Mythos 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 is the headline from Anthropic's quiet update to the Project Glasswing page on April 13, 2026 — and it reframes the entire conversation about what Claude Mythos Preview can actually do.
When Anthropic first announced Mythos Preview on April 7, the model scored 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 did not get smarter. It got more time.
This distinction matters more than most coverage acknowledges. For enterprise teams deciding how to deploy AI agents, the difference between "this model is not capable enough" and "this model needs a different time budget" is the difference between abandoning a project and shipping it.
What Actually Changed: From 82% to 92.1%
The original Mythos Preview launch on April 7, 2026 reported an 82% score on Terminal-Bench 2.0 and 77.8% on SWE-bench Verified. These numbers positioned Mythos as competitive but not clearly ahead of existing models on agent benchmarks.
The April 13 update changed two variables simultaneously. First, the benchmark itself was updated from Terminal-Bench 2.0 to Terminal-Bench 2.1. According to Anthropic's system card, Terminal-Bench 2.0 was "sensitive to inference latency" — meaning the benchmark's own timing mechanisms were penalizing models that thought longer before acting. The 2.1 update from the benchmark maintainers fixed this measurement artifact.
Second, Anthropic increased the timeout from thirty minutes to four hours. This is not a minor tweak. It is an eightfold increase in the compute budget available for each task.
The result: a jump from 82% to 92.1%. That is a 10.1 percentage-point improvement from changing the evaluation conditions, not the model.
Terminal-Bench 2.1: Why the Benchmark Update Matters
Terminal-Bench evaluates AI agents on real-world terminal tasks — the kind of work that software engineers do daily. Debugging production systems, configuring infrastructure, navigating complex codebases. Unlike benchmarks that test isolated reasoning, Terminal-Bench measures whether an agent can actually get things done.
The version 2.0 to 2.1 update addressed a specific flaw: tasks with fixed wall-clock timeouts were penalizing models with higher inference latency. A model that paused to reason deeply before acting was graded identically to a model that failed — both hit the timeout. This created a systematic bias against deliberate, multi-step reasoning.
For context, expert human engineers complete Terminal-Bench tasks in varied timeframes. Some take minutes; others take hours. Constraining AI agents to thirty minutes while allowing humans unlimited time is not a fair comparison — it is a measurement error.
The 2.1 fix acknowledged this reality. And the impact on Mythos Preview's score was dramatic.
The Compute-Time Paradigm Shift
The Mythos result illustrates a broader pattern emerging across AI research: test-time compute scaling. The idea is straightforward — instead of building bigger models (more parameters, more training data), you give existing models more time to think during inference.
This matters for three reasons:
Cost structure changes. Training a larger model costs millions upfront. Giving an existing model more inference time costs proportionally per task. For enterprises, this shifts AI spending from capital expenditure to operational expenditure — a fundamentally different budget conversation.
Quality becomes adjustable. A team can run the same model at thirty minutes for routine tasks and four hours for critical ones. This is analogous to how engineering teams assign different review depths to different pull requests. Not every task needs maximum compute.
Evaluation frameworks need updating. If a model scores 82% at thirty minutes and 92.1% at four hours, which number matters? The answer depends entirely on how you plan to use it. Teams evaluating AI agents with fixed short timeouts are systematically underestimating model capability.
At Context Studios, we see this dynamic play out in client projects regularly. An AI agent that seems to fail on a complex task often succeeds when given a longer execution window. The capability was always there — the constraint was time, not intelligence.
What This Means for Enterprise AI Teams
The Mythos 92.1% result has immediate practical implications for how organizations should approach AI agent deployment:
Re-evaluate rejected tools. If your team tested an AI agent and dismissed it as "not accurate enough," check the timeout configuration. A model that failed at two minutes may succeed at twenty.
Budget compute time explicitly. Agent platforms like OpenClaw and similar frameworks allow configurable timeouts per task. Start treating inference time as a first-class resource, like CPU or memory.
Match time budgets to task criticality. Security audits, code reviews, and architecture decisions deserve longer compute windows than formatting fixes or log analysis. This is not about spending more — it is about spending proportionally.
Benchmark your own workflows. Run the same AI agent on the same task at five different timeout values (1 minute, 5 minutes, 15 minutes, 30 minutes, 2 hours). Plot the accuracy curve. Most teams have never done this, and the results are often surprising.
The eleven organizations with access to Mythos Preview through Project Glasswing — including cybersecurity firms and government agencies — are likely already discovering that their initial evaluations underestimated the model by giving it too little time.
Why Most Teams Are Evaluating AI Wrong
The Mythos score revision exposes a systemic problem in how the industry evaluates AI agents. Most evaluation frameworks use fixed, short timeouts because they were designed for chat-style interactions — where a user expects a response in seconds, not hours.
But AI agents are not chatbots. They are autonomous workers that operate on task timescales, not conversation timescales. Evaluating an agent with a thirty-minute cap is like evaluating a junior developer by only measuring what they produce in their first half hour. You would miss the work that requires deep understanding.
Three evaluation practices need to change:
-
Use variable timeouts. Report scores at multiple time budgets, not just one. The relationship between time and accuracy is the most valuable signal.
-
Separate capability from speed. A model that solves 92% of problems in four hours is more capable than one that solves 75% in thirty minutes — even if the faster model is more practical for certain use cases.
-
Test on your actual workload. Generic benchmarks like Terminal-Bench provide directional signal, but the only benchmark that matters is your own production data.
Frequently Asked Questions
What is the actual Mythos Preview Terminal-Bench score?
Mythos Preview scored 92.1% on Terminal-Bench 2.1 with a four-hour timeout, up from 82% on Terminal-Bench 2.0 with a thirty-minute timeout. Both numbers are accurate — they reflect different evaluation conditions, not different models.
Did Anthropic change the model between 82% and 92.1%?
No. The same Mythos Preview model produced both scores. The difference came from two changes: an updated benchmark version (2.0 to 2.1) that fixed latency-related measurement issues, and an increased timeout (30 minutes to 4 hours).
Can anyone access Claude Mythos Preview?
As of April 2026, Mythos Preview is restricted to eleven organizations through Project Glasswing, which focuses on cybersecurity applications. There is no public API access or pricing page. Anthropic has not announced a general availability timeline.
What does this mean for teams using Claude Opus or Sonnet?
The compute-time scaling pattern applies broadly, not just to Mythos. Teams running Claude Opus 4.6 or Sonnet 4.6 for agent tasks should experiment with longer timeouts — the same model often performs significantly better with more time to reason through complex problems.
How should enterprises adjust their AI evaluation process?
Test at multiple timeout values, separate capability metrics from speed metrics, and benchmark on your actual production workload rather than relying solely on public benchmark scores.
The Bottom Line
Mythos Preview's jump from 82% to 92.1% is not a story about a model getting better. It is a story about an industry learning to measure capability more accurately. The model was always this capable. We were just not giving it enough time to show it.
For AI teams, the actionable takeaway is concrete: before concluding that a model cannot handle your use case, increase the timeout by 4-8x and test again. The results may change your entire deployment strategy.
The era of evaluating AI agents like chatbots is ending. The teams that adjust their evaluation frameworks first will find capabilities their competitors are still dismissing as impossible.