Anthropic released Claude Opus 5 on July 24, 2026. It costs exactly what Opus 4.8 cost. That’s the whole pitch in one sentence: same price, a lot more model.
If you’re paying for Claude right now, here’s the direct answer. Opus 5 is Anthropic’s new mid-frontier workhorse. It gets close to Claude Fable 5 on coding and knowledge work, stays at $5 per million input tokens and $25 per million output tokens, and becomes the default model on Claude Max and the top model on Claude Pro starting today. For almost everyone not doing multi-day autonomous research, it’s the model you should be using.
We spent the past two days going through Anthropic’s own benchmark data, the system card, and early reactions from companies like Cursor, Devin, and Box. Here’s what actually matters.
What Is Claude Opus 5?
Claude Opus 5 is Anthropic’s latest release in the Opus line, sitting one tier below Claude Fable 5 and Claude Mythos 5. Anthropic built it to close the gap with its top-end models on coding and professional work, while keeping Opus-level pricing.
This is the fourth Claude model release in about two months, following Mythos 5, Fable 5, and Sonnet 5 in June. <cite index=”5-1″>Anthropic said Opus 5 was easier to use than its predecessors, requiring less back and forth, and that it verifies its own work and recovers from errors without intervention.</cite>
The launch also lands in the middle of a rocky stretch for Fable 5. That model got hit with temporary U.S. export controls in mid-June after security researchers found ways around its safeguards, went dark for about two weeks, and came back with tighter security on June 30. Opus 5 is Anthropic’s answer to a simple problem it created for itself: Fable 5 is powerful but expensive and briefly unavailable, and a lot of paying customers just wanted something dependable in between.
What Changed From Opus 4.8
Anthropic isn’t shy about the numbers here, and they’re bigger than a typical point release.
<cite index=”11-1″>On Frontier-Bench v0.1, Opus 5 surpasses all other models and more than doubles Opus 4.8’s performance at a lower cost per task. On CursorBench 3.2, at max effort, the model performs within 0.5% of Fable 5’s peak score, but at half the cost per task.</cite> That second number is the one worth sitting with. Anthropic’s own flagship model barely beats Opus 5 on a real coding benchmark, and Opus 5 does it for half the money.
The gains aren’t limited to coding either. <cite index=”11-1″>On ARC-AGI 3, a benchmark built around solving novel problems, Opus 5’s score is three times as high as the next-best model. On Zapier’s AutomationBench, which tests whether a model can finish real business tasks start to finish, Opus 5’s pass rate is roughly 1.5 times the next-best model at the same cost, and even its lowest effort setting beats every other model.</cite>
On OSWorld 2.0, a computer-use benchmark, <cite index=”11-1″>Opus 5 outperforms every other model at any given cost, and beats Fable 5’s best result at just over a third of the cost.</cite>
Science and research work improved too. <cite index=”11-1″>Opus 5 beats Opus 4.8 across every life sciences evaluation Anthropic tracks, including structural biology, organic chemistry, and bioinformatics, with the biggest jump on organic chemistry tasks like reading molecular structures from spectroscopy data.</cite>
We’re not just repeating Anthropic’s marketing here. Early access partners echoed the same story with specifics. Devin’s team said Opus 5 gets close to Fable-level results at half the cost and specifically called out debugging and root-cause work. Cursor’s co-founder Sualeh Asif said it lands just under Fable 5 on CursorBench “with many of the same behaviors.” Box reported an 8% overall accuracy jump over Opus 4.8, with data analysis up 11% and due diligence work up 17%.
Claude Opus 5 vs Fable 5 vs Sonnet 5 vs Opus 4.8
This is the table most people actually need. Here’s how the current Claude lineup shakes out after the Opus 5 launch.
| Model | Best for | Price (input / output per million tokens) | Where it wins | Where it loses |
|---|---|---|---|---|
| Claude Opus 5 | Daily coding, agentic work, enterprise tasks | $5 / $25 | Cost-to-performance, verification and self-correction, safest Claude yet | Trails Mythos 5 on cybersecurity and biology research |
| Claude Fable 5 | Multi-day autonomous work, capability ceiling | $10 / $50 | Still Anthropic’s smartest general model, narrow CursorBench edge | Twice the cost, history of token burn complaints |
| Claude Sonnet 5 | High-volume, cost-sensitive workloads | $3 / $15 (intro $2 / $10 through Aug 31) | Cheapest way to stay in the Claude ecosystem | Not built for the hardest agentic or research tasks |
| Claude Opus 4.8 | Teams pinned to a validated eval suite | $5 / $25 | Known, tested behavior for existing pipelines | Opus 5 beats it on nearly every published benchmark at the same price |
One important detail buried in the release notes: <cite index=”9-1″>thinking is on by default for Opus 5, and because max_tokens caps thinking plus response text together, workloads that ran 4.8 without thinking need their limits revisited.</cite> If you’re calling the API directly, check that before you flip a production pipeline over.
