I read the Ars Technica piece on Claude Opus 5's launch the morning it dropped, and something in the comments stopped me cold. Not the usual "wake me when it's AGI" dismissals or the philosophical arguments about sentience. No, what caught my attention was a thread where engineering managers were comparing token costs per task the way they'd compare cloud instance pricing.
Someone had run the numbers: on CursorBench 3.2, Opus 5 at max effort lands within 0.5% of Anthropic's own Fable 5 — the model that launched four months ago as their absolute ceiling — at literally half the cost per task. On OSWorld 2.0, it doesn't just match Fable 5; it surpasses it, at just over a third of the cost. And on ARC-AGI 3, that notoriously hard reasoning benchmark, it scores three times the next-best model's result.
A year ago, the question everyone asked about a new model was "how smart is it?" The discourse around Opus 5 suggests the question has quietly shifted to "how cheap is it?" That's a bigger deal than any benchmark score.
The numbers that matter for production
Let me be concrete about why this changes things. Opus 5 costs $5 per million input tokens and $25 per million output tokens. That's the same price as Opus 4.8, but the performance difference is stark. On Frontier-Bench v0.1, it more than doubles Opus 4.8's score at a lower cost per task. Not marginally better — twice the result for less money.
I've been building AI agent teams for production workflows, and the single hardest constraint has always been cost-per-call. You want to route a complex code review to a capable model, but you also need to run it fifty times a day. At Fable 5 prices, that math doesn't work for most teams. At Opus 5 prices, it starts to.
The consistency improvement matters just as much. Lovable's CEO noted that Opus 5 shows "far less variance run to run." If you've ever tried to put an LLM in a production pipeline, you know that variance is the silent killer. A model that nails a task 90% of the time but flails on the remaining 10% for unpredictable reasons is a model you can't automate around without expensive fallback logic. Lower variance means simpler orchestration, fewer retries, more predictable latency.
The effort setting changes the architecture
Opus 5 ships with something called an "effort setting" — a dial that lets you trade intelligence for token savings. You can dial down for straightforward tasks like classification or summarization, and crank it up for complex reasoning or code generation.
This is actually the feature I'm most interested in, because it collapses a design decision that's been splitting teams for two years. Right now, most production AI systems use model routers — tools like Cursor's internal router or Meta's emerging infrastructure — that select a different model based on task difficulty. Cheap model for simple stuff, expensive model for hard stuff. But that introduces architectural complexity: different APIs, different latency profiles, different failure modes.
An effort setting on a single model means you can use one endpoint for everything. The trade-off is that you're locked into one provider's ecosystem. Worth it? For a team running thousands of calls per day, I suspect yes — the operational simplicity of a single model with a tunable intelligence dial outweighs the flexibility of a multi-model routing strategy. But ask me again when open-weight models like Kimi K3 ($15/M output tokens) close the quality gap further.
What this means for how we build
The real story of mid-2026 isn't that models got smarter. It's that frontier-level intelligence crossed a threshold where it becomes economically viable to use everywhere, every day. Not just for the big query — the one where you carefully craft a prompt and wait for a considered response — but for the hundreds of small, cheap, ephemeral calls that make up a real agent workflow: the intermediate reasoning steps, the self-correction loops, the parallel exploration of multiple approaches.
Opus 5 at $25/M output is cheap enough that those patterns no longer feel extravagant. And that changes the architecture of the systems we build. You stop asking "can I afford to call a model here?" and start asking "what happens if I call a model here three times in parallel and pick the best result?"
The Lindy AI research lead made an observation that stuck with me: the biggest gains on Opus 5 are on "harder, vaguer tasks." That's exactly the territory where human judgment used to be irreplaceable. If the cost of automating a vague, hard task drops by 2-3x, the set of things worth automating expands dramatically.
I don't know where this ends. Model routers are already emerging as a layer above individual models. Open-weight competitors are nipping at the price floor. If Anthropic can profitably sell near-frontier intelligence at these prices today, what happens when the next iteration drops the cost another 2x? Do we end up in a world where the bottleneck isn't model capability but our ability to design workflows that use it efficiently?
That's the question I keep circling back to. The capability race got us here. But the efficiency race is what will determine what we actually build.