Opus 5’s Big Promise: Cheaper Power, Uneven Smarts
Claude Opus 5 is Anthropic’s latest large language model that aims to outperform its predecessor Fable 5 on most benchmarks while charging roughly half the per‑token price, with especially strong gains in knowledge work, agentic search performance, and novel problem‑solving but weaker results in legal question answering and multidisciplinary reasoning tasks. The headline story is hard to miss: Anthropic has released a model that beats Fable on most of the company’s own benchmarks for half the price. In a market where every major vendor is racing to cut LLM token pricing, Opus 5 lands at USD 5 (approx. RM23) per million input tokens and USD 25 (approx. RM115) per million output tokens, identical to Opus 4.8 and half what Anthropic charges for Fable 5. Cheaper, safer, and “agentic” sounds like an easy upgrade—but the details show a more complicated trade‑off.
How Fable’s Turbulence Set the Stage for Opus 5
To understand Opus 5, you have to see it as a reaction to Fable 5’s rapid rise and equally rapid troubles. Claude Fable 5 is arguably the most influential LLM to launch in recent memory, marking the biggest leap in capabilities since Opus 4.5 debuted in November 2025. After launching on June 9, Fable went dark internationally three days later under Commerce Department export controls, then returned July 1. By July 22 it was the subject of a White House accusation that a Chinese lab had covertly distilled it, a claim some experts dispute. Last month, Anthropic blocked access to Fable 5 for users worldwide after the US government expressed concerns about how its offensive cybersecurity capabilities could be used by foreign countries, restoring access several weeks later with restrictions. Against that backdrop, Opus 5 is positioned as a safer, more controllable flagship—one that trades some cutting‑edge capability for a calmer risk profile.
Benchmarks: Agentic Search Wins, Law and Reasoning Losses
On paper, Claude Opus 5 benchmarks look strong. Anthropic says Opus 5 comes close to or even exceeds Fable 5’s performance in most areas for roughly half as much per token. The new model outperforms Fable 5 on tasks like knowledge work, novel problem solving, and agentic search, where it can plan, search, and synthesize across sources more effectively. But the story turns once you move into specialized domains: Opus 5 falls somewhat short of Fable 5 in answering legal questions and performing multidisciplinary reasoning without additional tools. It is also substantially behind Mythos 5 at exploiting cybersecurity vulnerabilities, reflecting Anthropic’s choice to dial back offensive capability. Even on classic knowledge tests the gains are mixed. On the closed‑book AA‑Omniscience benchmark, Opus 5 was 11% more accurate than Opus 4.8, while its hallucination rate was 6% higher—a reminder that more confident models can be more confidently wrong.
Token Pricing Meets Task Economics: Overthinking Has a Cost
Opus 5’s biggest marketing number is its token price, but the more interesting story is how it thinks. In terms of pricing, Opus 5 lands at USD 5 (approx. RM23) per million input tokens and USD 25 (approx. RM115) per million output tokens, while GPT‑5.6 Sol sits at USD 5 (approx. RM23) and USD 30 (approx. RM138), and Kimi K3 undercuts both at USD 3 (approx. RM14) and USD 15 (approx. RM69), with cache hits billed at USD 0.30 (approx. RM1.38). Yet per‑token price turned out to be a poor predictor of per‑task cost. In private head‑to‑head testing on a language‑engineering task, Sol at medium effort matched or beat Opus 5 at every effort tier, mapping audio to text with higher transcription accuracy (96.1% versus 90.7% on the harder unit). Opus 5 billed less per output token than Sol under batch pricing, USD 12.50 (approx. RM58) per million against USD 15 (approx. RM69), yet cost 21% more per unit at its cheapest setting and 80% more at high effort because it generated 1.5 to 2.5 times as many output tokens. Adaptive thinking is on by default, and sometimes it simply thinks too much.
Safety, Effort Levels, and the Limits of “More Thinking”
Anthropic’s own documentation quietly admits what the benchmarks hint: Opus 5’s higher‑effort modes can backfire. Its migration guide in the Opus 5 system card warns that max effort can produce diminishing returns and overthinking on simpler tasks. On FrontierCode, scores fell above high effort because the model made unnecessary refactors and other out‑of‑scope edits, a problem largely fixed by a brief instruction to stay within the requested scope. External pilot users likewise reported cases where higher effort made the model perform worse. During pre‑deployment testing, Anthropic concluded that Opus 5 showed lower rates of misaligned behavior than its predecessors, exhibiting lower deceptive behavior and proving harder to trick into misuse, while avoiding reckless actions that could create hard‑to‑reverse side effects. The tension is clear: more agentic search performance and adaptive thinking, but also more opportunities for hallucination and over‑editing. Opus 5 is safer and cheaper, but it is not automatically smarter.
The lesson from Claude Opus 5 benchmarks is that headline comparisons—“beats Fable 5 on most metrics at half the price”—hide important weaknesses in legal expertise, multidisciplinary reasoning, and cost per completed task. In this release, Anthropic has prioritized agentic search and safety over raw specialized capability. That may be the right trade‑off for many deployments, especially after Fable 5’s turbulent run, but it makes Opus 5 less of a universal upgrade and more of a targeted tool. For teams that care about precise legal answers or tightly scoped reasoning, Fable 5 and other rivals like GPT‑5.6 Sol remain competitive alternatives. Opus 5 is a step forward—but a selective one.
