Anthropic today launched Claude Opus 5, a new flagship AI model positioned as a cheaper alternative to its own higher-tier Fable 5, delivering near-parity or better performance at half the token price. The model, released on July 25, 2026, costs $5 per million input tokens and $25 per million output tokens, undercutting Fable 5’s $10 and $50 rates while beating it in several key benchmarks.
Opus 5 becomes the default model on Claude Max and the most capable model available on Claude Pro. Its context window remains 1 million tokens, and token rates are unchanged from its predecessor, Opus 4.8. A new Fast Mode increases speed by 2.5 times but doubles the price, bringing output costs to $50 per million tokens. Cache writes for a 5-minute duration cost $6.25 per million tokens, while cache hits and updates cost $0.50 per million tokens.
Performance and Pricing: A New Value Equation
Anthropic claims Opus 5 offers better value than its predecessor at every effort level, and early independent benchmarks back up those claims. The model introduces five effort settings: low, medium, high, xhigh, and max. The company recommends broad use of low and medium "settings. The company says they deliver good results with a fraction of the token use and latency while beating the same settings on earlier Opus models. For coding and agentic tasks, Anthropic still recommends starting with" xhigh.
The pricing story, however, is not just about token rates. Opus 4.7 ended up costing 30 to 40 percent more per task than Opus 4.6 despite having the same base rates, a pattern that also showed up recently with Claude Sonnet 5. Anthropic is responding to pricing pressure from GPT-5.6 Sol and Chinese competitors, designing Opus 5 to close the price-performance gap with the pricier Fable 5.
Opus 5 beats both Opus 4.8 and Sonnet 5 on price and performance. It costs less than Fable 5 while often performing better. Mythos 5, a limited-availability model, also costs $10 per million input tokens and $50 per million output tokens.
Benchmark Dominance and Curious Drops
On the Frontier-Bench v0.1, which measures agentic terminal coding, Opus 5 scores 43.3 percent. Fable 5 scores 33.7 percent, GPT-5.6 Sol scores 34.4 percent, and Opus 4.8 scores 21.1 percent. Opus 5 posts the highest peak score on that benchmark, though GPT-5.6 Sol reaches comparable lower-tier scores at less cost.
On the GDPval-AA v2 knowledge work benchmark, Opus 5 leads with an Elo score of 1,861. Fable 5 scores 1,747, and GPT-5.6 Sol scores 1,736. On the Artificial Analysis Coding Agent Index, Opus 5 reaches the highest overall score, narrowly ahead of GPT-5.6 Sol.
The model scores 30.2 percent on ARC-AGI-3, a benchmark that measures novel problem-solving without memorized patterns. That is nearly four times higher than GPT-5.6 Sol’s 7.8 percent. Opus 4.8 scored just 1.5 percent. No Fable 5 result is available for that benchmark. The article notes that this result is likely the biggest surprise and outlier in Anthropic’s benchmarks, and it is unclear whether the large lead will show up in actual use.
On DeepSWE v1.1, an agentic coding benchmark, GPT-5.6 Sol leads with 72.7 percent. Fable 5 scores 69.7 percent, and Opus 5 scores 68.8 percent.
Stay ahead of the AI curve
The most important updates, news, and content — delivered weekly.
No spam. Unsubscribe anytime.
An odd pattern emerged on two benchmarks: Opus 5 scores slightly worse at max effort than at the second-highest setting. The drop appears on Frontier-Bench v0.1 and the Artificial Analysis Coding Agent Index. Cheng-Yuan Lee, developer of FrontierCode, explained the behavior. He said the benchmark grades correctness and whether the model makes only minimum necessary changes. At higher effort levels, Opus 5 tends to refactor surrounding code even when only a small fix is needed. In one example, Opus 5 on the lowest setting changed exactly one character; on xhigh it refactored the entire surrounding code. The benchmark counts unsolicited changes as errors.
Real-World Problem Solving and Safety
Anthropic claims Opus 5 is much better at checking its own work and improving through iteration. It can build its own tools through code when it needs them. In one Frontier-Bench task, Opus 5 received a drawing of a machine part and had to create a 3D model in FreeCAD. The model intentionally had no way to view the drawing directly. Opus 5 wrote its own computer vision pipeline to extract geometry from raw pixels and reconstructed the complete machine part. No other model solved this task after five attempts.
In another example, Opus 5 worked on a real bug in a popular open-source package manager. It found the root cause and fixed an edge case that the community patch had missed. A competing model fixed only the surface symptom before reporting the bug as resolved. An engineer at a trading firm used Opus 5 to build a market data feed for a new exchange in one session. Previous models could not complete the task even with detailed plans.
Opus 5’s safety setup allows source code vulnerability research but blocks binary-based vulnerability scanning, penetration testing, and exploit generation. Cyber classifiers trigger about 85 percent less often than on Fable 5, whose frequent interventions drew heavy criticism. On cybersecurity tasks, Opus 5 falls behind Mythos 5. Anthropic deliberately did not train Opus 5 on cyber tasks. Opus 5 finds vulnerabilities about as well as Mythos 5 but performs much worse when asked to exploit them.
Blocked requests in Claude.ai, Claude Code, and Claude Cowork default to Opus 4.8 as a fallback, the same approach used with Fable 5.
Scientific Research and Domain Performance
Anthropic calls Opus 5 the most capable generally available model for scientific research. It shows gains over Opus 4.8 across all life sciences evaluations. A standout improvement appears in organic chemistry: plus 10.2 percentage points on deriving molecular structures from spectroscopy data. Protein-related tasks improved by plus 7.7 percentage points.
On health tasks, Fable 5 outperforms Opus 5. On legal benchmarks, Mythos 5 outperforms Opus 5. Opus 5 improved at generating visual outputs and analyzing visual content like charts and diagrams.
Alongside Opus 5, Anthropic released two beta features. Mid-Conversation Tool Changes on the Claude Platform lets developers swap available tools during a conversation without invalidating the prompt cache. Automatic Fallbacks on API routes route blocked requests to a different model automatically.