# What Is the Pareto Frontier in AI? Why OpenAI Leads It > The Pareto frontier is the set of AI models no rival beats on both cost and intelligence. On the October 2026 index, OpenAI holds 9 of its 12 points. Source: https://capitalandcompute.net/blog/what-is-the-pareto-frontier-in-ai/ Published: 2026-10-29 Updated: 2026-10-30 The Pareto frontier in AI is the set of models that no rival beats on cost and intelligence at once. On the Artificial Analysis Intelligence Index v5, read October 28 and 29, 2026, OpenAI holds 9 of its 12 points. That is the sense in which OpenAI “dominates” it. Every point on the frontier between $0.07 and $1.27 per task is a GPT-6.1 Terra or GPT-6.1 Sol setting. The top of the frontier is not OpenAI’s: Claude Opus 5.5 and Gemini 4 Carbon hold the two highest points, and OpenAI’s own flagship, GPT-6 Astra, is not on the frontier at any effort setting. Key takeaways * A model is on the Pareto frontier when no other model is both cheaper and smarter. Everything else is dominated: you could pay less and get more. * On Artificial Analysis's v5 index, measured as cost per task across 62 costed effort settings of 20 models, 12 settings are on the frontier: 9 from OpenAI, 2 from Anthropic, 1 from Google. * GPT-6.1 Sol at max effort scores 55 for $1.27 a task. Claude Sonnet 5.5 at max scores 54 for $10.02 on the same $2/$10 rate card, so the gap is tokens spent, not price. * OpenAI does not own the top. Claude Opus 5.5 (xhigh, 57) and Gemini 4 Carbon (59) sit above Sol, and those last 2 to 4 points cost about 5 to 6 times more per task. ## What does “Pareto frontier” mean for AI models? The idea comes from economics. A situation is Pareto efficient when nothing can be made better without making something else worse, as the [Wikipedia entry on Pareto efficiency](https://en.wikipedia.org/wiki/Pareto%5Fefficiency) summarizes the welfare-economics definition, named after the Italian economist Vilfredo Pareto. For AI models, the two things being traded are usually capability and cost. Plot every model as a dot, with a benchmark score on the vertical axis and price on the horizontal axis. A model is **dominated** if some other dot sits above it and to its left: higher score, lower price. The models left over, the ones nothing dominates, form the frontier. Join them and you get a staircase running from the cheap, weak corner to the expensive, strong one. Three things follow from that definition, and each one is commonly misread: 1. **The frontier is not the best model.** The cheapest model with any score at all is on the frontier. So is the single highest scorer, whatever it costs. Being on it means “nothing beats this on both axes”, not “this is good”. 2. **The frontier depends on the axes.** Swap cost for speed and a different set of models qualifies. Swap one benchmark for another and the order changes again. 3. **The frontier is a snapshot.** One release or one price cut redraws it. Gemini 4 Carbon joined it on October 28, 2026, and pushed Claude Opus 5.5 at max effort off the top. The framing became standard in AI evaluation partly through a 2024 arXiv preprint, [_AI Agents That Matter_](https://arxiv.org/abs/2407.01502) (Kapoor, Stroebl, Siegel, Nadgir and Narayanan), which argued that agent leaderboards reporting accuracy alone reward expensive systems, and that results should be shown as accuracy-cost Pareto curves instead. ## Which models are on the Pareto frontier right now? The figure below plots the [Artificial Analysis models leaderboard](https://artificialanalysis.ai/leaderboards/models) as read twice on October 28 and 29, 2026, with matching values. The score is the Intelligence Index v5, which [Artificial Analysis announced on October 27, 2026](https://artificialanalysis.ai/articles/artificial-analysis-intelligence-index-v5): nine evaluations across four categories, about 60% of it on private datasets. The cost is Artificial Analysis’s cost per index task at each provider’s list