GPT-6.1 Sol replaces GPT-6 Sol after seven days with near-Astra intelligence at quarter the cost

OpenAI's Sol tier has been overhauled after only seven days: GPT-6.1 Sol replaces GPT-6 Sol, and lands one point below the more expensive GPT-6 Astra on the Artificial Analysis Intelligence Index while costing less than a quarter as much per task. The findings come from Artificial Analysis, whose evaluation of the new model reached the Hacker News front page on 29 September.

Benchmark gains

According to Artificial Analysis, GPT-6.1 Sol adds four Intelligence Index points over GPT-6 Sol and five over GPT-5.6 Sol, leaving it a single point behind GPT-6 Astra. The largest single jump came on Terminal-Bench 4.0, where the model gained 12 points, followed by an eight-point accuracy improvement on AA-Omniscience. Other moves include six points on GDP.pdf, five on both Humanity's Last Exam and GDPval-AA v2.1, and four on AA-Briefcase v1.1, with the agentic knowledge-work evaluations showing the steadiest progress.

Accuracy improved alongside honesty: the measured hallucination rate on AA-Omniscience dropped from 60% to 54%. That is a meaningful shift, though still high enough to matter for anything requiring factual grounding. On AA-Briefcase the model climbed roughly 80 Elo, driven by stronger rubric and analytical-quality scores, while its presentation score slipped slightly.

Pricing and cost efficiency

List prices are unchanged: $2 per million input tokens and $10 per million output tokens, identical to GPT-6 Sol. The effective cut comes from caching, with the cache-read discount rising from 90% to 95%. Artificial Analysis calculates that this makes the blended price for agentic workloads slightly lower than GPT-6 Sol's, extending a price slide that began when GPT-6 Sol launched at half the price of GPT-5.6 Sol.

Measured per Intelligence Index task at maximum effort, GPT-6.1 Sol costs $0.72, against $3.26 for GPT-6 Astra, $1.05 for GPT-6 Sol and $1.99 for GPT-5.6 Sol. That is 31% cheaper than the model it replaces and 64% cheaper than the generation before. Artificial Analysis also reports that every effort level of the new model extends the cost-efficiency Pareto frontier, meaning no other model delivers the same intelligence score for less money.

The trade-off: token use

The gains are not free. GPT-6.1 Sol generates roughly 10–30% more output tokens than GPT-6 Sol across effort settings. Because its intelligence score rose as well, the low and medium effort tiers remain Pareto optimal for token efficiency, but anyone billed on raw output volume, or working close to context limits, will notice the increase.

Coding results

On the Artificial Analysis Coding Agent Index, GPT-6.1 Sol gains three points over GPT-6 Sol at max effort and sits two points behind GPT-6 Astra. At the xhigh effort setting it scores one point above GPT-6 Astra while costing less than 15% as much per task, a six-point gain over GPT-6 Sol at max effort and three points better than its own max setting. Artificial Analysis found the xhigh configuration outperforming max outright, an unusual inversion that suggests effort tuning now matters as much as model choice for coding agents.

Why it matters

A seven-day replacement cycle shows how quickly mid-tier models are becoming perishable goods. Near-flagship capability at a quarter of the per-task cost reshapes the economics of agentic and coding workloads, where cost per completed task, not cost per token, is the number that decides what ships to production. The caveats deserve equal weight: more output tokens per answer and a 54% hallucination rate on AA-Omniscience mean the headline price understates total cost for workloads that are sensitive to verbosity or factual reliability. For buyers, the practical lesson is that the Sol tier is now the value pick for most jobs, and that committing deeply to any single point release in this lineup carries real depreciation risk.