Claude Opus 5 landed today, July 24, 2026, and Anthropic is making a bold claim: this is the most capable everyday AI model they’ve ever shipped — and it costs exactly the same as the model it replaces. At $5 per million input tokens and $25 per million output tokens, Opus 5 matches Opus 4.8‘s price while delivering performance that, on several benchmarks, beats everything else on the market including Anthropic’s own more expensive flagship.
That last part is worth sitting with. Anthropic positions Fable 5 as its frontier intelligence model — the top of the stack. And yet on benchmarks like Frontier-Bench v0.1 and OSWorld 2.0, Opus 5 either matches or outright beats Fable 5 at a fraction of the cost per task. This isn’t a minor update. It’s a direct challenge to the assumption that “best performance” and “reasonable cost” can’t live in the same model.
How We Got Here: The Opus Lineage and the Cost Problem
Anthropic has been running two parallel tracks for a while now: the Opus line, built for capable, cost-efficient everyday work, and a frontier line for maximum capability. The tension between those two tracks has defined much of the AI pricing conversation in 2025 and 2026.
Developers and enterprises kept running into the same wall: frontier models were often too expensive to run at scale, but lighter models left real capability on the table for complex tasks. OpenAI ran into this with GPT-4o versus o3. Google faced it with Gemini Ultra versus Flash. The industry has been searching for the sweet spot — and Anthropic is claiming Opus 5 found it.
The previous model, Opus 4.8, was already well-regarded for agentic tasks and long-context reasoning. Opus 5 doesn’t just improve on it incrementally. On ARC-AGI 3 — a notoriously difficult test of novel problem-solving — Opus 5 scores three times higher than the next-best model. That’s not a marginal gain. On Zapier’s AutomationBench, Opus 5 completes business automation tasks end-to-end at roughly 1.5 times the pass rate of competitors, for the same cost. Even at its lowest effort setting, it outperforms every other model on that benchmark.
What Claude Opus 5 Actually Does Better
Software Engineering: The Numbers Are Striking
Opus 5’s headline performance is in coding. On CursorBench 3.2 at maximum effort, it lands within 0.5% of Fable 5’s peak score — at half the cost per task. On the AA Coding Agent Index, it sets a new state-of-the-art. The team at Devin (Scott Wu’s AI software engineering company) confirmed it approaches Fable-level performance on their internal FrontierCode 1.1 benchmark, with particular strength on debugging and root-cause analysis.
One anecdote from Anthropic’s testing stands out: given a technical drawing of a machine part with no way to directly view the image, Opus 5 wrote its own computer vision pipeline to extract geometry from raw pixels, then rebuilt the full 3D FreeCAD model from scratch. No competing model could solve the same task after five attempts. That’s not just impressive benchmark performance — it’s the kind of agentic improvisation that signals something qualitatively different about how the model approaches hard problems.
Knowledge Work, Science, and Business Tasks
Beyond code, the gains extend across a surprisingly wide range of domains:
- Life sciences: Opus 5 outperforms Opus 4.8 across all of Anthropic’s life sciences evaluations, with the biggest jump in organic chemistry — 10.2 percentage points higher on inferring molecular structures from spectroscopy data. Protein sequence analysis is up 7.7 percentage points.
- Computer use: On OSWorld 2.0, Opus 5 beats Fable 5’s best result at just over a third of the cost. For anyone building agents that interact with real desktop environments, that ratio matters a lot.
- Financial modeling: Early users at trading firms report 9 percentage points higher accuracy on financial modeling tasks, with a third fewer tool calls and 60% less time than Opus 4.8.
- Legal work: One legal AI platform saw gains specifically in corporate governance and arbitration, with Opus 5 achieving similar performance to Opus 4.8’s maximum reasoning — but with 26% fewer tokens on average.
- Business automation: On Zapier’s internal AutomationBench, Opus 5 hit a 100% pass rate on churn-prevention workflows that previous Claude models couldn’t complete at all.
Visual Output and Agentic Behavior
Anthropic is also highlighting stronger visual generation capabilities. They’ve published two demos: an interactive wind tunnel visualization showing airflow over aerodynamic objects, and an interactive cell illustration. These aren’t just party tricks — they’re signals that Opus 5 can produce production-quality visual artifacts in a single session, which matters for front-end development and data visualization workflows.
The agentic story is the most interesting thread here. Lovable’s co-founder reports 22% improvement on agentic coding tasks over Opus 4.7, with far lower variance run-to-run. JetBrains’ head of AI describes the model as catching its own logical faults during planning rather than after the fact. That’s a meaningful behavioral shift — moving from reactive error correction to proactive reasoning. For developers building on top of these models, consistency and self-correction are often more valuable than raw peak performance.
Safety and Alignment: Anthropic’s Most Careful Model Yet
Anthropic is making a pointed claim about alignment: Opus 5 scores 2.3 on their automated behavioral audit’s overall misalignment metric — lower than Opus 4.8, Sonnet 5, and Fable 5. They’re calling it their most aligned model to date, with the lowest rates of deceptive behavior and the best resistance to manipulation into misuse.
p>On cybersecurity specifically, the picture is nuanced. Opus 5 wasn’t trained on cyber tasks, but general capability improvements have brought it close to their more specialized Mythos 5 model on vulnerability identification. The key distinction: it lags substantially behind Mythos 5 on exploit development — turning a found vulnerability into an actual threat. Anthropic has built classifiers that allow vulnerability discovery in source code but block binary-based scanning, penetration testing, and exploit generation. The Cyber Verification Program gives vetted enterprises and researchers access to a version with fewer restrictions.
