Anthropic | Pricing & Benchmarks
Opus 5 lands cheaper, keeps pace
On July 24, 2026, Anthropic unveiled a new model, Opus 5. It's priced at $5 per million input tokens and $25 per million output tokens — half the price of the current top-tier model, Fable 5 ($10/$50). Yet it matched or beat Fable 5 on 7 of 12 major benchmarks.
Same price tag, more performance under the hood
What surprised people about Opus 5's pricing wasn't a price hike — there wasn't one. Its rate of $5 input / $25 output per million tokens is identical to the prior-generation model, "Opus 4.8." In its official announcement, Anthropic positioned Opus 5 as delivering performance close to Fable 5 at half the price.
Behind this sits an escalating price war among frontier labs. Rival GPT-5.5 is reportedly priced at $15 input / $60 output per million tokens — over three times Opus 5's rate. According to CNBC's coverage, Opus 5 became the default model on Claude Max and the top-tier model on Claude Pro. It's fair to read this in the context of a crumbling assumption: that a high-performance model must be expensive. What used to look like a binary choice between performance and price now looks like a line Anthropic is trying to redraw.
That said, price cuts alone aren't new — OpenAI and Google have both repeatedly revised pricing in the past. What's different this time is that Anthropic held the price steady while packing in performance that beats the prior top-tier model. Read as an industry trend, the competition seems to be shifting from a simple price war toward a race to deliver more capability at the same cost.
The price gap, in numbers
7 wins out of 12 — not a clean sweep
Opus 5 leads overall, but Fable 5 holds its ground by a hair in one coding-specific benchmark.
On Frontier-Bench v0.1, Opus 5 scored 43.3% versus Fable 5's 33.7% — a 9.6-point gap. On GDPval-AA v2, which mimics enterprise knowledge work, Opus 5 also led with 1,861 points against Fable 5's 1,747. But on the coding-specific SWE-bench Pro, Fable 5 held onto the top spot by just 0.8 points — half the price doesn't mean dominance everywhere. In an interview with VentureBeat, an Anthropic spokesperson described Opus 5 as "your daily driver, the model you hand complex work to and review when it's done."
| Fable 5 | Opus 5 |
|---|---|
| Frontier-Bench v0.1: 33.7% | Frontier-Bench v0.1: 43.3% |
| GDPval-AA v2: 1,747 pts | GDPval-AA v2: 1,861 pts |
| SWE-bench Pro: narrow lead | SWE-bench Pro: 2nd, off by 0.8pt |
| $10 in / $50 out | $5 in / $25 out |
Hand it complex work,
then check it
when the work is done.
Heavy users feel it,
light users won't notice
Developers & agent operators
Developers who make heavy API calls through coding agents or batch jobs should see their monthly bill fall by nearly half. Switching from Fable 5 to Opus 5 alone changes the cost structure while holding — or improving — output quality.
Enterprise AI leads
Opus 5 is also offered as the top-tier model on Claude Pro, so internal users on a subscription get the upgrade at no extra cost. Teams paying by usage on the API have the strongest case to evaluate switching.
Individuals & light users
People who only use Chat or a Pro plan occasionally rarely think about per-token pricing in the first place. The savings are real on paper, but they likely won't notice much difference.
What to weigh, and what's still unresolved
Developers currently running Fable 5 in production should first run their own replacement tests on benchmarks close to their actual workload. Even where Opus 5 leads on aggregate scores like Frontier-Bench or GDPval-AA v2, individual tasks can flip the other way — SWE-bench Pro is exactly that case. Rather than switching on aggregate scores alone, a staged rollout is more realistic: (1) route a few days of real tasks to Opus 5, (2) measure actual latency and cost, (3) flip the default once quality holds up or improves.
Anthropic has intentionally trained Fable 5 to a higher standard than Opus 5 for offensive cybersecurity research and autonomous biology tasks, and still expects Fable 5 to be used in those two domains. Deciding to switch purely because Opus 5 is cheaper is premature — it's worth checking whether your own use case touches those exception areas. It's also worth noting these figures come from Anthropic's own internal benchmarks; independent third-party verification is still in its early stages.