Quick Answer
Anthropic released Claude Opus 5 on July 24, 2026 at $5 and $25 per million input and output tokens, the same price as Opus 4.8 and half what Fable 5 costs. Anthropic reports it approaches Fable 5 intelligence on coding and knowledge work, and it is now the default model on Claude Max.
Key Takeaways
- Pricing is unchanged from Opus 4.8 at $5 input and $25 output per million tokens, half of Fable 5’s rates
- Every headline benchmark figure comes from Anthropic’s own testing, so treat the comparisons as vendor-reported
- A new effort setting lets you trade thoroughness against token spend on a per-request basis
- Anthropic’s own charts show Opus 5 still trailing Fable 5 on legal and health evaluations, and GPT-5.6 Sol on one coding benchmark
- It is now the default on Claude Max and the strongest model on Claude Pro, so most paying users get it without changing anything
What is Claude Opus 5?
Claude Opus 5 is Anthropic’s new mid-tier model, released on Friday, July 24, 2026, and available immediately across the Claude apps, the Claude API under the model ID claude-opus-5, and major cloud platforms. It joins a lineup of four models released in under two months: Mythos 5 at the restricted top, Fable 5 as the most capable public model, Opus 5 in the middle, Sonnet 5 below.
The positioning is the interesting part. Anthropic is not pitching Opus 5 as a new flagship. In its launch announcement, the company describes it as a model designed to be used every day, one that comes close to Fable 5’s frontier intelligence at half the price. Anthropic product leader Dianne Penn told Reuters that users should pick Opus 5 for value and reserve Fable 5 for days-long autonomous projects.
That framing tells you where the industry has moved. The competition is no longer about which model tops a leaderboard, but about which one delivers acceptable quality at a cost per task a business can run at scale. Opus 5 is a bet that most work does not need the frontier.
How much does Claude Opus 5 cost?
Opus 5 costs $5 per million input tokens and $25 per million output tokens. That is identical to Opus 4.8, its predecessor, which means the generational improvement arrives at no additional cost per token.
Here is how it sits against the rest of the current Claude lineup:
| Model | Input (per MTok) | Output (per MTok) |
|---|---|---|
| Claude Fable 5 | $10 | $50 |
| Claude Opus 5 | $5 | $25 |
| Claude Opus 4.8 | $5 | $25 |
| Claude Sonnet 5 | $2 (intro, to Aug 31) | $10 (intro, to Aug 31) |
Sonnet 5’s introductory rates run through August 31, 2026, after which it moves to $3 and $15. Anthropic also offers Opus 5 in a Fast mode running roughly 2.5 times the default speed at twice the base price, on the Claude Platform and through Claude Code usage credits. Cached input and batch processing are both discounted, so a well-engineered workload can land below the headline figure.
One detail matters for anyone comparing tiers on contract terms rather than benchmarks: Anthropic says Opus 5 carries no data retention requirements for general access, unlike its two most capable models. For regulated buyers, that can decide the question before a single benchmark is read.
What do the Claude Opus 5 benchmarks actually show?
Before the numbers, the caveat that should travel with all of them: these are Anthropic’s own evaluations, run on Anthropic’s own harnesses. They are self-reported. Independent replication takes weeks, and until it arrives, treat every figure below as a vendor claim rather than settled fact.
With that said, the self-reported results are strong. On Frontier-Bench v0.1, a software engineering evaluation, Anthropic says Opus 5 surpasses all other models and more than doubles Opus 4.8’s score at a lower cost per task. On CursorBench 3.2 at maximum effort, the company reports it lands within half a percentage point of Fable 5’s peak score while costing half as much per task. On ARC-AGI 3, a novel-problem-solving test, Anthropic claims a score three times the next-best model’s.
Two figures stand out for business buyers rather than researchers. On Zapier’s AutomationBench, which measures end-to-end completion of real business tasks, Anthropic reports a pass rate around 1.5 times the next-best model at the same cost per task. On OSWorld 2.0, a computer-use benchmark, it says Opus 5 beats Fable 5’s best result at just over a third of the cost.
