On 22 September, Anthropic released Claude Opus 5.5. OpenAI answered the same afternoon with GPT-6 Sol and GPT-6 Luna. Nine models had already pushed Terminal-Bench 2.1, the standard test of how well an AI agent handles real command-line work, past 87 to 92 percent. The benchmark's maintainers retired it days later. Its replacement, Terminal-Bench 4.0, is harder and fairer. On it, the same frontier models top out at 59.6 percent. That is not a story about weaker models. It is a story about how much a passing score was hiding.
A leaderboard number tells you less than you think
Terminal-Bench 4.0 was built by the Laude Institute with Stanford researchers and open source contributors. It covers 66 tasks across software, machine learning, science, operations, security, hardware and media work, and it deliberately removed eight tasks from the old version that were saturated, refusal-prone, publicly solved or affected by unresolved quality issues. Claude Opus 5.5 and GPT-6 Astra now share the top score, at 59.6 percent.
That reset matters alongside research from UC Berkeley's Centre for Responsible, Decentralized Intelligence, published in April this year. Researchers built a scanning agent that audited eight major AI agent benchmarks, including SWE-bench, Terminal-Bench, WebArena and GAIA. Every single one could be exploited to near-perfect scores without solving a single task, through weaknesses such as shipped answer keys, weak string matching and judges that trust unverified output. A benchmark score tells you how a model plays one specific, static test. It tells you far less about whether it will do the job reliably in your own environment.
Passing once is not the same as working
Separate research from Microsoft's Copilot Studio team, published in August, makes the reliability gap concrete. Testing GPT-5.4 on 507 stateful business tasks, agents succeeded on the first attempt 65.36 percent of the time, but succeeded on all 20 repeated attempts only 25.25 percent of the time. Most of that failure, 77.5 percent, happened at the tool-call layer rather than in the final answer. An agent that can find a working path once is not the same as an agent that takes that path consistently. For anyone deploying agents in production, a single pass@1 figure on a vendor's leaderboard tells you almost nothing about whether the agent will hold up on Tuesday, then Wednesday, then a hundred times after that.
Governance is starting to catch up
On 1 September, OWASP's GenAI Security Project refreshed its 2026 Top 10 for LLM Applications and moved Excessive Agency up to the third-ranked threat. Alongside it, the project accepted a new Agent Control Standard aimed at runtime enforcement, not just pre-deployment testing. That direction matches what the benchmark data shows. Evaluation cannot be a one-off gate before launch. It has to run continuously, and it has to constrain what an agent is allowed to do, not only check what it did afterwards.
If the only evidence you have that your agent works is a single leaderboard run, you have less evidence than you think. EvalSpec does not run or trace agents today. It scores the responses your system gives to a fixed set of test cases, and stores each result with a record hash so it cannot be quietly changed afterwards.
Try EvalSpec free - evaluate your AI system across all 5 dimensions. No code required.
Start your free evaluation