Claude Opus 4.8 cuts error omission to 3.7% while Mythos model arrives in weeks

Anthropic has officially released Claude Opus 4.8, a model update that secures first place in five of six core benchmarks while maintaining the existing pricing structure of $5 per million input tokens and $25 per million output tokens. The release marks a strategic departure from pure performance chasing, as the company explicitly positions 'trustworthiness' as the primary value proposition for enterprise adoption. Data compiled by Woofun AI shows that in code summarization honesty tests, the model's failure to flag its own errors plummeted from 19.7% in version 4.7 to just 3.7% in 4.8, representing a roughly fivefold improvement in self-correction capabilities. This metric is critical because it addresses the specific risk of 'silent errors,' where a model provides a polished but incorrect answer that users fail to detect due to its apparent completeness.

Beyond individual model accuracy, Opus 4.8 introduces dynamic workflows within Claude Code, enabling the autonomous orchestration of dozens to hundreds of child agents in a single session. These sub-agents operate in parallel and include adversarial agents designed to rebut results before they are presented to the user, effectively automating the 'Agent Teams' concept introduced in version 4.6. In due diligence tests, the model achieved a literal zero in 'error reporting flawed results' and eliminated 'lazy investigations,' reducing overconfident wrong answers by approximately 11 times. Woofun AI notes that this architectural shift allows the system to expose its own uncertainty, thereby reducing the cognitive load on human operators who must otherwise verify every output manually.

Despite these gains, the model does not dominate every category. In the Terminal-Bench 2.1 test, which evaluates long-horizon agent tasks via terminal operations, GPT-5.5 retains the lead with a score of 78.2% compared to Opus 4.8's 74.6%. Anthropic acknowledged this deficit transparently on the release card, highlighting a persistent divide where GPT-5.5 functions as a superior terminal operator while Opus 4.8 excels as a professional engineer in coding, reasoning, and knowledge work.

Furthermore, the 244-page system card reveals honest regressions not featured in the marketing materials, including a slight dip in the GPQA Diamond expert science test from 94.2% to 93.6% and reduced resistance to prompt injections in computer use scenarios.

The model demonstrates significant leaps in specific high-value domains, particularly in mathematics and long-context reasoning. In the USAMO 2026 competition held in March, Opus 4.8 scored 96.7%, a 27-point increase from the 69.3% achieved by version 4.7, with no data contamination issues since the competition occurred post-training cutoff. In million-token image reasoning tests, it scored 68.1, substantially outperforming both its predecessor at 40.3 and GPT-5.5 at 45.4. Woofun AI analysis suggests that these improvements in long-context handling are pivotal for complex enterprise tasks where maintaining coherence over vast datasets is more valuable than raw speed.

A critical paradigm shift is evident in token efficiency and multi-agent performance. In the most challenging coding tests, Opus 4.8 operating at its lowest effort setting outperforms version 4.7 at its highest effort setting, allowing users to achieve peak performance with significantly lower token costs. In multi-agent web research tasks, a single agent scores 84.3%, lagging slightly behind Gemini's 85.9%, but a conductor-scheduled team of sub-agents raises the score to 88.5%, achieving the top-reported result in one-fifth of the time. This capability underscores the transition from single-model reliance to orchestrated agent ecosystems.

Looking ahead, Anthropic has confirmed that the Mythos-class model, currently restricted to 52 audited institutions, will become available in the coming weeks. Mythos represents a controlled cutting-edge layer that surpasses Opus 4.8 in nearly all metrics, including a 93.9% score on SWE-bench Verified and the ability to generate executable exploits against browser targets, a task where Opus 4.8 succeeds less than 10% of the time. This creates a two-tier market structure: a commodified, widely accessible Opus 4.8 competing with open-source models, and an expensive, restricted Mythos layer serving as infrastructure for high-stakes operations. The rapid release cadence, moving from Opus 4.6 in February to 4.7 in April and now 4.8, indicates that the industry is moving toward a future where model migration and rigorous validation are more valuable than mastering a single static version.