On the science side, Anthropic reports gains on every internal life sciences evaluation, the largest being organic chemistry tasks such as inferring molecular structures from spectroscopy data, up 10.2 percentage points on Opus 4.8, and protein variant-effect prediction, up 7.7 points.
The company also publishes a footnote most coverage skips: the Frontier-Bench effort plot comes from an internal run on a specific harness, averaged over five attempts per task, with Opus 4.8 serving as the fallback when safety classifiers refused a request. That disclosure makes the rest of the numbers more credible, not less.
Where does Claude Opus 5 lose?
A launch table with no losses is a marketing document, so this is worth dwelling on. Anthropic’s own published comparison, as reported in Fortune’s coverage and other launch reporting, includes evaluations where Opus 5 does not come first.
According to those published charts, Opus 5 trails Fable 5 on a held-out legal agent benchmark and on HealthBench Professional, a health-domain evaluation. On DeepSWE v1.1, a coding benchmark, OpenAI’s GPT-5.6 Sol reportedly comes out ahead of both Opus 5 and Fable 5. Anthropic also states plainly that Opus 5 remains behind Mythos 5 on cybersecurity and on biology research, and it discloses a slightly higher factual hallucination rate than Opus 4.8.
None of that undermines the release. It does mean the honest summary is state of the art on several evaluations, competitive on most, behind on a few, rather than a clean sweep. If your workload resembles the benchmarks where it loses, test before switching. Our roundup of the top AI models tracks the wider field.
What is the effort setting, and why does it matter?
This is arguably a bigger deal than any individual benchmark score. Opus 5 exposes an effort setting that lets you tell the model how hard to think on a given request, trading thoroughness against token consumption and latency.
In practice that turns model selection into a dial rather than a switch. Instead of routing a cheap classification job to a small model and a hard refactor to a large one, you can run both on Opus 5 at different effort levels and let cost follow task difficulty. Anthropic’s charts plot performance against cost per task at each level, showing the whole trade-off curve rather than one point on it.
The practical implication: if you built a router around model tiers, the effort dial changes the economics of every branch. That is a rebuild worth scheduling, not a setting worth ignoring.
Anthropic also shipped two beta platform features. Mid-conversation tool changes let developers swap which tools Claude can access without invalidating the prompt cache. Automatic fallbacks route API requests flagged by safety classifiers to another model instead of refusing outright, reducing the hard failures that make safety filters painful to build around.
How does Claude Opus 5 handle security and safety?
Anthropic says its automated behavioral audit found Opus 5 to be its most aligned model to date, with the lowest rates of deceptive behavior and the least susceptibility to being tricked into misuse among its current models. On that internal audit, Opus 5 scored 2.3 on overall misaligned behavior, where lower is better, against 2.85 for Opus 4.8, 2.81 for Fable 5 and 3.35 for Sonnet 5. Again: Anthropic’s own audit, Anthropic’s own scale.
The cybersecurity picture is more nuanced and more interesting. Anthropic says it deliberately did not train Opus 5 on cyber tasks, but the model improved at them anyway as a byproduct of general capability gains. Its system card reports that Opus 5 comes close to Mythos 5 at finding software vulnerabilities while remaining substantially behind at exploiting them, which is the gap that determines whether a capability is defensively useful or offensively dangerous.
Because of that gap, Opus 5 ships with lighter cyber guardrails than Fable 5. Anthropic expects its classifiers to intervene around 85% less often, while still blocking binary-based vulnerability scanning, penetration testing and exploit generation. Flagged requests fall back to Opus 4.8 by default in the consumer apps. Enterprises and researchers in Anthropic’s Cyber Verification Program get access to a less restricted version. If you are weighing what any of this means for your own threat model, our explainer on agentic AI security risks covers the broader category.
How does it compare to Fable 5, Sonnet 5 and GPT-5.6?