price, which captures how many tokens a model actually burns, not just its rate card. Each reasoning-effort setting is its own dot, because each one is a separate price-and-score choice a buyer can make. The frontier is computed from the data, not drawn by hand. __AI models: Intelligence Index v5 against cost per task, October 2026__ | Model | Intelligence Index v5 | Cost per index task | On the cost-efficiency frontier | | ---------------------- | --------------------- | ------------------- | ------------------------------- | | Gemini 4 Carbon (high) | 59 | $7.76 | Yes | | Opus 5.5 (max) | 58 | $10.04 | No | | Opus 5.5 (xhigh) | 57 | $6.05 | Yes | | Astra (max) | 57 | $6.24 | No | | Sol (max) | 55 | $1.27 | Yes | | Sonnet 5.5 (max) | 54 | $10.02 | No | | Sol (xhigh) | 52 | $0.76 | Yes | | Opus 5.5 (high) | 52 | $3.46 | No | | Terra (max) | 51 | $0.96 | No | | Astra (medium) | 51 | $2.73 | No | | Sol (high) | 50 | $0.59 | Yes | | Gemini 4 Argon (high) | 50 | $7.68 | No | | Sol (medium) | 46 | $0.36 | Yes | | Kimi K3.1 Pro | 46 | $3.68 | No | | Astra (low) | 43 | $1.42 | No | | Muse Wonder (max) | 43 | $6.95 | No | | Terra (high) | 42 | $0.26 | Yes | | Sonnet 5.5 (high) | 41 | $1.64 | No | | Terra (medium) | 38 | $0.17 | Yes | | Haiku 5.5 (max) | 36 | $0.41 | No | | Sol (low) | 35 | $0.16 | Yes | | Opus 5.5 (low) | 35 | $0.98 | No | | Grok 4.7 (high) | 35 | $6.18 | No | | DeepSeek V4.1 Flash | 32 | $0.60 | No | | GLM-5.3 (max) | 32 | $5.01 | No | | Terra (low) | 29 | $0.07 | Yes | | MiniMax-M3.1-Flash | 29 | $0.49 | No | | Luna (max) | 28 | $0.15 | No | | Haiku 5.5 (low) | 19 | $0.04 | Yes | | Luna (low) | 10 | $0.01 | Yes | Artificial Analysis Intelligence Index v5 against cost per index task, log scale. Each dot is one model at one reasoning-effort setting. The dashed staircase is the Pareto frontier: 12 settings, 9 of them OpenAI. Every frontier point from $0.07 to $1.27 is GPT-6.1 Terra or GPT-6.1 Sol. Gemini 4 Carbon is plotted although only Google's Fairwind cyber-defender program can use it.Source: [Artificial Analysis models leaderboard, Intelligence Index v5, read October 28 and 29, 2026; frontier computed by this site](https://artificialanalysis.ai/leaderboards/models) The 12 frontier settings, in order of cost: | Frontier point | Lab | Index v5 | Cost per task | Gain over the previous point | | ----------------------- | --------- | -------- | ------------- | ---------------------------- | | GPT-6 Luna (low) | OpenAI | 10 | $0.01 | Cheapest scored setting | | Claude Haiku 5.5 (low) | Anthropic | 19 | $0.04 | +9 points for 3 cents | | GPT-6.1 Terra (low) | OpenAI | 29 | $0.07 | +10 points | | GPT-6.1 Sol (low) | OpenAI | 35 | $0.16 | +6 points | | GPT-6.1 Terra (medium) | OpenAI | 38 | $0.17 | +3 points for 1 cent | | GPT-6.1 Terra (high) | OpenAI | 42 | $0.26 | +4 points | | GPT-6.1 Sol (medium) | OpenAI | 46 | $0.36 | +4 points | | GPT-6.1 Sol (high) | OpenAI | 50 | $0.59 | +4 points | | GPT-6.1 Sol (xhigh) | OpenAI | 52 | $0.76 | +2 points | | GPT-6.1 Sol (max) | OpenAI | 55 | $1.27 | +3 points | | Claude Opus 5.5 (xhigh) | Anthropic | 57 | $6.05 | +2 points for 4.8x the cost | | Gemini 4 Carbon (high) | Google | 59 | $7.76 | +2 points, restricted access | The frontier was computed over all 62 effort settings with both a v5 score and a cost on the leaderboard, of which 30 are plotted. Every unplotted setting is dominated by a plotted frontier point, so leaving it out does not change the staircase. The leaderboard lists Claude settings as “with fallback”; they are labelled here by effort level only. ## Why does OpenAI dominate the Pareto frontier? Three reasons, and the first one matters most. **OpenAI’s models use far fewer tokens per task.