Claude Opus 5: Near-Fable Intelligence at Half the Cost

On July 24, Anthropic released Claude Opus 5 - a model that comes "close to the frontier intelligence of Claude Fable 5 at half the price," according to their announcement. The reaction on Hacker News was immediate and massive: the story hit 1,378 points and 746 comments within hours.

Here is what is in the release, what the HN thread got right and wrong, and why Opus 5 might be the most practical model Anthropic has shipped this year.

What Claude Opus 5 actually is#

Opus 5 is not a new frontier model. Anthropic is explicit: "Claude Opus 5 is not more capable overall than our most capable general-access model, Claude Fable 5." It is an Opus-class model that has been significantly upgraded to close the gap with Fable while staying at the Opus 4.8 price point ($5/M input tokens, $25/M output tokens).

The model tops the Artificial Analysis Intelligence Index with a score of 61, ahead of GPT-5.6 Sol (59) and Fable 5 (60). It also claims state-of-the-art results on Frontier-Bench v0.1 and GDPval-AA. On CursorBench 3.2 at max effort, it performs within 0.5% of Fable 5's peak score at half the cost per task.

The benchmark that caught the most attention on HN is ARC-AGI-3, where Opus 5 scores 30.2% - roughly three times the next-best model. ARC-AGI measures a model's ability to solve novel problems from few examples, and the jump from every previous model is large enough that commenters questioned whether the benchmark had been compromised through training-data contamination. Anthropic's system card addresses this obliquely, noting that Opus 5 was not specifically trained on ARC-style tasks.

What HN is saying#

The HN thread is worth reading in full at news.ycombinator.com/item?id=49038433. Several themes emerged:

Cost vs. capability tension. The most-upvoted reaction was confusion: if Fable 5 is still the most capable model, why does Opus 5 exist? The answer came from multiple commenters pointing out that Fable 5 is twice the price and not included in Claude Pro subscriptions. One commenter laid out the math: "Half the price of Fable 5 and usable with 100% of your subscription means roughly 4x the usage." Another noted that Fable 5 on long coding tasks with subagents "will easily chew through hundreds of dollars in a single run."

No data retention requirement. Several commenters flagged that Opus 5 does not have the 30-day data retention requirement that applies to Fable 5 and Mythos 5. The release confirms this: "Consistent with prior Opus models, Opus 5 does not have data retention requirements for general access." For regulated industries and enterprise deployments, this is a significant practical advantage.

ARC-AGI-3 sparks debate. The 30.2% score on ARC-AGI-3 generated the most technical discussion. One commenter asked whether models are "actually developing fluid intelligence" or if benchmarks are being trained toward. A counterpoint noted the $20k total cost to achieve that score versus GPT-5.6 Sol's 7.8% at $20k per task, suggesting the capability jump is real and cost-efficient.

Model routing acceleration. A top comment observed that the proliferation of models at different price-performance points means "model routing is the fastest growing segment in AI right now." With Opus 5, Sonnet 5, Fable 5, and Mythos 5 all available simultaneously, the routing decision is no longer about picking one model - it is about the infrastructure cost of picking the wrong tier for a given task.

Skepticism about benchmarks. A handful of comments accused Anthropic of "manipulative" benchmark presentation, pointing to inconsistent highlighting of leading values across tables. Others pushed back, noting the benchmarks come with methodology footnotes and that the system card is a 190-page PDF with detailed evaluation descriptions.

Dev-to-dev take: Opus 5 changes the daily-driver equation#

The most interesting thing about Opus 5 is not the benchmark scores. It is that Anthropic has now created a clear three-tier pricing ladder for production AI workloads - Sonnet for cheap throughput, Opus for everyday agentic work, and Fable (or Mythos) for the hardest problems. That ladder did not exist before. Opus 4.8 was a meaningful step down from Fable 5; Opus 5 closes that gap to the point where many teams will never need to pay Fable prices.

The mid-conversation tool changes feature - where you can swap tools mid-chat without invalidating the prompt cache - is a quality-of-life improvement that matters more than most benchmark points. It makes agent loops where the model picks up new tools mid-stream dramatically cheaper.

The automatic fallback to Opus 4.8 on safety classifier refusals is also welcome. Fable 5's refusal rate has been a pain point for production teams. If Opus 5's classifiers are 85% less restrictive than Fable 5's (per the announcement), and flagged requests drop down to 4.8 instead of hard-blocking, the effective uptime of your agent pipeline goes up significantly.

On the cybersecurity front, Anthropic deliberately did not train Opus 5 on offensive cyber tasks, and the model remains well behind Mythos 5 on exploit development. For most developers doing defensive work or vulnerability discovery in their own source code, the permissive source-code scanning allowance is exactly what they needed.

The Opus 5 vs. Fable 5 decision#

The decision tree is straightforward:

• Use Opus 5 as your default for coding assistants, agent loops, and production API workloads. It is close enough to Fable on quality and half the price.
• Use Fable 5 when you need the absolute ceiling on novel reasoning, long-horizon agent tasks, or when you are working on problems where Opus 5's benchmarks show a gap.
• Use Sonnet 5 for cheap bulk work, classification, and high-throughput tasks where latency matters more than peak quality.
• Consider model routing if you have mixed workloads. Opus 5's effort levels (low, medium, high, xhigh, max) give you fine-grained control over cost vs. quality on a per-request basis.

Sources#

• Anthropic: Introducing Claude Opus 5 - official announcement with full benchmarks
• Anthropic: Claude Opus 5 System Card - 190-page safety and capability evaluation
• HN discussion: Claude Opus 5 - 746 comments and counting
• Artificial Analysis: Intelligence Leaderboard - independent benchmark aggregation
• Claude Opus 5 prompting guide - Anthropic's official optimization tips

Continue Reading#

• Claude Opus 4.8 Is an Agent Honesty Release - What the previous Opus generation shipped and how it changed agent trustworthiness
• Best Claude Model Now That Fable 5 Is Disabled - A decision guide for the post-Fable model landscape
• Frontier Model API Pricing June 2026 - How Opus 5 pricing compares against the full field of GPT, Gemini, and open-weight models
• Handling Fable 5 Refusals: A Guide to the Fallback API - Production patterns that now apply equally to Opus 5's classifier architecture
• Anthropic Model Naming Explained - Understanding where Opus fits in the Haiku-Sonnet-Opus-Fable-Mythos spectrum