Where Opus 5 Still Falls Short
Anthropic is upfront about two gaps, and we’d rather flag them than pretend Opus 5 is perfect.
<cite index=”11-1″>Opus 5 remains behind Mythos 5 on cybersecurity tasks. It comes close to Mythos 5 at finding vulnerabilities, but is considerably less successful at developing exploits from them.</cite> That’s a safety choice, not a bug. Anthropic deliberately avoided training Opus 5 on offensive cyber tasks, the same approach it took with Opus 4.8.
<cite index=”11-1″>On biology, Opus 5 uses safeguards similar to Opus 4.8, making it Anthropic’s most capable generally available model for scientific research, but it still shows real limitations on long-running, autonomous research tasks — the exact area where Mythos 5 remains stronger.</cite> If your work involves days-long unsupervised research agents, Fable 5 or Mythos 5 access is still the better call.
Also worth knowing: Opus 5 is the newest model, not the smartest one on paper. That title stays with Fable 5. Anthropic’s own positioning is that most real work happens in a “middle band” of difficulty, where near-frontier intelligence delivered cheaply beats paying frontier prices for headroom you rarely use.
Is Claude Opus 5 Safer?
Yes, and Anthropic backs it with data instead of a slogan. <cite index=”11-1″>On Anthropic’s automated behavioral audit, Opus 5 scored as the company’s most aligned model to date, adhering to Claude’s Constitution better than Opus 4.8, Sonnet 5, or Fable 5, with the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse.</cite>
That matters practically, not just ethically. <cite index=”11-1″>Anthropic expects Opus 5’s cyber safety classifiers to intervene about 85% less often than Fable 5’s do, meaning fewer false-positive refusals for legitimate security research work.</cite> If you’ve been fighting with Claude blocking benign code-review or pentesting-adjacent tasks, that number alone might be reason enough to switch.
Should You Switch to Claude Opus 5?
Here’s our honest take after reading through everything Anthropic published plus what early users are actually saying.
Switch now if you’re on Opus 4.8. There’s no real downside. Same price, better performance on nearly every published benchmark, and it’s already the default on Claude Max.
Consider dropping down from Fable 5 if your workloads are day-to-day coding, business automation, or research assistance rather than genuinely autonomous multi-day agent runs. You’ll save half your API bill and give up less than a percentage point on the hardest coding benchmark.
Stay on Fable 5 or get Mythos access if your work touches cybersecurity research, biosecurity, or long-horizon autonomous agents where the capability ceiling actually matters more than cost.
Stick with Sonnet 5 if you’re running high-volume, low-complexity tasks. Opus 5 is a better model, but Sonnet 5’s $3/$15 pricing (or $2/$10 through the end of August) is still the better economics for simple, repetitive work.
The bigger story here isn’t really about one model. It’s that Anthropic shipped four new models in under two months, and the AI race has visibly shifted from “who’s smartest” to “who gives you the most intelligence per dollar.” Opus 5 is Anthropic’s clearest statement yet that it thinks that second question is the one that actually decides who wins enterprise customers.
FAQ SECTION:
Q1: What is Claude Opus 5? A1: Claude Opus 5 is Anthropic’s newest AI model, released July 24, 2026. It sits between Sonnet 5 and Fable 5 in capability, built to deliver near-Fable-level performance on coding and professional tasks at half the price.
Q2: How much does Claude Opus 5 cost? A2: Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same price as its predecessor, Opus 4.8. A Fast mode is available at twice that price for roughly 2.5x the speed.
Q3: Is Claude Opus 5 better than Claude Fable 5? A3: Not overall. Fable 5 remains Anthropic’s smartest model and keeps a narrow lead on CursorBench and long-horizon autonomous tasks. But Opus 5 gets within 0.5% of Fable 5’s peak coding score at half the cost, making it the better value for most day-to-day work.
Q4: Is Claude Opus 5 available for free users? A4: No. Opus 5 is the default model on Claude Max and the strongest model available on Claude Pro. Free-tier users do not get Opus 5 access.
Q5: Does Claude Opus 5 replace Opus 4.8? A5: Effectively yes. Opus 5 delivers better performance at the same price across nearly every benchmark Anthropic published, so there’s little practical reason to keep using Opus 4.8 unless you have a validated evaluation pipeline built specifically around it.
Q6: What are the weaknesses of Claude Opus 5? A6: Opus 5 trails Claude Mythos 5 on cybersecurity exploit development and long-running autonomous biology research. Anthropic deliberately limited its offensive cyber capabilities as a safety measure, so this is by design rather than an oversight.
