OpenAI's GPT-6.1 Terra joins the board at #3 as the references move to policyengine-us 2.60.0
PolicyBench added OpenAI's GPT-6.1 Terra to its 2026-10-27 snapshot, the day after OpenAI released it, bringing the board to 51 models. GPT-6.1 Terra gets 93.1% of answers right and ranks #3 of 51. Only GPT-6 Sol (95.2%) and Claude Opus 5.5 (94.1%) score higher, and GPT-5.6 Sol, which also rounds to 93.1%, trails by less than a tenth of a point. On the bounded score, which gives partial credit for amounts close to the reference, GPT-6.1 Terra ranks #5. GPT-6.1 Sol, which OpenAI released on September 29, gets 90.2% of answers right (#10). PolicyBench counts a dollar amount as right when it falls within $1 of the reference PolicyEngine computes, and an eligibility answer as right when it matches. It weights each answer by household impact.
PolicyBench runs GPT-6.1 Terra as it runs GPT-6.1 Sol, on OpenAI's Responses API with the board's forced answer tool at OpenAI's default reasoning effort, and it returned reasoning tokens on all 100 households. At OpenAI's price of $1 per million input tokens and $5 per million output tokens, the 100 households cost $1.25, against $2.30 for GPT-6.1 Sol and $2.71 for GPT-6 Sol. GPT-5.6 Terra, the Terra model before it, gets 87.6% (#18) and cost $3.89.
The release also moves PolicyBench's references from policyengine-us 2.41.0 to 2.60.0, the newest release when PolicyBench began recomputing them on October 23. The move changes four scored references, each for law in force before the references were frozen. An elderly Californian who lives alone, whose SNAP was $0, now gets the $24 monthly minimum, $288 a year: California's categorical eligibility for CalFresh applies no net income test, even to a household with an elderly member, and the previous version applied one (policyengine-us #10036). Three households in Pennsylvania and Arkansas have too much income for free or reduced-price school lunch, but those states give students at covered schools a free breakfast, so the output asking whether PolicyEngine returns free school meal support is now yes for each (policyengine-us #10238). The new version models both breakfast laws for a student at a covered school and leaves out their conditions: Pennsylvania reimburses schools for them to the extent funds are available, and Arkansas serves them on request.
Nothing is newly excluded, and no excluded output returns to scoring, so every model is still scored on 1,935 of its 1,984 requested outputs. When PolicyBench checked PyPI on 2026-10-29, the newest policyengine-us release was 2.78.0, and it gives the same value as 2.60.0 for all 1,935; it moves one excluded output, which keeps the value it was decided on.
Against the previous release, 47 of the 50 earlier models fall, by 0.03 to 0.07 points, and three rise by 0.02. No model answers yes for any of the three households' school meals, and only three give the Californian's new SNAP: GPT-5.5, GPT-6 Astra and GPT-6.1 Sol, the three that rise. Claude Sonnet 5.5 falls from #4 to #6, behind GPT-6.1 Terra and GPT-6 Astra.