Against Fable 5, the trade is cost versus ceiling. Fable 5 remains Anthropic’s most capable generally available model and the one the company recommends for long autonomous runs. It costs twice as much and, per Anthropic’s own charts, still leads on some domain evaluations. Fable 5 has also had a more complicated year, having briefly been subject to US export controls before access was restored.
Against Sonnet 5, the trade runs the other way. Sonnet 5 is substantially cheaper and remains the sensible default for high-volume, lower-difficulty work. The effort dial narrows that gap, but Sonnet still wins on raw throughput economics.
Against OpenAI’s GPT-5.6 family, the comparison is genuinely contested. Anthropic’s charts put Opus 5 ahead on several coding and knowledge-work evaluations and behind on at least one. Neither vendor’s numbers should be taken as neutral. If you want a practical rather than a benchmark comparison, our guides to the three major assistants for coding and the broader Claude versus ChatGPT versus Gemini comparison are better starting points than any leaderboard.
One further piece of context: asked about Kimi K3, the open-weight model from China-based Moonshot AI, Penn told Reuters it remains to be seen how open-weight models perform on complicated real-world work. The Trump administration has alleged that Kimi K3 was distilled from Anthropic’s Fable model. That allegation is contested and, as of publication, no public evidence supporting it has been released.
Who should actually switch to Claude Opus 5?
If you are on Claude Max, the decision is already made: Opus 5 is the default. Claude Pro subscribers get it as the strongest available model. Consumer users need do nothing.
For developers and teams on the API, the case is strongest if you are currently running Opus 4.8, since the upgrade is free at the token level and Anthropic reports improvement across the board. If you are on Fable 5 for cost reasons you were never comfortable with, Opus 5 is the obvious thing to test, particularly for agentic coding and computer-use workloads where the reported cost-per-task advantage is largest.
The case is weakest if your workload sits in a domain where Anthropic’s own charts show Fable 5 ahead, or if you have tuned prompts tightly around Opus 4.8. Run a held-out evaluation on your actual tasks before migrating production workloads. Vendor benchmarks are a reason to test, not a reason to switch. If you are using it through the CLI, our guide to Claude Code covers the setup, and our roundup of the best AI coding assistants puts it in context against the alternatives.
FAQ
How much does Claude Opus 5 cost?
Claude Opus 5 costs $5 per million input tokens and $25 per million output tokens, the same as its predecessor Opus 4.8 and half the price of Claude Fable 5, which runs at $10 and $50. A Fast mode that runs roughly 2.5 times quicker is available at twice the base price on the Claude Platform and through Claude Code usage credits.
Is Claude Opus 5 better than Fable 5?
It depends on the task. Anthropic’s own published evaluations show Opus 5 leading on several coding and knowledge-work benchmarks while still trailing Fable 5 on some domain-specific evaluations, including legal and health tests. Anthropic recommends Fable 5 for long, highly autonomous projects and Opus 5 for everyday value. All these figures are vendor-reported.
When was Claude Opus 5 released?
Anthropic released Claude Opus 5 on Friday, July 24, 2026. It was available immediately across the Claude consumer apps, the Claude API under the model ID claude-opus-5, and major cloud platforms. It is the company’s fourth Claude 5 model release in under two months, following Mythos 5, Fable 5 and Sonnet 5.
What is the effort setting in Claude Opus 5?
The effort setting lets you tell Opus 5 how thoroughly to work on a given request, trading intelligence against token consumption, cost and speed. Lower effort produces faster, cheaper results; higher effort spends more tokens for better outcomes. It turns model choice into a per-request dial rather than a fixed tier selection.
Is Claude Opus 5 safe to use for security work?
Anthropic says Opus 5 approaches its restricted Mythos 5 model at finding software vulnerabilities but remains substantially behind at building working exploits. Its cyber guardrails are lighter than Fable 5’s, still blocking binary-based scanning, penetration testing and exploit generation. Enterprises needing fewer restrictions can apply to Anthropic’s Cyber Verification Program.