Claude Opus 5 Is the Smartest AI in the World — And It Costs 50x More Than It Should

Claude Opus 5 is the smartest AI model in the world. It might also be the most expensive mistake your team makes this quarter.

That's not a hot take — it's arithmetic. Anthropic dropped Opus 5 yesterday, and within hours it climbed to #1 on the Artificial Analysis Intelligence Index with a score of 61, edging out its own bigger sibling Claude Fable 5 (60) and GPT-5.6 Sol (59). The benchmarks are genuinely impressive: it triples the next-best model on ARC-AGI 3, hits state-of-the-art on Frontier-Bench, and matches Fable 5 on CursorBench at half the cost.

But here's the number nobody's putting front and center: $2.03 per task. That's what Artificial Analysis measured as the average cost to complete a single intelligence benchmark task with Opus 5 at max effort. For comparison, DeepSeek V4 Pro does the same work for $0.04. gpt-oss-120b costs $0.06. Kimi K3 — an open-weights model you can download and run yourself — clocks in at $0.95.

The gap between the smartest model and the cheapest competent model has never been wider. And for anyone building AI agents, that gap is the entire ballgame.

What Opus 5 Actually Does Well

Let me give credit where it's due, because the model is genuinely remarkable in specific ways.

The biggest gains are on agentic tasks — the kind where a model needs to make decisions, use tools, and recover from mistakes across multiple steps. On ARC-AGI 3, a benchmark that tests whether models can solve genuinely novel reasoning problems they've never seen before, Opus 5 scores three times higher than the next-best model. That's not incremental improvement. That's a categorical shift.

On OSWorld 2.0, a computer-use benchmark where the model controls an actual desktop environment, Opus 5 beats every other model at any given cost point. It surpasses Fable 5's best result at just over a third of the price. On Zapier's AutomationBench — which measures whether models can complete real business tasks end to end — Opus 5's pass rate is roughly 1.5x the next-best model for the same cost, and it's the only model that hits 100% on the tasks it completes.

The anecdotal evidence from early-access users is telling. One engineer gave Opus 5 a drawing of a machine part and asked it to reconstruct it as a 3D FreeCAD model — but deliberately didn't give it any way to view the drawing. Opus 5 wrote its own computer vision pipeline from scratch to extract geometry from the raw pixel data, then built the model. No competing model could solve the same task after five attempts.

Another early tester gave it a real bug in a popular open-source package manager. Opus 5 found the root cause and fixed an edge case that the community's own patch had missed. A competing model — from another major lab — fixed the surface symptom, declared the bug resolved, and moved on.

A third report came from an engineer at a trading firm who used Opus 5 to build a market data feed for a brand-new exchange in a single session. Previous models couldn't complete the task at all, even with extensive plans. Opus 5 couldn't find a live feed to validate against, so it built its own test harness to verify that its code parsed the exchange's data format correctly. It invented its own QA pipeline mid-task.

That's not a party trick. That's the kind of self-directed problem-solving that separates a model you can trust with real work from one you have to babysit. The consistency matters too — Lovable reported 22% improvement over the previous Opus on their hardest agentic coding tasks, with "far less variance run to run." In production agent workflows, consistency is often more valuable than peak performance. A model that's brilliant 80% of the time and unreliable the other 20% is harder to build on than one that's solid 100% of the time.

The Price Problem Nobody Wants to Talk About

Here's where it gets uncomfortable.

Opus 5 costs $5 per million input tokens and $25 per million output tokens. At max effort, it's verbose — it thinks a lot before answering, which means it generates more tokens per task than most models. Artificial Analysis spent $3,835.51 just to evaluate it on their intelligence index. That's 100 million output tokens for a single benchmark suite. Let that sink in for a moment.

Let's translate to something concrete. If you're running an AI agent that makes 50 model calls to complete a task — a typical workload for anything beyond a simple chatbot — and each call generates a few thousand tokens of reasoning plus output, you're looking at significant spend per user interaction. Multiply by hundreds of tasks per day across a product team, and the math gets painful fast. A startup burning $500/month on API costs can hit five figures before they notice, especially with a verbose reasoning model that "thinks" through every step.

The Hacker News thread on the release is full of people doing this math in real time and not liking the results. One developer tested Kimi K3 on a security auditing benchmark and burned through $100 in extra usage credits at 11% completion. He noted that the same tasks cost a fraction as much with other models. Another commenter pointed out that while Anthropic's own benchmarks frame Opus 5 as cheaper than Sonnet per completed task, third-party measurements from Artificial Analysis put it at 1.25x Sonnet's cost and roughly 2x GPT-5.6's cost per task.

The benchmark-versus-reality gap is the dirty secret of the AI industry. Vendors optimize their cherry-picked comparisons — "Opus 5 is cheaper than its predecessor per task!" sounds great until you realize "per task" means "on the specific benchmarks we chose, at the effort level we selected, using our own token counts." Independent evaluators try to measure honestly. Users experience something different from both, because real workloads look nothing like benchmarks. A model that looks cost-efficient on Frontier-Bench might bleed you dry on your specific agentic workflow, and vice versa.

There's also the data retention angle, which the HN crowd flagged immediately. Claude Fable 5 — Anthropic's absolute frontier model — requires a 30-day data retention period for API access. Your prompts, your code, your data — it sits on their servers for a month. Opus 5 doesn't have this requirement. For a lot of organizations, that single difference makes Opus 5 the practical choice regardless of the intelligence gap. The smartest model in the world is useless if compliance won't let you use it.

