Claude Opus 5 matches Fable 5 performance at half the token price

Anthropic announced Claude Opus 5, a new flagship model designed to deliver near-Fable 5 performance at substantially lower cost. Opus 5 is priced at $5 per million input tokens and $25 per million output tokens—exactly half the cost of Fable 5 ($10 and $50 respectively)—while maintaining the same 1 million-token context window as its predecessor, Opus 4.8. The model becomes the default on Claude Max and the most capable option on Claude Pro. A new Fast Mode feature increases speed by 2.5x but doubles the price to $10 per million input tokens and $50 per million output tokens.

The release arrives amid competitive pressure from GPT-5.6 Sol and Chinese competitors. The body notes that pricing alone tells an incomplete story: token efficiency—the actual tokens consumed to complete a task—varies significantly across models and effort levels. Users can now trade performance against token consumption through five effort settings: low, medium, high, xhigh, and max. Anthropic recommends broad use of low and medium settings for general tasks, claiming they deliver good results with a fraction of the token use and latency compared to earlier Opus models, while recommending xhigh specifically for coding and agentic work. Anthropic reports that Opus 5 offers better value than its predecessor at every effort level, though the body notes an anomaly: on two benchmarks (Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index), Opus 5 scores slightly worse at the max effort setting than at the second-highest setting, despite the higher cost.

Opus 5's benchmark performance demonstrates strong positioning in coding and knowledge work. On Frontier-Bench v0.1, Opus 5 reaches 43.3% on agentic terminal coding, outpacing Fable 5 (33.7%), GPT-5.6 Sol (34.4%), and its predecessor Opus 4.8 (21.1%) by wide margins. On the knowledge work benchmark GDPval-AA v2, Opus 5 achieves an Elo score of 1,861, ahead of Fable 5 (1,747) and GPT-5.6 Sol (1,736). The most striking result comes on ARC-AGI-3, which measures novel problem-solving without memorized patterns: Opus 5 scores 30.2%, nearly four times higher than GPT-5.6 Sol (7.8%) and far ahead of Opus 4.8 (1.5%). The body flags this as likely the biggest surprise and notes uncertainty about whether such a large lead will translate to actual use. However, Opus 5 does not win on all fronts. On agentic coding via DeepSWE v1.1, GPT-5.6 Sol leads with 72.7%, followed by Fable 5 (69.7%) and Opus 5 (68.8%). Fable 5 and Mythos 5 outperform Opus 5 on health and legal tasks respectively. Anthropic deliberately did not train Opus 5 on cybersecurity tasks, and it falls behind Mythos 5 on exploit development while performing comparably on vulnerability discovery. Anthropic also reports improvements in visual output generation and visual content analysis.

Anthropoc highlights Opus 5's ability to build its own tools and check its own work through iteration. In one Frontier-Bench task, Opus 5 received a drawing of a machine part and was asked to create a 3D model in FreeCAD, but the interface intentionally blocked direct viewing of the drawing. Opus 5 responded by writing its own computer vision pipeline to extract geometry from raw pixels and reconstructing the complete machine part—a task no other model solved after five attempts. In another example, Opus 5 identified the root cause of a bug in a popular open-source package manager and fixed an edge case the community patch had missed. A trading firm engineer reportedly used Opus 5 to build a market data feed for a new exchange in a single session, a task previous models could not complete.

Regarding safety, Opus 5's cyber filters allow source code vulnerability research but block binary-based vulnerability scanning, penetration testing, and exploit generation. Anthropic reports that its cyber classifiers trigger roughly 85 percent less often than on Fable 5, responding to criticism of Fable 5's frequent interventions. Blocked requests in Claude.ai, Claude Code, and Claude Cowork default to Opus 4.8 as a fallback. Anthropic describes Opus 5 as the most capable generally available model for scientific research, showing gains across all life sciences evaluations—with particular improvements in organic chemistry (plus 10.2 percentage points on deriving molecular structures from spectroscopy data) and protein-related tasks (plus 7.7 percentage points). The release includes two beta features: Mid-Conversation Tool Changes, which allows developers to swap available tools during a conversation without invalidating the prompt cache, and Automatic Fallbacks on the API, which routes blocked requests to a different model automatically.

Anthropic's release of Claude Opus 5 represents a direct response to competitive and pricing pressure in the large language model market. The body notes that Anthropic is specifically responding to pricing pressure from GPT-5.6 Sol and Chinese competitors, and designed Opus 5 to close the price-performance gap with Fable 5, which commands a significantly higher cost per token. By pricing Opus 5 at exactly half the input and output token rates of Fable 5 while delivering equal or superior performance on multiple benchmarks, Anthropic is repositioning its product lineup to offer better value without sacrificing capability.

However, the article emphasizes that raw token rates mask a more complex cost story. Token efficiency—the actual number of tokens consumed to complete a task—varies substantially across models and effort levels. The body notes that Opus 4.7 ended up costing 30 to 40 percent more per task than Opus 4.6 despite identical base rates, and a similar pattern appeared with Claude Sonnet 5. This means users cannot assume that Opus 5's lower per-token price automatically translates to lower per-task cost. The five effort levels (low, medium, high, xhigh, max) allow users to make explicit performance-versus-efficiency trade-offs; Anthropic reports that Opus 5 achieves better value at every effort level compared to its predecessor, and recommends users start with low or medium for general work and xhigh for specialized coding tasks.

Opus 5's benchmark performance tells a nuanced story: it leads decisively on coding and novel problem-solving (the ARC-AGI-3 result of 30.2%—nearly 4× higher than GPT-5.6 Sol—is flagged by the body as a notable outlier with unclear real-world implications), and surpasses Fable 5 on knowledge work. However, the article makes clear that Opus 5 does not win everywhere—Fable 5 and Mythos 5 retain advantages in health and legal domains, and GPT-5.6 Sol leads on the DeepSWE v1.1 coding benchmark. The release also includes technical improvements: Opus 5 can now build its own tools to solve problems (the body describes it writing a computer vision pipeline when direct image viewing was blocked), and its safety classifiers trigger roughly 85 percent less often than Fable 5's, a response to Fable 5's heavy criticism for excessive interventions.

No comments yet. Be the first to share your thoughts!