** GPT-6.1 Sol and Claude Sonnet 5.5 carry the same list price, $2 per million input tokens and $10 per million output. At max effort Sol scores 55 for $1.27 a task. Sonnet 5.5 scores 54 for $10.02\. With the rate card identical, nearly all of that 7.9x gap is how many tokens each model spends to finish the same tasks (input mix and caching differ a little too). Artificial Analysis’s v5 announcement reports that every measured model now generates more output tokens per task than under v4.3, with the median rising from 35,000 to 59,000\. A model that reaches the same answer in fewer tokens moves left on the chart without any price cut. | Same $2/$10 rate card | Index v5 | Cost per task | Cost per index point | | ------------------------- | -------- | ------------- | -------------------- | | GPT-6.1 Sol (max) | 55 | $1.27 | 2.3 cents | | GPT-6.1 Sol (high) | 50 | $0.59 | 1.2 cents | | Claude Sonnet 5.5 (max) | 54 | $10.02 | 18.6 cents | | Claude Sonnet 5.5 (xhigh) | 50 | $4.09 | 8.2 cents | **The lineup covers the whole curve.** OpenAI sells three current tiers, [GPT-6.1 Terra](/blog/gpt-6-1-terra-pricing-benchmarks/), GPT-6.1 Sol and [GPT-6 Luna](/blog/gpt-6-sol-luna-pricing-benchmarks/), and each has five or six effort settings. That is 17 priced points on the leaderboard, and the effort dial lets the same model step up the staircase. Sol is on the frontier at all five of its settings. Terra is on it at low, medium and high. **Prices kept falling.** OpenAI released GPT-6.1 Terra on October 26 at $1 and $5 per million tokens, half the GPT-6.1 Sol rate. In [its launch article](https://artificialanalysis.ai/articles/gpt-6-1-terra-pushes-the-cost-efficiency-frontier), Artificial Analysis said Terra “pushes the cost efficiency frontier” and called it Pareto-optimal at low and medium effort on the older v4.3 index. On v5, Terra’s high setting is on the frontier too. The open-weight models are off the frontier at their first-party prices No open-weight model is on the frontier on this read. DeepSeek V4.1 Flash scores 32 for $0.60 and MiniMax-M3.1-Flash scores 29 for $0.49, while GPT-6.1 Terra at low effort scores 29 for $0.07\. Self-hosting changes the cost axis, so this applies to the providers’ own APIs as Artificial Analysis prices them. ## Where is OpenAI not on the frontier? At the top, and with its own flagship. **GPT-6 Astra is dominated at every effort setting.** At max effort Astra scores 57 for $6.24 a task. [Claude Opus 5.5](/blog/claude-opus-5-5-pricing-benchmarks/) at xhigh scores the same 57 for $6.05, 19 cents less. Astra’s lower settings are dominated by OpenAI’s own Sol: Astra at high scores 55 for $3.24, and Sol at max scores 55 for $1.27\. So the model OpenAI positions as its strongest is the one its buyers have the least reason to pick on these two axes. [GPT-6 Astra](/blog/gpt-6-astra-pricing-benchmarks/) launched at $10 and $50 per million tokens, five times the Sol rate. **The highest scores belong to Google and Anthropic.** [Gemini 4 Carbon](/blog/gemini-4-carbon-pricing-benchmarks/) leads at 59 for $7.76, which dominates Claude Opus 5.5 at max (58 for $10.04). Carbon is only available to cyber defenders in Google’s Fairwind program, so for most buyers today the top usable point is Opus 5.5: 57 at xhigh, or 58 at max. **The top of the curve is steep.** From Sol at max to Opus 5.5 at xhigh, the score rises 2 points and the cost per task rises 4.8x. Artificial Analysis’s v5 announcement puts it from the other side: Sol at max “scores within 4 points of Claude Opus 5.5 (max) at about an eighth of the cost per task.” Whether those last points are worth 5 to 8 times the spend depends on the task, which is why the frontier keeps both models on it. **Speed is a different frontier.** In an [August 2026 post on X](https://x.com/ArtificialAnlys/status/2085083490056589784), Artificial Analysis reported that only two labs occupied its time-per-task frontier on the index then in use: every frontier model under two minutes was OpenAI’s, and four of the six above it were Anthropic’s. Its time measure divides output tokens by output speed, so it rewards the same token efficiency that drives the cost chart. It does not include tool-call or network time, a limitation critics raised at the time. That chart predates v5 and the October releases. ## Should you measure the frontier in cost per token or cost per task? Cost per task, when the source has it. A per-token price only tells you the rate. It says nothing about how many tokens a model spends, and that is where models differ most. Two examples from the same leaderboard read show how far the two measures diverge: * **Same price, 7.9x different cost.