This matters in context. As AI models get better at security research, the question of dual-use capability becomes harder to sidestep. Anthropic’s decision to document exactly where the model’s capabilities sit — and where classifiers intervene — is more transparent than most competitors have been. Whether those guardrails are calibrated correctly is a conversation the security community will have in the coming weeks. For more on the broader industry conversation around AI safety at the capability frontier, OpenAI’s long-horizon safety lessons offer useful context.
Pricing, Access, and What’s Actually New for Developers
Opus 5 is available today across all platforms. The pricing structure:
- Standard: $5 per million input tokens, $25 per million output tokens — same as Opus 4.8
- Fast mode: ~2.5x the default speed, at twice the base price
- Claude Pro: Opus 5 is now the strongest model available on this tier
- Claude Max: Opus 5 becomes the new default model
- API access: Available as
claude-opus-5with no data retention requirements for general access
Two beta features ship alongside the model. First, mid-conversation tool changes on the Claude Platform — developers can now swap which tools Claude can access within a conversation without invalidating the prompt cache. That’s a practical quality-of-life fix that agentic developers have been asking for. Second, automatic fallbacks on the API, so flagged requests route to another model rather than simply failing. In Claude.ai, Claude Code, and Claude Cowork, flagged requests fall back to Opus 4.8 by default.
The effort-setting system from Opus 4.8 carries over, letting developers tune the model toward speed and cost savings or maximum capability depending on the task. This is increasingly standard — OpenAI does something similar with reasoning effort levels in the o-series — but Anthropic’s benchmark data suggesting Opus 5 beats competitors even at lower effort settings is a meaningful differentiator if it holds in real workloads. For a sense of how competitors are approaching enterprise AI capability gaps, it’s worth reading about NTT DATA’s experience deploying OpenAI Codex for incident analysis.
What This Means for Developers and Businesses
Here’s the practical read: if you’re running Opus 4.8 in production today, the upgrade path is essentially free. Same price, meaningfully better performance on most task categories. The effort settings give you levers to actually reduce costs below Opus 4.8 levels on simpler tasks while reserving higher compute for the hard stuff.
For teams evaluating whether to use Opus 5 versus Fable 5: the gap has narrowed significantly. On coding benchmarks, Opus 5 is within a rounding error of Fable 5 at half the cost. Unless you specifically need Fable 5’s edge on biology research or want access to Mythos 5’s cybersecurity depth, Opus 5 is probably the right default for most enterprise workloads starting today.
The life sciences improvements are also worth flagging for researchers. A 10-point jump on organic chemistry tasks and meaningful gains on protein analysis suggest Opus 5 could become a genuine research tool — not just a coding assistant wearing a lab coat. The push toward AI in scientific research is accelerating across the industry, and Anthropic is positioning Opus 5 squarely in that space.
Frequently Asked Questions
What is Claude Opus 5 and how does it differ from Opus 4.8?
Claude Opus 5 is Anthropic’s newest general-purpose AI model, replacing Opus 4.8 as the default on Claude Max and the top model on Claude Pro. It delivers substantially better performance across coding, knowledge work, and scientific tasks at the same price point — $5 per million input tokens, $25 per million output tokens.
How does Claude Opus 5 compare to OpenAI’s and Google’s leading models?
On the benchmarks Anthropic cites, Opus 5 outperforms GPT-4-class models and Gemini Ultra-class models on coding (Frontier-Bench, CursorBench), business automation (Zapier AutomationBench), and computer use (OSWorld 2.0). Independent verification of these claims is still pending, but early-access partner feedback from companies like Zapier, Lovable, and Box aligns with the benchmark story.
Is Claude Opus 5 safe for enterprise use involving sensitive data?
Consistent with prior Opus models, Opus 5 has no data retention requirements for general API access. Anthropic also reports it as their most aligned and least manipulation-susceptible model to date. For regulated industries, the Cyber Verification Program provides a pathway to more permissive access with additional vetting.
When is Claude Opus 5 available and how do I access it?
Opus 5 is available today, July 24, 2026, across Claude.ai, the Claude API (as claude-opus-5), Claude Code, and Claude Cowork. Fast mode is also available at twice the standard price for approximately 2.5x speed. No waitlist — it’s live now.
Anthropic’s pricing discipline here is the most interesting long-term signal. They’ve delivered a model that genuinely challenges their own flagship on several dimensions without charging more for it. If that pattern holds across future generations, it accelerates the timeline at which frontier-quality AI becomes accessible to mid-market businesses and independent developers — not just well-funded enterprises with the budget to run Fable 5 at scale. The real test will be whether the benchmark gains translate to the messy, ambiguous tasks that real users throw at these models every day. Based on the partner testimonials, early signs point toward yes.
