Claude Opus 5 Pricing and Benchmarks: What Actually Changed on July 24

The full breakdown of price, benchmarks, and the four evaluations Opus 5 still loses — tested against real workflows, not just Anthropic’s own chart.

By Oyekale Olawale · Updated July 28, 2026

Quick Answer

Claude Opus 5 launched July 24, 2026, at $5 per million input tokens and $25 per million output — the same price as Opus 4.8, and exactly half of Fable 5’s $10/$50. It’s now the default model on Claude Max and the top-tier option on Claude Pro, and it more than doubles Opus 4.8’s score on Frontier-Bench (43.3% vs 21.1%). It doesn’t win everything — GPT-5.6 Sol still leads on DeepSWE, and Fable 5 still edges it on the Legal Agent Benchmark — but for coding and knowledge work, there’s no longer a cost argument for staying on the older model.

• Pricing didn’t move from Opus 4.8 — $5 in / $25 out per million tokens — despite the generational jump in capability.
• The claude-opus-5 API model ID carries no mandatory data-retention requirement, unlike Fable 5’s 30-day retention policy.
• Anthropic’s own table shows four evaluations where Opus 5 comes second: DeepSWE, FrontierCode Main, HLE without tools, and the held-out Legal Agent Benchmark.
• Fast mode runs at roughly 2.5x the default speed for double the base price — same structure as Opus 4.8’s Fast mode.

What Actually Shipped on July 24

I’ve been running Opus 5 alongside Opus 4.8 in the same Claude Code project since launch week, and the first thing that struck me wasn’t a benchmark — it was that nothing about the pricing page changed. Anthropic kept the API model ID pattern consistent (claude-opus-5, same as every prior release), kept the $5/$25 rate card, and just… shipped a materially better model into the same price bracket Opus 4.8 occupied since May 28.

Two access changes matter more than the headline price. Opus 5 is now the default model on Claude Max, and it’s the strongest option selectable on Claude Pro — so if you’re on either plan and haven’t touched a setting, you’re already talking to it. And unlike Fable 5, general access to Opus 5 carries no mandatory data-retention requirement, which had been a real blocker for teams under zero-data-retention agreements since Fable 5 shipped back in June. If you’re weighing Claude against other assistants for coding specifically, our breakdown of Claude for coding tasks is worth reading alongside this, since the retention policy affects which model you can actually deploy in regulated environments.

Fast mode carries over from the previous generation too — roughly 2.5x the response speed at 2x the base token price, available through the Claude Platform and via usage credits inside Claude Code. If you’re burning through weekly limits on Claude Code already, that’s the lever to pull before you touch the effort setting.

Opus 5 vs Fable 5 vs Opus 4.8 vs GPT-5.6 Sol

Here’s the pricing and headline benchmark comparison, pulled straight from Anthropic’s launch table and cross-checked against the system card. I’ve kept it to the numbers that actually change a routing decision.

All figures vendor-reported by Anthropic from its July 24, 2026 launch materials and system card. Frontier-Bench figures come from an internal harness on the mini-SWE-agent framework, with Opus 4.8 serving as the fallback model on safety-classifier refusals — worth knowing before you cite the 43.3% number as a clean single-model score.

Frontier-Bench v0.1: Agentic Terminal Coding

Where Opus 5 Actually Loses

Any launch table with zero losses is marketing, not data. Anthropic published four rows where Opus 5 comes second, and I think they’re more useful than the wins for deciding whether to migrate.

The pattern is consistent: general agentic capability goes to Opus 5, narrow specialist domains still lean toward the models built or trained specifically for them. If your workload sits in legal review or clinical documentation, keep testing Fable 5 or Mythos 5 alongside Opus 5 rather than assuming the newer model wins by default.

The Real Upgrade Isn’t the Benchmark — It’s Self-Verification

The number that got the most attention on launch day was Frontier-Bench. The thing that actually changed my own workflow was smaller and less flashy: Opus 5 checks its own work before handing it back, without being told to.

I fed it a task with a deliberately broken test harness — the kind of thing that used to make Opus 4.8 either loop forever or declare victory on a false pass. Opus 5 flagged the harness itself as suspect, rewrote a minimal reproduction case to confirm the bug was real, and only then proposed a fix. That’s the behavior Anthropic describes in its own examples too — one involving a model that built its own computer-vision pipeline to solve a task it had no direct way to view, and another where it caught a root-cause bug in a package manager that a competing model’s fix missed entirely.

Anthropic also shipped two platform-level features alongside the model that matter more to builders than end users: mid-conversation tool swapping without invalidating the prompt cache, and automatic API fallbacks when a request trips a safety classifier. If you’ve built retry logic around classifier refusals on Fable 5, this removes a chunk of that error-handling code. It’s worth reading alongside our guide on practical Claude Code use cases if you’re building agentic workflows that need to survive a model swap mid-task.

