GDPx says the best agent scores 12%, and OpenAI's LifeVal publishes no score at all
Two agent benchmarks landed on the same Monday, and they could not be more different.
SemiAnalysis launched GDPx, which covers 11 domains picked for economic value: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Each domain comes from a partner RL-environment startup. The key design choice: every model runs in its own native harness, Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, OpenCode for everything else, with tasks of up to 8 hours. The best model scores 12%, and that is Opus 5.5. Gemini 4 Carbon wins biology, GPT-6 Astra wins RSI and materials science, and on finance and legal "basically every model scores exactly 0". Two sharper findings: safety classifiers drag down Opus 5.5 and Astra on cyber, bio and alignment tasks, and when SemiAnalysis ran its task-quality checks on Terminal Bench 4.0, it flagged nearly half of the tasks as broken. Worth knowing: the environments come from paid vendors, and SemiAnalysis sells compute economics research.
OpenAI went the other way with LifeVal, a benchmark of everyday personal-agent tasks drawn from US and multinational time-use surveys plus a new survey of 3,995 people: caregiving, health, school enrollment, household admin, shopping, government services. OpenAI says it has used LifeVal internally for a year to improve browser use, computer use and connectors. It shows example task scores like 93.1% on kindergarten enrollment, but no aggregate score, no model comparison, no task count and no dataset.
Put together, the contrast is the story. One benchmark says frontier agents fail almost every long professional task and publishes the per-harness breakdown. The other says the consumer agent is doing great and keeps the data to itself. If you build agents, the GDPx method, scoring the model inside its real harness, is the one to copy.
GDPx: https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities LifeVal: https://openai.com/index/lifeval/
← Back to all articles
SemiAnalysis launched GDPx, which covers 11 domains picked for economic value: chip design, cybersecurity, biology, finance, legal, quant trading, recursive self-improvement, hardware design, materials science, alignment and software engineering. Each domain comes from a partner RL-environment startup. The key design choice: every model runs in its own native harness, Claude Code for Opus 5.5, Codex for GPT-6 Astra, Antigravity for Gemini 4 Carbon, OpenCode for everything else, with tasks of up to 8 hours. The best model scores 12%, and that is Opus 5.5. Gemini 4 Carbon wins biology, GPT-6 Astra wins RSI and materials science, and on finance and legal "basically every model scores exactly 0". Two sharper findings: safety classifiers drag down Opus 5.5 and Astra on cyber, bio and alignment tasks, and when SemiAnalysis ran its task-quality checks on Terminal Bench 4.0, it flagged nearly half of the tasks as broken. Worth knowing: the environments come from paid vendors, and SemiAnalysis sells compute economics research.
OpenAI went the other way with LifeVal, a benchmark of everyday personal-agent tasks drawn from US and multinational time-use surveys plus a new survey of 3,995 people: caregiving, health, school enrollment, household admin, shopping, government services. OpenAI says it has used LifeVal internally for a year to improve browser use, computer use and connectors. It shows example task scores like 93.1% on kindergarten enrollment, but no aggregate score, no model comparison, no task count and no dataset.
Put together, the contrast is the story. One benchmark says frontier agents fail almost every long professional task and publishes the per-harness breakdown. The other says the consumer agent is doing great and keeps the data to itself. If you build agents, the GDPx method, scoring the model inside its real harness, is the one to copy.
GDPx: https://newsletter.semianalysis.com/p/gdpx-measuring-frontier-model-capabilities LifeVal: https://openai.com/index/lifeval/
Comments