Anthropic just shipped its fourth new model in under two months. On July 24, 2026, the company released Claude Opus 5, positioning it as the model most people will actually use day to day — not because it’s the smartest thing Anthropic has ever built, but because it gets close enough to their smartest model, Claude Fable 5, while costing half as much. If you’ve been following our coverage of Claude Sonnet 5, Opus 5 is the next chapter in the same story: Anthropic is no longer competing purely on raw intelligence. It’s competing on how much of that intelligence you can afford to use every single day.
This matters beyond Anthropic’s own product line. It signals where the entire AI industry is heading in the second half of 2026 — away from headline-grabbing “smartest model” claims and toward a much more practical question: which model gives a business the most usable intelligence per dollar, at the volume real work actually requires?
Here’s everything worth knowing about Claude Opus 5 — what it is, what changed, how it benchmarks, what it costs, and who should actually care.
What Is Claude Opus 5?
Claude Opus 5 is the newest release in Anthropic’s Opus tier, which sits below the company’s top-of-the-line Fable 5 and restricted-access Mythos 5 models, but above the Sonnet and Haiku lines in raw capability. It’s the fourth model Anthropic has released since June 2026, following Mythos 5, Fable 5, and Sonnet 5 — an unusually fast release cadence even by 2026’s standards.
Anthropic describes Opus 5 as “a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price.” That’s a deliberate positioning choice. Rather than trying to reclaim the “smartest model” title, Anthropic built Opus 5 specifically for the middle band of everyday professional work — the tasks that make up the bulk of what businesses and individuals actually use AI for, as opposed to the rare, extremely hard problem that needs the absolute frontier model.
It’s designed to be used constantly rather than sparingly: building presentations and reports from a rough outline, working through data analysis and financial models, and researching across dense source material to produce something close to ready-to-use output. It also verifies its own work and recovers from errors with less need for a human to step in mid-task.
What’s Actually New
A few features distinguish Opus 5 from its predecessor, Opus 4.8, beyond a straightforward capability bump:
An effort toggle for cost control. Opus 5 lets users choose between low, medium, and high effort settings, letting teams trade capability for cost depending on the task. Routine, high-volume calls can run on low effort to save money, while genuinely difficult prompts can be pushed to high effort when it matters. Anthropic’s own guidance suggests starting at extra-high effort for coding and agentic work and high effort elsewhere, then adjusting downward from there — since even Opus 5’s lower effort settings reportedly outperform the equivalent settings on Opus 4.8.
A fast mode. For latency-sensitive use cases, Opus 5 offers a mode that runs at roughly 2.5 times the standard speed for double the base price — a straightforward dial for teams that need speed more than they need to save money on a given task.
Far fewer safety-classifier interventions. Anthropic says it expects Opus 5’s safety classifiers to intervene around 85% less often than they do for Fable 5. When a request is flagged inside Claude.ai, Claude Code, or Claude Cowork, it now falls back to Opus 4.8 by default rather than blocking the request outright, and API users can enable the same fallback behavior.
No data retention requirement. Consistent with earlier Opus models, Opus 5 carries no data retention requirement for general access — a detail Anthropic flagged specifically for enterprise customers with strict zero-data-retention policies.
A Cyber Verification Program (CVP) for restricted work. Enterprises and researchers already enrolled in Anthropic’s CVP get access to a version of Opus 5 with fewer cybersecurity-related safeguards, intended to unblock legitimate security research that the model’s default guardrails would otherwise interfere with.
Benchmark Performance: How Opus 5 Actually Stacks Up
Anthropic’s benchmark claims for Opus 5 are aggressive, and several independent write-ups have started confirming the broad strokes. On Frontier-Bench v0.1, an agentic terminal coding benchmark, Opus 5 scores 43.3%, more than double Opus 4.8’s 18.7% and clearly ahead of Fable 5’s 33.7% — all at a lower cost per task. On GDPval-AA v2, a human-graded knowledge-work evaluation, Opus 5 leads outright with an Elo score of 1,861, ahead of both Fable 5 (1,747) and OpenAI’s GPT-5.6 Sol (1,736).
On ARC-AGI-3, a benchmark built around novel problem-solving, Opus 5 reportedly scores roughly three times higher than the next-best model. On OSWorld 2.0, a computer-use benchmark, it surpasses Fable 5’s best result at just over a third of the cost. Independent aggregator Artificial Analysis places Opus 5 at the top of its Intelligence Index as well, ranking it first out of 187 tracked models, narrowly ahead of Fable 5 and GPT-5.6 Sol.