The Open-Source Counterargument

Here's the thing that should make Anthropic nervous, even on their victory lap.

Look at the Artificial Analysis leaderboard again. The intelligence gap between Opus 5 (61) and the best open-weights models is shrinking fast. Kimi K3 scores 57 — four points behind the most expensive proprietary model on the planet. GLM-5.2 from Zhipu scores 51. MiniMax-M3 hits 44. DeepSeek V4 Pro sits at 44 as well.

Now look at the cost-per-task column one more time. DeepSeek V4 Pro: $0.04. gpt-oss-120b: $0.06. MiniMax-M3: $0.12. Kimi K3: $0.95.

The math is brutal and unambiguous. DeepSeek V4 Pro achieves 72% of Opus 5's intelligence score for 2% of the cost. That's not a close race — it's a fundamentally different value proposition. If your task doesn't absolutely require frontier-level reasoning, paying 50x more for marginal intelligence gains is a budget mistake, not a technology decision.

The caveat, and it's a real one, is that intelligence index scores don't capture everything. Opus 5's edge on agentic benchmarks — the novel reasoning on ARC-AGI, the tool use on Terminal-Bench, the self-correction on real coding tasks — is where proprietary models still pull ahead meaningfully. Running an agent that needs to navigate a desktop, debug its own code across a codebase, and make judgment calls about trade-offs is categorically different from answering knowledge questions on a benchmark. Open-weights models are catching up on raw intelligence, but the gap on complex, long-horizon agentic workflows remains real and measurable.

How long that gap lasts is the multi-billion-dollar question. Six months ago, the top open-weights model was 15+ points behind the proprietary frontier. Today it's four. At this rate of convergence, the "frontier advantage" becomes very hard to justify at a 50x price premium by early next year.

What This Means If You're Building Agents

If you're building anything with AI agents — and at this point, who isn't — Opus 5's release crystallizes a strategic decision that's been building for months: when do you pay for frontier intelligence, and when do you route to something cheaper?

The answer that's emerging across the industry is model routing. Instead of committing to one model for everything, you use a cheap, fast model for the 80% of tasks that don't require frontier reasoning, and escalate to Opus 5 or Fable 5 only when the task genuinely demands it. The infrastructure for this is getting better — OpenRouter for multi-provider routing, intelligent fallback chains, cost-aware orchestration layers. The economics make it almost mandatory at any meaningful scale.

This is also why open-weights models matter more than ever, even if you never run them yourself. They're the floor. They're the pricing pressure that keeps proprietary providers from charging whatever they want. Without DeepSeek V4 Pro sitting at $0.04/task, Opus 5 at $2.03/task has no counterweight. The existence of competent, freely available open models is what makes the entire ecosystem competitive — and what prevents a return to the locked-in, vendor-controlled pricing that defined the early cloud era.

At CopperRiver, we run open-source models — GLM, DeepSeek, Qwen, MiniMax, Kimi — precisely because the cost-intelligence tradeoff is the right one for most agent workloads. You get 80%+ of the frontier for a fraction of the cost, your data stays on your machine, and you're not at the mercy of a vendor's pricing page changing overnight. There are absolutely tasks where Opus 5 is worth every penny — hard debugging, novel reasoning, complex multi-step workflows where mistakes are expensive and reliability is paramount. The point is being able to choose, not being locked into one provider's pricing tiers.

The Frontier Is Getting Crowded

Six months ago, there was one model at the top of the intelligence rankings, and everyone else was fighting for second. Today, the top four models — Opus 5 (61), Fable 5 (60), GPT-5.6 Sol (59), and Kimi K3 (57) — are separated by four points on a 100-point scale. The next tier — Grok 4.5 (54), GLM-5.2 (51), Muse Spark (51) — is only a few points further back.

This is what a maturing market looks like. The frontier isn't a single model anymore — it's a plateau. And plateaus are where pricing wars happen, where open-source catches up, and where the real differentiation moves from raw intelligence to things like reliability, cost-efficiency, inference speed, and ecosystem depth.

Opus 5 outputs 55 tokens per second. Gemini 3.6 Flash does 235. gpt-oss-120b does 268. On speed alone, Opus 5 is the 11th fastest model among the top 12 in intelligence. For real-time applications — agent loops, interactive coding, anything where latency matters — that's a significant drawback that doesn't show up in intelligence benchmarks.

The smartest model in the world that's also the slowest and most expensive is a niche product, not a default. It's the model you reach for when nothing else works, not the one you run on every request.

The Real Question

Opus 5 is the smartest model available today. That's a fact, verified by independent benchmarks. The title is real.

But enjoy it while it lasts — based on the current release cadence from every major lab, someone else will claim it next month. Maybe OpenAI. Maybe a Chinese lab that open-weights their model and charges nothing. The leaderboard has become a rotating door.

The real question isn't who's at the top of the intelligence rankings this week. It's who gives you the most intelligence per dollar, per task, for the specific work you actually need done. And increasingly, the answer to that question involves open-source models, smart routing, and the freedom to switch providers without rewriting your entire stack.

That's the story that matters. Not the benchmark win — the economics underneath it.

---

CopperRiver is a desktop AI assistant for Mac that runs open-source models locally, browses the web, executes terminal commands, and automates your workflows — without sending your data to the cloud. Plans start at $9/mo. Check it out.