Myth at 92.1%: The AI That Just Needs More Time
Give an AI agent four hours instead of thirty minutes and its benchmark score jumps ten points. That's the core message of Anthropic's quiet update to the Project-Glasswing page on April 13, 2026 — and it changes the entire discussion about what Claude Mythos Preview can actually do.
When Anthropic introduced Mythos Preview on April 7, the model achieved 82% on Terminal-Bench 2.0. Impressive, but not dominant. Six days later, with a longer timeout and a revised benchmark version, that number became 92.1%. The model didn't get smarter. It got more time.
This distinction is more important than most reporting concedes. For enterprise teams deciding on the use of AI agents, the difference between "this model isn't powerful enough" and "this model needs a different time budget" is the difference between project cancellation and delivery.
What Actually Changed: From 82% to 92.1%
The original launch of Mythos Preview on April 7, 2026, reported a score of 82% on Terminal-Bench 2.0 and 77.8% on SWE-bench Verified. The update on April 13 changed two variables simultaneously: the benchmark itself (2.0 to 2.1, fixing latency sensitivity) and the timeout (from thirty minutes to four hours).
The result: a jump from 82% to 92.1%. An improvement of 12.3 percentage points due to changed evaluation conditions, not a changed model.
Terminal-Bench 2.1: Why the Benchmark Update Matters
Terminal-Bench evaluates AI agents on real-world terminal tasks — debugging, infrastructure configuration, navigating complex codebases. The update from version 2.0 to 2.1 fixed a specific flaw: Tasks with fixed time limits systematically penalized models with higher inference latency.
A model that thought thoroughly before acting was rated the same as one that failed — both exceeded the timeout. Experienced engineers take different amounts of time for the same tasks. Limiting AI agents to thirty minutes while humans get unlimited time is not a fair comparison — it's a measurement error.
The Paradigm Shift in Compute Time
The Mythos result illustrates Test-Time Compute Scaling: Instead of building larger models, existing models are given more time to think. This changes the cost structure (operational expenditure instead of capital expenditure), makes quality adjustable (30 minutes for routine tasks, 4 hours for critical ones), and forces the updating of evaluation frameworks.
At Context Studios, we experience this dynamic regularly: an AI agent that seems to fail at a complex task is often successful if given a longer execution window. The ability was always there — the limitation was time, not intelligence.
What This Means for Enterprise AI Teams
The 92.1% result has immediate practical implications for AI Agent deployment:
Re-evaluate rejected tools. A model that failed at two minutes may succeed at twenty. Explicitly plan for compute time. Platforms like OpenClaw allow configurable timeouts per task. Adjust time budgets to task criticality. Security audits and code reviews deserve longer compute windows. Benchmark your own workflows. Run the same AI Agent with five different timeout values.
The eleven organizations with access via Project Glasswing — including government agencies — are likely already discovering that their initial assessments underestimated the model.
Why Most Teams Misjudge AI
AI Agents are not chatbots. They are autonomous workers, operating on task-level timescales. Evaluating an agent with a thirty-minute limit is like judging a junior developer solely on what they produce in their first half hour.
Three practices need to change: Use variable timeouts, separate capability from speed, and test on your own workload instead of relying on generic benchmarks.
Frequently Asked Questions
What is the actual Terminal-Bench score of Mythos Preview?
Mythos Preview achieved 92.1% on Terminal-Bench 2.1 with a four-hour timeout, compared to 82% on Terminal-Bench 2.0 with a thirty-minute timeout. Both figures are correct — they reflect different evaluation conditions.
Did Anthropic change the model between 82% and 92.1%?
No. The same Mythos Preview model produced both results. The difference resulted from the updated benchmark version and an increased timeout.
Can anyone access Claude Mythos Preview?
As of April 2026, Mythos Preview is limited to eleven organizations via Project Glasswing. There is no public API access.
What does this mean for teams using Claude Opus or Sonnet?
The compute-time scaling pattern applies generally. Teams using Claude Opus 4.6 or Sonnet 4.6 for agent tasks should experiment with longer timeouts.
How should companies adapt their AI evaluation process?
Test at multiple timeout values, separate capability metrics from speed metrics, and benchmark on your actual production workload.
Conclusion
The jump from 82% to 92.1% isn't a story about a model that got better. It's a story about an industry learning to measure capabilities more accurately. The model was always this capable. We just didn't give it enough time to show it.
The era of evaluating AI agents like chatbots is ending. The teams that adapt their evaluation frameworks first will find capabilities their competitors still dismiss as impossible.