That said, Opus 5 doesn’t win across the board, and Anthropic is fairly transparent about where it falls short. On SWE-bench Pro, the hardest real-repository software engineering benchmark, Opus 5 scores 79.2% — behind both Mythos 5 (80.3%) and Fable 5 (80.0%), though still well ahead of Opus 4.8’s 69.2%. On certain agentic coding evaluations like DeepSWE, GPT-5.6 Sol edges ahead. And on cybersecurity-related tasks specifically, Anthropic says Opus 5 remains behind Mythos 5, which is why cybersecurity researchers with legitimate needs are directed toward the CVP program rather than the general release.
The honest takeaway: Opus 5 isn’t Anthropic’s smartest model in every category, but it’s close enough on the benchmarks that matter most for everyday work — coding, knowledge tasks, and computer use — while costing meaningfully less than the model it’s chasing.
Pricing and Availability
Claude Opus 5 is priced at $5 per million input tokens and $25 per million output tokens — unchanged from Opus 4.8’s pricing, meaning users get a substantial capability upgrade for the same cost as before. For comparison, Fable 5 is priced at $10 per million input tokens and $50 per million output tokens, exactly double Opus 5’s rate.
The model is rolling out across all of Anthropic’s platforms starting immediately. It becomes the new default model on Claude Max, Anthropic’s premium consumer subscription tier, and the strongest model available on Claude Pro. Developers can access it through the Claude API using the model string claude-opus-5, and it’s also available through Claude Code and Claude Cowork.
For teams evaluating whether to upgrade, the practical math is straightforward: at standard pricing, Opus 5 delivers meaningfully better benchmark performance than Opus 4.8 at an identical price point, and comparable-or-better performance than Fable 5 at half the cost. The main reason to still reach for Fable 5 is for the longest, most demanding autonomous tasks, where Anthropic continues to position Fable 5 as the stronger option.
Opus 5 vs. Fable 5 vs. Sonnet 5 vs. Mythos 5
With four new Claude models released in under two months, knowing which one fits a given task is genuinely confusing. Here’s how they break down:
| Model | Best For | Input / Output Price (per 1M tokens) | Availability |
|---|---|---|---|
| Mythos 5 | Restricted, high-sensitivity research (cybersecurity, biology) | Not publicly listed | Vetted partners only |
| Fable 5 | Days-long autonomous projects, hardest coding tasks | $10 / $50 | General availability |
| Opus 5 | Daily professional work, coding, knowledge tasks | $5 / $25 | General availability |
| Sonnet 5 | Agentic coding at a lower cost, long tool-use sessions | $2 / $10 (intro pricing) | General availability |
If you’re choosing between Opus 5 and Sonnet 5 specifically, the decision mostly comes down to task complexity versus budget: Sonnet 5 remains the cheaper, faster option well-suited to long agentic coding sessions, while Opus 5 is built for tasks where the extra reasoning depth — financial modeling, dense research synthesis, high-stakes analysis — pays for itself.
What This Means for Businesses, Freelancers, and Developers
The effort-toggle feature is arguably the most practically important part of this release, because it turns “how good is this model” into “how good is it at the budget I’m willing to spend” — a much more useful question for anyone running AI at real volume rather than testing it on a handful of prompts.
For freelancers and small teams — a group we covered in detail in our ChatGPT vs. Gemini comparison — this kind of cost control matters more than benchmark bragging rights. A freelance analyst running dozens of client reports a month benefits far more from a dial that lets routine work run cheap and only escalates cost for the genuinely hard tasks than from a model that’s marginally smarter but priced without flexibility.
For developers, the built-in fallback behavior is worth knowing about specifically: because Opus 5’s safety classifiers intervene far less often than Fable 5’s, teams that previously avoided Anthropic’s top models for certain workflows due to over-triggered safety blocks may find Opus 5 meaningfully easier to build around, particularly for legitimate technical and security-adjacent use cases.
The Bigger Picture: Why Anthropic Is Racing This Fast
Opus 5’s launch follows a genuinely turbulent stretch for Anthropic’s top-tier models. Fable 5 and Mythos 5 were suspended for 18 days in June 2026 after a U.S. government directive tied to cybersecurity concerns — specifically, a technique researchers had documented for eliciting dangerous outputs from Fable 5. Export controls were lifted and access restored on July 1, 2026, but the episode underscored how much scrutiny frontier-level models are now facing from regulators, not just competitors.
Releasing four models in under two months — Mythos 5, Fable 5, Sonnet 5, and now Opus 5 — also reflects the broader competitive pressure across the industry. Google shipped Gemini 3.6 and two new Gemini 3.5 Flash variants around the same time, and OpenAI’s GPT-5.6 family reached general availability just weeks earlier. The AI labs are no longer just racing to build the smartest model; they’re racing to make near-frontier intelligence affordable enough that it becomes the default choice for ordinary daily work, not an occasional splurge.
