Anthropic launches cheaper Claude Opus 5 model
Anthropic has launched Claude Opus 5, a cheaper and more capable vision-language model that resolves key user complaints about Claude Fable 5 while setting new benchmark records.
Anthropic has launched Claude Opus 5, a vision-language model that processes up to 1 million input tokens and 128,000 output tokens at 52.8 tokens per second. It features five reasoning levels (low, medium, high, xhigh, max; defaulting to high), prompt caching from 512 tokens, and a fast mode at 2.5 times standard speed. It is the default for Claude Max subscribers ($100-200 per month) and available to Claude Pro subscribers ($20 per month). API pricing is $5/$0.50/$25 per million input/cached/output tokens, rising to $10/$1/$50 in fast mode.
On Artificial Analysis Intelligence Index, Claude Opus 5 at max reasoning scored 61 points ($2.03 per task), beating Claude Fable 5 at max reasoning with fallback (60 points, $2.75) and GPT-5.6 Sol at max reasoning (59 points, $1.54). On the Coding Agent Index, Opus 5 at xhigh reasoning tied GPT-5.6 Sol at max reasoning with a score of 67. It also led on GPTval-AA v2, AA-Briefcase, and MMMU-Pro. On ARC-AGI-3, Opus 5 at high reasoning scored 30.2 percent ($20.7k to run), beating GPT-5.6 Sol at max reasoning (7.8 percent, $25.1k). It resolved 43.5 percent on Terminal-Bench 3.0 and 26.2 percent on AutomationBench. On CursorBench 3.2, Opus 5 scored 70 percent ($8.23 per task) compared to Fable 5's 70.5 percent ($17.32).
The model eliminates the 30 day data retention policy and permits biology and chemistry queries that previously triggered fallbacks to Claude Opus 4.8. In geopolitical news, Michael Kratsios claimed on July 22 that Moonshot AI distilled Claude Fable 5 to build Kimi K3, which ranks fourth on the Intelligence Index. Treasury Secretary Scott Bessent threatened sanctions, though Braden Hancock of the Laude Institute disputed the timeline.
For practitioners, Claude Opus 5 is a highly capable workhorse. Because it excels at self-checking, developers should omit legacy verification instructions to avoid over-verification. While Fable 5 remains better for factual recall, Opus 5 offers a cheaper, highly effective alternative for complex workflows.
This is our own summary of reporting by The Batch
