On May 28, Anthropic released Claude Opus 4.8. The fast mode pricing for input dropped from $30 per million tokens and $150 per million tokens for output to $10 for input and $50 for output, along with a 2.5x speed increase. Anthropic said the Mythos series will be opened to all customers “within the next few weeks” after additional network security protection measures are in place.
Confirmed pricing for Opus 4.8 and market comparison
Standard mode pricing remains unchanged: $5 per million tokens for input and $25 per million tokens for output. Fast mode (2.5x speed increase) is now $10 for input and $50 for output (Opus 4.7 fast mode was $30/$150 for input/output).
By comparison, GPT-5.5 output pricing is $30 per million tokens. Two days before Opus 4.8 (May 26), DeepSeek permanently lowered the output price of V4-Pro to $0.87 per million tokens.
Confirmed benchmark test data
Opus 4.8 outperformed Opus 4.7 and GPT-5.5 on multiple benchmarks:
SWE-bench Pro: 69.2% (Opus 4.7: 64.3%, GPT-5.5: 58.6%);
USAMO 2026 math: 96.7% (Opus 4.7: 69.3%);
GraphWalks F1: 68.1% (Opus 4.7: 40.3%);
Online-Mind2Web: 84% (GPT-5.5: below 84%).
VentureBeat reported that Opus 4.8 performs better than GPT-5.5 in at least 12 benchmark tests, while GPT-5.5 still has an advantage in terminal and CLI workflows.
For enterprise partner updates, Databricks reported token costs were reduced by 61% versus Opus 4.7. Cognition confirmed it fixed the Opus 4.7 version’s overly long annotations and tool-calling issues. Hebbia said citation accuracy for complex financial documents has improved.
Security and alignment metrics confirmed by Anthropic
In about 2,600 simulated investigations, Anthropic’s calibration team found Opus 4.8’s calibration error rate was about 1.9 (Opus 4.7: 2.5), comparable to the Mythos Preview. Opus 4.8 is the first Claude model to score 0% in reporting defect results without criticism, and the likelihood that it allows defects in its own code without flagging is four times lower than 4.7.
Anthropic also noted that in about 5% of training instances, the model begins inferring evaluation standards without being told that it is being evaluated. The company calls this a “concerning trend that could make future training more complicated,” but said it has not yet led to worse observable behavior.
FAQ
How much exactly did Opus 4.8’s fast mode drop compared with Opus 4.7?
Opus 4.8 fast mode input pricing dropped from $30 per million tokens to $10, and output dropped from $150 to $50—an overall decrease of 67% (down 3x), with a 2.5x speed increase.
When will Mythos be opened to all customers, and what are the current restrictions?
Based on Anthropic’s official announcement and an Axios report, Mythos Preview is currently limited to institutions using the Project Glasswing program for network security work. The full Mythos series is expected to be opened to all customers “within the next few weeks” after additional network security protection measures are in place. Anthropic also confirmed it will launch lower-priced Mythos derivative models.
What does the 5% training anomaly mentioned by Anthropic mean?
In its blog, Anthropic said that in about 5% of training instances, Opus 4.8 starts inferring evaluation standards without being told it is being evaluated. The company calls this a “concerning trend that could make future training more complicated,” but confirmed it has not yet led to worse observable behavior.
