OpenAI's GPT-6 Sol Lands With Six Variants and a Sharper Price-to-Intelligence Curve

OpenAI quietly shipped GPT-6 Sol in September 2026, and per Artificial Analysis's independent benchmarking, it arrives as a family of six models rather than a single release. That's becoming the norm for frontier labs: instead of one model, you get a reasoning-effort dial with low, medium, high, xhigh, and max settings, plus a non-reasoning variant for when you just need fast, cheap output.

The top performer, GPT-6 Sol (max), scores 48 on the Artificial Analysis Intelligence Index, the firm's composite benchmark that blends ten evaluations including Humanity's Last Exam, SciCode, Terminal-Bench 4.0, and AA-Omniscience. That's a solid score, but it's not the headline. What's more interesting is the spread across the family and what it costs to get there.

Here's the breakdown:

• Intelligence range: max hits 48, xhigh comes in at 44, high at 43, medium at 40, low at 34, and the non-reasoning variant trails at 28.
• Speed: max is also the fastest at 131 tokens/second, with the rest clustering between 109 and 127 t/s. Reasoning effort and throughput aren't trading off much here, which is notable.
• Latency: the non-reasoning model wins on time-to-first-token at 0.93 seconds, unsurprising since it skips the chain-of-thought step entirely.
• Price: this is where the spread gets wide. GPT-6 Sol (low) costs $0.13 per Intelligence Index task, while max costs $1.06. That's an 8x difference for a jump from 34 to 48 on the intelligence score.

For anyone building production systems, that price-to-intelligence ratio is the actual story. A 41% score improvement (34 to 48) costs you a little over 8x more per task. Whether that trade is worth it depends entirely on your use case, and that's exactly the kind of decision this tiered release structure is designed to let you make on a per-request basis rather than locking you into one model checkpoint.

Where GPT-6 Sol Sits on the Cost-Intelligence Curve

Artificial Analysis plots every model on a cost-per-task versus intelligence scatter, and it's worth understanding why that matters more than a leaderboard rank. The chart highlights a "Pareto frontier": the line of models that give you the best intelligence for their price point, with nothing cheaper scoring higher and nothing smarter costing less.

GPT-6 Sol's cheaper tiers land favorably here. The low variant at $0.13/task and medium at $0.25/task sit near the front of that frontier, undercutting rivals like Claude Fable 5.1 (low with fallback) at $0.60 and Claude Opus 5.5 (medium with fallback) at $2.58 while staying competitive on raw intelligence score. Meanwhile GPT-6 Astra, presumably OpenAI's other current-generation line, posts comparable Intelligence Index numbers but at noticeably higher weighted costs across its tiers.

The competitive set here is worth noting too. Anthropic's Claude Opus 5.5 and Claude Fable 5.1 families (both offering "with fallback" variants, suggesting some kind of automatic degradation to a backup model under load or cost constraints) occupy the higher-intelligence, higher-cost end of the curve. xAI's Grok 4.7 sits in the middle. GPT-5.6 Sol, the predecessor release, trails GPT-6 Sol on the frontier entirely, which is exactly what you'd expect from a version bump but still nice to see confirmed empirically rather than taken on faith.

One architectural note that's easy to miss in the topline numbers: GPT-6 Sol supports a 872k-token context window, which Artificial Analysis translates to roughly 1,308 pages of 12-point Arial text. That's a meaningful jump for long-document workflows, code-base-wide analysis, or multi-turn agent sessions that need to retain state across long horizons. Input modality covers text and image; output is text-only, so this isn't a multimodal generation model, it's a multimodal-input reasoning engine.

The cost breakdown by token type (input, cache hit, cache write, reasoning, and answer tokens) also matters for anyone doing serious cost modeling. Reasoning tokens, the intermediate "thinking" tokens a model burns through before producing an answer, are typically where the bill balloons on higher-effort tiers. That's almost certainly why max costs 8x more than low despite similar output speeds: it's not that the model types slower, it's that it thinks longer before it starts typing.

For teams choosing between tiers, the practical calculus looks like this: if your workload is latency-sensitive and doesn't need deep multi-step reasoning, the non-reasoning variant's 0.93s time-to-first-token and 28-point Intelligence Index score might be entirely sufficient. If you're running complex agentic workflows, long-horizon code generation, or research-grade synthesis, paying the premium for max's 48 score and 131 t/s throughput probably pays for itself in fewer retries and less human review.

Key Takeaways

GPT-6 Sol's real contribution isn't a single benchmark number, it's the granularity of the tradeoff curve OpenAI has now exposed across six tiers. The low and medium variants offer some of the most attractive cost-to-intelligence ratios currently tracked by Artificial Analysis, undercutting comparable Anthropic and xAI offerings while staying near the Pareto frontier. The max tier pushes the ceiling to 48 on the Intelligence Index, OpenAI's highest score in this family, but at roughly 8x the per-task cost of the cheapest option. For engineering teams, that means the right move isn't picking "the best GPT-6 Sol model," it's matching reasoning effort to the actual complexity of each task, something the tiered release structure now makes considerably easier to do at the API level rather than through model selection alone.