Claude Opus 5 benchmarks beat Fable 5 at half the input price
Anthropic released Claude Opus 5 on July 24, pricing it below Fable 5 while posting stronger scores on key AI tests.
By Sofia Marchetti · Columnist
· 3 min read
Anthropic’s Claude Opus 5 benchmarks give developers and AI-heavy businesses a clear cost-performance comparison: the new model scored ahead of Claude Fable 5 on several major tests while costing half as much per million input tokens. For users paying for AI tools, that matters because model pricing can flow through to app costs, usage limits, and which features companies can afford to offer.
Anthropic released Claude Opus 5 on July 24. The company priced it at $5 per million input tokens, matching Claude Opus 4.8 and coming in at half the price of Claude Fable 5, according to Anthropic’s release details. An input token is a small unit of text sent into an AI model, so lower input pricing makes repeated prompts, long documents, and software-agent workflows cheaper to run.
How did Claude Opus 5 compare with Fable 5?
On Frontier-Bench v0.1, Anthropic said Claude Opus 5 scored 43.3%, compared with 33.7% for Claude Fable 5 and 34.4% for OpenAI’s GPT-5.6 Sol. Frontier-Bench v0.1 measures whether coding agents can finish real software engineering tasks from start to finish, with the score showing the share of tasks passed.
Claude Opus 5 also posted a large lead on ARC-AGI-3, according to Anthropic’s benchmark figures. The model scored 30.2%, while OpenAI’s GPT-5.6 Sol scored 7.8%. Fable 5 was not tested on that benchmark in the figures cited. ARC-AGI-3 is designed around new problem-solving puzzles, which are meant to reduce the chance that a model is relying on memorized training examples.
On GDPval-AA v2, a benchmark for knowledge-work tasks that uses Elo ratings, Anthropic reported Claude Opus 5 at 1,861. Fable 5 scored 1,747, while GPT-5.6 Sol scored 1,736. Elo is a ranking system, commonly associated with chess, that compares relative performance rather than giving a simple percentage score.
What is Claude Opus 5?
Claude Opus 5 sits in Anthropic’s Opus tier, which the company positions as its heavier-duty class of models. Anthropic’s lineup also includes Haiku for faster, lower-cost tasks, Sonnet as a middle tier, and a higher Mythos class that includes Claude Fable 5 for public users.
Anthropic introduced the Mythos tier this spring. The same product family includes Claude Mythos 5, a version with fewer restrictions made available through Project Glasswing to vetted cybersecurity researchers and critical infrastructure operators, according to the company’s described model structure.
The rollout also changes where Fable 5 fits for subscribers. Fable 5 launched June 9, was pulled globally three days later after a U.S. government emergency export control order cited a jailbreak vulnerability, and returned June 30. After its return, Fable 5 moved to a credits-only setup rather than being included in standard plans.
Claude Opus 5 is now the default model for Claude Max and the strongest option on Claude Pro, according to Anthropic’s release positioning. That makes Opus 5 the practical flagship for many paying Claude users, even though Fable 5 remains part of the higher Mythos class.
Lovable, a developer platform with millions of users, also tested Opus 5 internally. Fabian Hedin said in a statement shared by Anthropic that the model was up 22% over Opus 4.7 on Lovable’s hardest agentic coding tasks and showed less run-to-run variation.
This story draws on original reporting from Decrypt.
