On May 28, Anthropic released Claude Opus 4.8, available via API under the model name claude-opus-4-8. Pricing remains unchanged at $5 per million input tokens and $25 per million output tokens, matching the previous Opus 4.7.
But the official announcement highlights a different set of numbers.
It now admits when it's unsure
Anthropic claims Opus 4.8 is around four times less likely than its predecessor to allow flaws in code it has written to pass unremarked
. In plain terms, the model now proactively flags uncertain or potentially problematic code, rather than confidently delivering flawed output. This self-checking ability is arguably more impactful than benchmark improvements, as it saves developers the time of debugging confidently wrong code. The same logic extends to knowledge tasks: Opus 4.8 is more willing to indicate uncertainty and avoid unsupported assertions.
Benchmarks remain strong
Despite the honesty focus, Opus 4.8 delivers solid performance across key metrics:
- Online-Mind2Web (web navigation): 84%
- OSWorld-Verified (desktop navigation): 82.3%
- Legal Agent Benchmark: first model to exceed 10% under the
all tasks passed
criterion - CursorBench: surpasses previous Opus across all effort levels
- Finance Agent v2: outperforms Opus 4.7
The legal benchmark result is particularly noteworthy: all tasks passed
means every step in a task must be correct. While 10% may seem low, no previous model had crossed this threshold, highlighting the near-zero tolerance for error in such tasks.
Fast mode pricing slashed
Anthropic also revamped fast mode for Opus 4.8, offering 2.5x speed at $10 per million input tokens and $50 per million output tokens — three times cheaper than previous fast mode tiers. This reduces the premium for speed.
Additionally, Anthropic introduced:
- Dynamic workflows (research preview): launches hundreds of parallel sub-agents in Claude Code
- Effort slider: available in claude.ai and Cowork, allowing users to balance quality and speed
The parallel sub-agent capability, combined with the self-checking feature, points to a strategy of having multiple Claude instances monitor each other rather than working alone.
The real powerhouse is still coming
The announcement teases a more powerful model codenamed Mythos, currently limited to a few cybersecurity institutions. Anthropic says it needs to ensure safety before a wider release, expected in the coming weeks
. This pattern — releasing a stable flagship while hinting at a more advanced model — is familiar from Anthropic, with safety cited as the reason for delay.
Timing is strategic
The release comes as OpenAI prepares to file a confidential S-1 for its IPO, intensifying the rivalry between the two companies. By keeping Opus 4.8's price unchanged and reducing fast mode costs, Anthropic signals it can compete on both performance and value. Whether honesty
becomes the next battleground depends on how OpenAI and Google respond.
A model that says I'm not sure
may not sound exciting, but in practice, it could be more reassuring than a few extra benchmark points. The next-generation Mythos is expected in weeks.
Sources: Introducing Claude Opus 4.8 (Anthropic); Anthropic releases new model, CocoLoop, Opus 4.8 (Axios); Anthropic Debuts Claude Opus 4.8, Teases Upcoming Launch of Mythos-Class Models (Gizmodo)