Claude Opus 4.8 Debuts as Anthropic Pushes for ‘Improved Honesty’

Anthropic says its latest AI model is better at admitting when it might be wrong.

The artificial intelligence lab on Thursday released Claude Opus 4.8, an upgraded version of its leading generally available model. According to the company, the release delivers measurable improvements across coding, reasoning, and practical knowledge work benchmarks while keeping standard costs identical to its predecessor, Opus 4.7.

A primary focus of the Opus 4.8 rollout centers on model reliability and what Anthropic describes as improved honesty. The company noted that AI models frequently "jump to conclusions", confidently claiming to have made progress in their work despite the evidence being thin."

To combat this, the new model was trained to more readily surface gaps in its own confidence and avoid making unsupported assertions. Internal evaluations indicate that Opus 4.8 is roughly 4x less likely than Opus 4.7 to allow flaws in code it has written to pass unremarked.

Early testers have reported a greater tendency for the system to flag uncertainties about its work rather than projecting unearned confidence.

The upgrade also registered across-the-board gains on published benchmarks:

• Agentic Coding (Terminal-Bench 2.1): Rose from 64.3% to 69.2%.
• Multidisciplinary Reasoning with Tools: Improved from 54.7% to 57.9%.
• Agentic Computer Use: Shifted upward from 82.8% to 83.4%.
• Knowledge Work: Scores increased from 1,753 to 1,890.

Anthropic's alignment team said the model "reaches new highs on our measures of prosocial traits like supporting user autonomy and acting in the user’s best interest," lowering rates of deceptive or misaligned behaviors down to levels near its restricted preview models.

New controls and parallel subagents

Launching alongside the model are several tool updates aimed at giving users more command over operational costs and computing power.

A newly introduced effort control panel on claude.ai and Cowork allows users to manually adjust the intensity of Claude's processing. Selecting higher effort configurations prompts the model to think more frequently and deeply for complex assignments, consuming more tokens. Lower effort settings accelerate response times and preserve rate limits for simpler tasks.

For software development, Anthropic introduced a research preview feature called "dynamic workflows" in Claude Code. The functionality allows the system to organize, plan, and execute massive projects by deploying hundreds of parallel subagents in a single session. 

The company claims this allows Claude to execute codebase-scale migrations spanning hundreds of thousands of lines of code from kickoff to merge, using existing test suites as the benchmark for validation.

For developers using the Messages API, Anthropic added the capability to accept system entries inside the messages array. This allows engineers to adjust instructions, environment contexts, or token budgets mid-task without interrupting the prompt cache or waiting for a user turn.

Mythos looms in the background

Even as Anthropic rolls out Opus 4.8, the company is already pointing toward a more advanced family of AI systems.

Anthropic said it plans to release “a new class of model with even higher intelligence than Opus” in the coming weeks. Those systems are tied to Project Glasswing and Claude Mythos Preview, which are currently being tested by a small number of organizations for cybersecurity work. 

According to Anthropic, Mythos-class models require stronger safeguards before wider deployment because of their advanced cybersecurity capabilities.

Also read: Google’s Gemini 3.5 Flash and Antigravity 2.0 add another enterprise coding challenge for Anthropic and OpenAI.