GPT-5.6 Sol Nearly Matches the Best AI Model — at a Third of the Price

Now that GPT-5.6 is publicly available, the first independent numbers are in — and they explain why the launch has rivals nervous. According to Artificial Analysis, a platform that aggregates many AI tests into one score, GPT-5.6 Sol lands at 59 points on its Intelligence Index, just one point behind Claude Fable 5, the current leader at 60. The catch for Anthropic: Sol does it at about $1.04 per task, roughly a third of Fable 5’s $2.75.

In the platform’s new Coding Agent Index, which measures how well models handle real programming work with tools, Sol actually takes first place with 80 points when running in OpenAI’s Codex environment — ahead of Fable 5 in Claude Code at 77. The smaller GPT-5.6 variants are cheaper still: Terra costs about $0.55 per task and Luna $0.21. Official token prices per million run $5 in / $30 out for Sol, $2.50 / $15 for Terra, and $1 / $6 for Luna. Sam Altman also claims Sol uses up to 54 percent fewer output tokens on agentic coding tasks — reportedly, and from the CEO, so treat it as marketing until independently confirmed.

What’s behind this? “Cost per task” is becoming the number that matters more than raw benchmark scores. A model that scores slightly lower but burns far fewer tokens — the little text chunks AI companies charge for — can be much cheaper in practice. That’s the squeeze on Anthropic right now: its Fable 5 still leads on aggregate intelligence, but Meta’s new API, xAI’s Grok 4.5, and Chinese open-source models are all pushing prices down, and now OpenAI is matching top-tier quality at a fraction of the cost. Worth keeping expectations grounded, though: aggregated indexes hide a lot of nuance, and one point of difference on a 100-point scale says little about which model will feel better for your particular work.

What this means for you: If you subscribe to a chat plan, nothing changes today — these are developer API prices. But they trickle down: cheaper tokens mean providers can offer more usage per subscription, and tools built on these models get cheaper to run. If you’re choosing a model for a project, the practical takeaway is to stop asking “which model is smartest” and start asking “which is smart enough for this job at the lowest cost” — the gap at the top has become that narrow.

Sources

• Artificial Analysis: GPT-5.6 benchmark data
• The Decoder: GPT-5.6 Sol nearly matches Fable 5 on aggregated benchmarks
• CNBC: Sam Altman on GPT-5.6