
Frequently asked questions
When was Claude Opus 5 released?
July 24, 2026. Anthropic made it available the same day across the API, Claude.ai, Claude Code, Claude Cowork, and cloud platforms including Amazon Bedrock.
How much does Claude Opus 5 cost?
$5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. Fast mode costs double the base price and runs about 2.5x faster.
Is Opus 5 better than Fable 5?
Anthropic's benchmarks show Opus 5 within 0.5% of Fable 5 on CursorBench 3.2 at half the cost, and ahead of Fable 5 on OSWorld 2.0 at roughly a third of the cost. Anthropic positions it as cheaper and less restrictive than Fable in most use cases.
What is the model ID for the API and Bedrock?
On the Anthropic API the identifier is claude-opus-5. On Amazon Bedrock it is global.anthropic.claude-opus-5.
Do I need to change my code or pricing to upgrade from Opus 4.8?
No pricing change is required because rates are identical. The upgrade is a one-line model string swap, followed by a regression pass on your own evaluations to confirm the quality gain on your specific workloads.
What is the biggest practical improvement over Opus 4.8?
Reliability. Opus 5 verifies its own output and iterates until a task actually succeeds, which reduces failed or half-finished results in long autonomous agent loops and lowers the effective cost per completed task.
