Claude Opus 5: Benchmarks, Pricing, and How It Compares to Fable 5

Anthropic released Claude Opus 5 on July 24, 2026, at $5 per million input tokens and $25 per million output tokens, the same rate as Opus 4.8 and exactly half the price of Fable 5. The company's positioning is straightforward: near-frontier intelligence for the price you were already paying. The API model ID is claude-opus-5, the context window is 1M tokens, and it's the new default on Claude Max and the strongest model on Claude Pro.

This article covers what the benchmarks actually show, where Opus 5 leads and where it doesn't, and how it performs on two real-world coding tasks against Fable 5, GPT-5.6 Sol, and Kimi K3.

What Anthropic says about Opus 5

Anthropic describes Opus 5 as "a thoughtful and proactive model that comes close to the frontier intelligence of Claude Fable 5 at half the price," with state-of-the-art performance on coding and knowledge work benchmarks including Frontier-Bench and GDPVal-AA. They note it "remains behind Mythos 5 on cybersecurity tasks," which is consistent with the pattern across the Claude 5 family.

One practically significant detail: Opus 5 does not have data retention requirements for general access. Unlike Fable 5, which carries retention requirements that restricted its use on certain confidential benchmarks, Opus 5 can be tested and deployed more freely in privacy-sensitive contexts. For teams handling proprietary data, that's a meaningful operational difference.

Thinking is on by default in Opus 5. Disabling it above the high effort setting returns a 400 error. A fast mode is available at approximately 2.5x the default speed for $10/$50 per million tokens, accessible on the Claude Platform and through usage credits in Claude Code.

Benchmark results

Anthropic's internal benchmarks

On Frontier-Bench v0.1, the agentic terminal coding benchmark, Opus 5 scores 43.3%, nearly 10 points ahead of Fable 5 at 33.7% and more than double Opus 4.8's 21.1%. That's a substantial coding improvement over both predecessors.

The ARC-AGI-3 result is harder to ignore. Opus 4.8 scored 1.5% on this novel problem-solving benchmark. GPT-5.6 Sol held the previous high at 7.8%. Opus 5 scores 30.2%, independently verified by the ARC Prize Foundation at 30.16%.

The ARC Prize Foundation noted that Opus 5 used "advanced logical reasoning to turn ARC-AGI-3 layouts into algebraic notation," meaning it developed its own mathematical representation of the problem rather than pattern-matching directly. That's a qualitatively different kind of performance than the models that came before it on that benchmark.

Anthropic also ran Opus 5 on all six IMO 2026 problems without tools or an agent harness. A three-model judge panel scored all 24 generated solutions correct, with human experts independently grading one pre-specified solution per problem at 7/7. The 42/42 total is gold-medal level, above the 29/42 cutoff.

Artificial Analysis independent benchmarks

On the Artificial Analysis Coding Index, Opus 5 places second, 0.3 points behind GPT-5.6 Sol and 1.5 points ahead of Fable 5. On the Agentic Index, Opus 5 takes first place, ahead of both Fable 5 and GPT-5.6 Sol. On the Intelligence Index, it also leads, scoring 61 against Fable 5 at 60 and GPT-5.6 Sol at 59.

The independent data broadly corroborates Anthropic's claims. Opus 5 is genuinely at or near the top of the field on the benchmarks that matter most for developer workflows.

Pricing and cost in practice

The headline pricing comparison is clean: Opus 5 at $5/$25 per million tokens versus Fable 5 at $10/$50. Same input cost as GPT-5.6 Sol ($5), slightly cheaper output ($25 vs $30).

One nuance worth noting: the "half the price" claim is accurate on a per-token basis, but Opus 5 can be more verbose than Fable 5 on complex tasks, using more tokens to produce more detailed responses. In the coding tests below, the total API cost for Opus 5 came out close to Fable 5's on a per-task basis because of this. On average across benchmark runs, the cost-per-task advantage is real, but for any specific task, the verbosity can narrow the gap.

Building a 3D Formula 1 racing game

The first coding test: generate a complete, browser-based F1 racing game as a single self-contained HTML file using Three.js from a CDN. The prompt specified a drivable player car with chase-camera and responsive physics, at least three AI opponents, a lap counter and timer, engine sounds via the Web Audio API, barriers and trackside detail, a minimap, and a HUD, all running at 60fps with no build step.

Opus 5 produced a visually detailed and functional game. The car models were polished, the lap counter and minimap worked correctly, and the driving physics felt responsive. Two minor issues: the track rendered in the wrong direction and the car could drive under the grass. Neither broke the core experience.

Fable 5 also produced a solid result. The track layout was well-designed and a camera-wobble effect added some immersion. Car models were noticeably more basic compared to Opus 5's output.

GPT-5.6 Sol struggled significantly. Car models were reasonable, but there was no racetrack, just a flat green plane. The game was unplayable.

Kimi K3 produced a result comparable to GPT-5.6 Sol: functional cars, no complete track environment.

Both Anthropic models were clearly in a different tier for this kind of complex creative single-shot generation.

Building a full-stack personal finance dashboard

The second test: build a personal finance management dashboard with account tracking across savings, checking, investments, and debts; transaction history; financial goals; fund transfer functionality; and seeded backend data. No authentication required.

Opus 5 produced a clean, modern dark-mode application with properly structured frontend and backend. The UI organized content across separate pages: Overview, Accounts, Transactions, Goals. All features were functional. Tech stack: React with React Router on the frontend, Node.js/Express with SQLite on the backend. Straightforward, sensible choices.

Fable 5 produced a functional application with a less polished UI and some less useful data visualizations. It built its own rudimentary routing library from scratch instead of using React Router, an unnecessary choice. Backend was similar: Node.js with SQLite.

GPT-5.6 Sol produced a visually clean light-mode interface. All content was on a single long-scrolling page rather than separate views. Tech stack was the most ambitious: Next.js with Drizzle ORM configured for Cloudflare Workers deployment. More complex than the task required.

Kimi K3 produced a decent UI that looked somewhat dated. The major weakness was the backend: it used a flat JSON file for storage rather than a proper database, which limits scalability and reliability.

Opus 5 produced the most complete, well-architected result across both tests.

Safeguards

Anthropic describes Opus 5's cyber classifiers as expected to "intervene around 85% less often" than Fable 5's. This makes Opus 5 more permissive for legitimate security research, such as finding vulnerabilities in source code, while still blocking overtly offensive activities like binary-based vulnerability scanning and exploit generation. The restrictions are described as proportionally similar to Opus 4.8.

Where Opus 5 fits

Opus 5 makes Fable 5 harder to justify for most everyday workloads. It matches or beats Fable 5 on the benchmarks that matter most for coding and agentic tasks, costs half as much on a per-token basis, and carries no data retention requirements. The cases where Fable 5 still has an edge are narrow: primarily Mythos-class cybersecurity tasks and specific scenarios where Fable 5's particular strengths are needed.

For developers who were already paying Opus 4.8 prices, Opus 5 is a straightforward upgrade: same pricing, substantially better performance across coding, reasoning, and agentic benchmarks.