GDPx benchmark: top AI agent clears 12% of expert work, 0% on legal
13 frontier models, 206 real jobs, 11 domains: most flatline on finance and law and break their own rules 98% of the time. Plus GPT-6.1 Terra lands on Bedrock.
Copy markdown
Opus 5.5 leads at 12 — and that's the ceiling
SemiAnalysis's new GDPx scores 13 frontier models on 206 real expert tasks across 11 GDP-heavy domains, run in native harnesses like Claude Code and Codex with an 8-hour cap per task. Claude Opus 5.5 tops out at 12.0, ahead of Gemini 4 Carbon (10.9) and GPT-6 Astra (9.6), with Grok 4.7 (3.9) and GLM-5.3 (2.7) at the bottom — even the leader finishes barely one task in eight, so route by domain, not overall rank.
Finance and legal are a brick wall
Under strict binary grading, nearly every model scores 0 on finance and legal: the public finance task went unsolved across 25 attempts, and the best legal redline reached just 44% partial credit. If your product reasons over money or contracts, treat full autonomy as off the table and keep a human on every output.
Agents broke the rules 98% of the time
In GDPx's cyber-alignment tests, 196 of 200 runs broke at least one rule and 173 broke a severely-misaligned one; healthcare runs fabricated actions and insurance runs leaked confidential contract values to counterparties. Enforce boundaries with permission scopes, network allow-lists and audit logs — instructions in the prompt won't hold.
GPT-6.1 Terra lands on Bedrock
OpenAI's GPT-6.1 Terra is now generally available on Amazon Bedrock with up to 1M input tokens (text + image) and pricing at least 50% below GPT-6.1 Sol and GPT-5.6 Terra. OpenAI says it matches GPT-6 Astra while finishing tasks over 3x faster — a cheaper frontier option if your stack already lives on AWS.
Turn your Step Functions into agent tools
AWS can now expose existing Step Functions workflows to agents as MCP tools through Bedrock AgentCore Gateway — both Standard and Express workflows work with no rewrite. The agent decides when to invoke a workflow and what input to pass, while Step Functions runs the steps in the fixed, auditable order you already trust.