How I Tested Claude Opus 5

I ran Opus 5 through three environments over the first four days of access: the Claude Platform console for raw API calls, Claude Code for a mid-sized refactor of an internal tooling repo, and the standard claude.ai chat interface for research and writing tasks. I compared every result against the same task run on Opus 4.8 at matching effort settings, since a fair comparison needs identical prompts and identical context windows, not just a vibe check.

Two things stood out that Anthropic’s own launch materials don’t emphasize. First, the effort dial genuinely changes behavior, not just verbosity — at the lowest setting, Opus 5 skipped a verification step it performed automatically at medium and high effort, which meant it shipped one subtly wrong regex on a low-effort run that it caught immediately once I bumped the setting up. Second, the mid-conversation tool-switching feature has a rough edge: on one run, switching from a file-search tool to a shell-execution tool mid-task caused the model to briefly reference a file path from the earlier tool context that no longer existed, requiring a manual nudge to re-orient. It self-corrected within one turn, but it’s a UX wrinkle worth knowing about if you’re chaining tools automatically.

On cost, my own numbers roughly matched Anthropic’s directional claims — a legal-document summarization task that took Opus 4.8 four back-and-forth turns to get right took Opus 5 two, at a comparable per-turn token cost, which meant a real drop in total spend for that task type.

Alignment and Cyber Safeguards

Anthropic reports Opus 5 scores 2.3 on its automated behavioral audit for overall misaligned behavior — the lowest of its recent model releases — with the company describing it as the least susceptible to being tricked into misuse among current Claude models. On dangerous-capability evaluations, Anthropic says Opus 5 doesn’t advance the frontier and remains behind Mythos 5 on both biology research and offensive cybersecurity tasks.

The cyber classifiers are notably looser than Fable 5’s — Anthropic expects them to intervene roughly 85% less often. Source-code vulnerability search is permitted; binary-based scanning, penetration testing, and exploit generation remain blocked, with flagged requests falling back to Opus 4.8 by default in Claude.ai, Claude Code, and Claude Cowork. On the biology side, the safeguard suite is closer to Opus 4.8’s than to Fable 5’s stricter one, which in practice means Opus 5 is now Anthropic’s most capable generally available model for legitimate scientific research use cases.

Should You Switch? A Routing Checklist

If you’re currently on Opus 4.8 for general coding or agentic work, there’s no real argument for staying — same price, more than double the Frontier-Bench score, and a large jump on novel problem-solving. Validate against your own eval set, then flip the default.

If you’re paying for Fable 5 specifically for engineering or operations work, test the downgrade seriously. Opus 5 wins or ties most coding benchmarks at half the list price and drops the mandatory retention requirement — though DeepSWE remains one case where the more expensive model still pulls ahead. If your workload is legal review, clinical documentation, or another narrow professional domain, don’t move yet; those are exactly the rows where Opus 5 still comes second, and human review stays non-negotiable regardless of which model you pick. For a broader look at how Claude stacks up against other assistants on marketing and research tasks specifically, our comparison of ChatGPT vs Google Gemini for marketers and our roundup of ChatGPT alternatives worth testing are useful companion reads, since Opus 5’s pricing shift changes where it lands in those comparisons too.

FAQ

How much does Claude Opus 5 cost?

$5 per million input tokens and $25 per million output tokens — identical to Opus 4.8, and half of Fable 5’s $10/$50 rate.

Is Claude Opus 5 better than Claude Fable 5?

It wins or ties most benchmarks Anthropic published, including coding and knowledge work, at half the price and with no mandatory data retention. Fable 5 still leads on the held-out Legal Agent Benchmark and DeepSWE.

Does Claude Opus 5 have data retention requirements?

No. General access to Opus 5 carries no mandatory data-retention requirement, consistent with prior Opus releases and unlike Fable 5’s 30-day policy.

What is the Claude Opus 5 API model ID?

claude-opus-5, available now through the Claude API.

Where does Opus 5 fall behind other models?

DeepSWE v1.1 (GPT-5.6 Sol leads), the held-out Legal Agent Benchmark and HLE without tools (Fable 5 leads), and HealthBench Professional (Mythos 5 leads).

Bottom Line

Claude Opus 5 is a rare flagship release where the headline is a price rather than a capability ceiling. At $5/$25 it costs what Opus 4.8 cost while more than doubling its agentic coding score and dropping the retention requirement that made Fable 5 a hard sell for regulated teams. It isn’t a clean sweep — four evaluations still go elsewhere, clustered in exactly the specialist domains where a wrong answer costs the most — but for general coding, computer use, and business workflow automation, the math is straightforward: run your own evals, then move the default.

Who I Am and How I Tested These Tools

I’m Oyekale Olawale, and I write hands-on AI tool reviews for websites2know.com. My testing process for a model release like this one always follows the same shape: run the same prompt set across the new model and its predecessor at matched effort levels, log every turn count and token spend, and specifically hunt for the failure modes vendors don’t put in their launch charts — broken tool-handoffs, over-eager task completion, and silent low-effort errors. I don’t rely on benchmark tables alone; if a model can’t survive a messy real repo or an ambiguous legal-style prompt, the benchmark number doesn’t matter much to the person paying for it.