On July 24, Anthropic released Claude Opus 5. The line that jumped out of the announcement was: frontier intelligence of Claude Fable 5 at half the price.
A lot of Chinese coverage translated this as “Opus 5 cuts prices in half.” That summary is half right and half wrong.
Half right because Anthropic’s own marketing line is exactly that: “frontier intelligence of Claude Fable 5 at half the price.” Half wrong because Anthropic did not cut the token price. Opus 5 still lists at $5 per million input tokens and $25 per million output tokens, identical to Opus 4.8. Fast mode is still 2x the base price on the Claude Platform and through usage credits in Claude Code.
So where does the “half price” come from?
It comes from cost per task. On CursorBench 3.2 at max effort, Opus 5 finished within 0.5% of Fable 5’s peak score, but at half the cost per task. On OSWorld 2.0, it surpassed Fable 5’s best result at just over a third of the cost. On ARC-AGI 3, its score is three times the next-best model, and the next-best here is not Opus 4.8 but other labs’ frontier models.
In other words, what got cut in half is not the API rate. It is unit intelligence cost. You pay less for the same business output, because the model is more capable, uses fewer reasoning tokens, and finishes the job on the first pass without rework.
What is actually changing
After reading through the Anthropic post and the early-access customer quotes, I think this shift is being undersold.
First, the API price table is frozen, the capability price table is moving. This has been the pattern across Opus 4.5, 4.7, 4.8, and now Opus 5. Anthropic runs a calibrated playbook: leave the per-token rate untouched, then push the cost-saving story into “fewer reasoning tokens, more accuracy, more stability.” Customer Box reports Opus 5 runs 8% higher than Opus 4.8 on their internal evals, with data analysis up 11% and due diligence up 17%. Customer Lovable reports +22% on hardest agentic coding tasks. A legal workflow evaluation reports 26% fewer generated tokens at max reasoning versus Opus 4.8. A trading firm reports “roughly a seventh of the reasoning tokens and under half the latency” of Opus 4.8.
Read these together and the message is clear: Anthropic has broken the otherwise abstract promise of “smarter, cheaper models” into individually verifiable small metrics.
Second, “half the price” is relative to Fable 5, not to Opus 4.8. This matters. Fable 5 is Anthropic’s frontier flagship and was already priced higher, so “close to Fable 5 capability at half its cost” is a relative position against the top tier, not a general price cut. Anthropic also gave a separate number for Opus 4.8: on Frontier-Bench v0.1, Opus 5 more than doubled Opus 4.8’s performance at a lower cost per task. So relative to the previous generation Opus, the capability is doubled and the per-task cost is compressed — that is the industry-meaningful version of “prices falling, capability rising.”
Third, more verbose does not mean better value. AI Pricing Guru ran Opus 5 through 49 deterministic reasoning, extraction, coding, and pricing tasks on launch day: 49/49 accuracy (tied with Opus 4.8), but 1,989 completion tokens vs. 630 for Opus 4.8. With the token rate unchanged, the per-call bill could actually rise. This is a counterintuitive reminder: token price unchanged but real spending up is exactly the failure mode you hit when you instrument API spend instead of task-level spend.
What this means for builders
I wrote earlier about Cursor Router, which is about how an IDE routes work across multiple models to compress cost per commit. Cursor Router answers “which model should do this task.” Opus 5 answers “the bill for the same task just got cut in half.” On the surface these are different problems; underneath they point at the same underlying shift: the model layer’s economic metric is moving from per-token to per-task.
The upside of per-task thinking is that you do not have to bet on the model behaving itself. At max effort on CursorBench 3.2, Opus 5 finishes more jobs, uses fewer tokens, and the bill and success rate both look good. The catch is that you have to actually measure it. Reading the API price table alone gives you “Opus 5 did not get cheaper,” which is the wrong conclusion.
Three concrete moves for anyone using or considering Claude.
First, put cost per task into your evaluation matrix. If you run Claude Code, Cursor, or Devin with a harness, run Opus 5 and Opus 4.8 against the same accepted-patch / task-completion / review-time / retry-rate metrics. Anthropic’s automated behavioral audit gave Opus 5 a misaligned-behavior score of 2.3, the lowest of recent models, which means it is less likely to wander and quietly inflate per-task cost through rework.
Second, set thresholds on reasoning effort instead of defaulting to max. Even at its lowest effort setting, Opus 5 passes more tasks than any other model. That means everyday tasks do not need to be cranked up; what you save is not the token rate but the total reasoning tokens. Anthropic’s own number is Opus 5 generates 26% fewer tokens than Opus 4.8 at max reasoning. If you were already running low effort, the bill improvement compounds further.
Third, plan for Opus 5’s fallback path. In Claude.ai, Claude Code, and Claude Cowork, requests that get flagged by the safety classifiers fall back to Opus 4.8 by default. Most products do not notice, but if you call Opus 5 directly and budget by a fixed per-task cost, the fallback adds variance. Early customers Box, Lovable, Notion, Cognition, and Zapier all report strong numbers, but they all repeat the same line: run it on your real workflow before you change your default.
The inflection I see
Two capabilities in Opus 5’s story convinced me more than any benchmark number.
One is agent self-verification and iteration. Anthropic describes Opus 5 as “much stronger at verifying its work and iterating carefully until it succeeds.” The concrete example: on a Frontier-Bench task, Opus 5 was given a drawing of a machine part and asked to write code to rebuild it; the model wrote its own computer vision pipeline, pulled the geometry out of the raw pixels, and reconstructed the full part. Cursor’s early testing reports Opus 5 is significantly stronger than Opus 4.8 on difficult debugging and root-cause analysis.
The other is judgment. Early customer Igor Ostrovsky (Co-Founder) wrote flat out: “what stands out about Claude Opus 5 is judgment.” Another customer, Zimu Li, gave a concrete scene: during a rearchitecting session, Opus 5 pushed back on a design Zimu proposed, and did not fold when Zimu insisted. That is more revealing than any benchmark. As agents go deeper into decision flows, judgment is worth more than completion rate, because being wrong once costs more than being slow once.
Neither of these lives on the token price sheet. Both can only be verified through per-task evaluation and per-decision review. So the real change on July 24 is not “Opus 5 is cheaper than Opus 4.8.” It is that the yardstick inside enterprises for choosing a model is shifting from “how much per million tokens” to “how much per completed business task.”
That line connects back to Cursor Router and earlier routing tools, and forward to the standardization of agent harnesses and the productization of per-task billing. The model layer is giving up the center of the token-economics conversation and pushing the cost-saving responsibility onto the harness and the toolchain. That is the bill everyone — model vendors, agent products, builders — will have to redo over the next year.