Anthropic is also reportedly preparing for a public listing later this year, which adds a business dimension to this release cadence: demonstrating that near-frontier intelligence can be delivered profitably at consumer-friendly prices is exactly the kind of story that matters heading into an IPO.
Frequently Asked Questions
Is Claude Opus 5 better than Claude Fable 5? Not universally — Fable 5 remains Anthropic’s most capable model overall, particularly for the longest autonomous tasks and the hardest coding benchmarks. But Opus 5 matches or beats Fable 5 on several major evaluations, including Frontier-Bench and GDPval-AA, at half the price, making it the better practical choice for most everyday work.
How much does Claude Opus 5 cost? $5 per million input tokens and $25 per million output tokens — the same price as its predecessor, Opus 4.8, but with a substantial performance upgrade.
Is Claude Opus 5 available to Claude Pro and Max users? Yes. It’s the new default model on Claude Max and the strongest model currently available on Claude Pro.
What is the effort toggle in Claude Opus 5? A setting that lets users choose low, medium, or high effort per task, balancing cost against capability. Lower effort settings are cheaper and faster; higher settings push more reasoning into harder problems.
Does Claude Opus 5 retain user data? No. Consistent with previous Opus models, Opus 5 has no data retention requirement for general access, which matters for enterprise customers with strict compliance policies.
How does Claude Opus 5 compare to GPT-5.6 Sol and Gemini? On several key benchmarks, including GDPval-AA v2, Opus 5 currently leads OpenAI’s GPT-5.6 Sol. GPT-5.6 Sol edges ahead on some specific agentic coding evaluations like DeepSWE. Independent aggregator Artificial Analysis currently ranks Opus 5 first overall on its Intelligence Index, narrowly ahead of both Fable 5 and GPT-5.6 Sol.
A Few Things Worth Being Skeptical About
Every one of these benchmark numbers comes from Anthropic’s own testing harness, run under Anthropic’s own conditions — a detail worth keeping in mind before treating any single score as gospel. Independent verification is starting to catch up, and where it has (like Artificial Analysis’s Intelligence Index), the broad ranking largely holds. But a few caveats are worth flagging for anyone making a purchasing decision based on these numbers alone.
First, the Frontier-Bench figures Anthropic published were run on an internal harness, and in at least one documented case, Opus 4.8 served as the silent fallback whenever safety classifiers refused a request meant for Opus 5 or Fable 5 — meaning some of the reported score may reflect a blend of models rather than Opus 5 in isolation. Second, effort-level defaults matter enormously to real-world cost: Anthropic’s own guidance to start at extra-high effort for coding work means the headline benchmark scores likely reflect a more expensive configuration than what most cost-conscious teams will actually run day to day. And third, “half the price of Fable 5” is true on paper, but the actual cost of a workflow depends heavily on how many tokens a task consumes at a given effort level — a cheaper per-token rate doesn’t always translate to a cheaper finished task if the lower-cost model needs more tokens or more retries to get there.
None of this means the headline claims are wrong. Multiple outlets running their own smaller-scale tests have found Opus 5 performing competitively with Fable 5 on practical coding tasks. But as with any lab-reported benchmark, the sensible approach is to treat Anthropic’s numbers as a strong starting signal, then validate against your own actual workload before committing a team’s budget to a switch.
How to Start Using Claude Opus 5
Getting access to Opus 5 doesn’t require any special sign-up beyond an existing Claude subscription or API account. Claude Max and Claude Pro subscribers on claude.ai will see Opus 5 available immediately, with Max users defaulting to it automatically. Developers building on the Claude API can switch to it by specifying claude-opus-5 as the model string in their API calls, and teams using Claude Code or Claude Cowork will find it available as a selectable model within those tools as well.
For teams currently on Opus 4.8, the practical move is straightforward: since pricing is identical, there’s little reason not to test Opus 5 against an existing workload immediately. For teams currently paying Fable 5 pricing for tasks that don’t strictly need Fable 5’s absolute ceiling — which, based on Anthropic’s own benchmark breakdown, describes a large share of everyday coding and knowledge-work tasks — Opus 5 is worth a direct side-by-side comparison before the next billing cycle.
The Bottom Line
Claude Opus 5 isn’t trying to be the smartest AI model in the world — that title still belongs to Fable 5, at least for now. What it’s trying to be is the model that makes the most economic sense for the overwhelming majority of daily professional work: reports, analysis, research, and coding tasks that don’t need the absolute frontier, but do need to be fast, reliable, and affordable at real volume. With pricing unchanged from Opus 4.8 and benchmark scores that more than double it in several categories, Opus 5 is a strong argument that the real AI competition in the second half of 2026 has shifted from “who’s smartest” to “who delivers the most usable intelligence per dollar” — and for now, Anthropic has a genuinely competitive answer to that question.