** GPT-6.1 Sol and Claude Sonnet 5.5 both list at $2/$10, yet cost $1.27 and $10.02 per task at max effort. * **Same model, 10x different cost.** Claude Opus 5.5 scores 35 at low effort for $0.98 and 58 at max for $10.04, a tenfold spread that Artificial Analysis also notes in its v5 announcement. A frontier drawn on per-token prices, like the blended-price chart on the [Capital & Compute homepage](/), answers “which model has the best rate card for its score”. A per-task frontier answers “which model finishes the work cheapest for its score”. For agentic and reasoning workloads, where token counts swing by an order of magnitude, the second question is the one your bill answers. The [GPT-6.1 Terra launch](/blog/gpt-6-1-terra-pricing-benchmarks/) and the [October 2026 price cuts](/blog/why-are-openai-and-anthropic-cutting-prices/) are both cases where the rate card and the per-task cost tell different stories. ## How do you use the Pareto frontier to pick a model? 1. **Set your quality floor first.** Decide the lowest score, on a benchmark close to your workload, that you can accept. The frontier does not choose this for you. 2. **Take the cheapest frontier point above that floor.** If you need an index of 45 or so, Sol at medium (46, $0.36) is the frontier choice; Opus 5.5 at medium scores 48 but costs $2.45. 3. **Treat effort as a dial, not a fixed model choice.** Most frontier points are effort settings of two OpenAI models. Running one model at several efforts, routed by task difficulty, can track the staircase without switching vendors. 4. **Re-check the benchmark that matches your work.** A general index is an average. On a coding-specific board the frontier can differ: on Terminal-Bench 4.0, Claude Opus 5.5 led on accuracy in October 2026, as the [GPT-6 Astra analysis](/blog/gpt-6-astra-pricing-benchmarks/) shows. 5. **Re-read after every release.** The frontier moved twice in the last week of October 2026 alone, with Terra on October 26 and Carbon on October 28\. The [AI model leaderboard](/ai-model-leaderboard/) and the [Intelligence Index v5 changes](/blog/artificial-analysis-index-v5-which-models-dropped/) track those moves. ## Frequently asked questions What is the Pareto frontier in AI? It is the set of AI models that no other model beats on two measures at once, usually capability and cost. A model is on the frontier when nothing is both cheaper and higher-scoring. Every other model is dominated: some alternative gives you more for less. The frontier changes with each release and price cut, and depends on which benchmark and which cost measure you use. Does OpenAI dominate the AI Pareto frontier? In the low and middle part of the cost range, yes. On the Artificial Analysis Intelligence Index v5, read October 28 and 29, 2026, 9 of the 12 frontier points are OpenAI settings, and every frontier point from $0.07 to $1.27 per task is GPT-6.1 Terra or GPT-6.1 Sol. The top two points are Claude Opus 5.5 at xhigh (57) and Gemini 4 Carbon (59), and GPT-6 Astra is not on the frontier at any effort. Why is GPT-6.1 Sol so much cheaper per task than Claude Sonnet 5.5? Both list at $2 per million input tokens and $10 per million output, but at max effort Sol costs $1.27 per Artificial Analysis index task against $10.02 for Sonnet 5.5, for a score of 55 against 54\. With the same rate card, the difference comes almost entirely from how many tokens each model spends to finish the tasks. Is the model on the Pareto frontier the best model? Not necessarily. The frontier includes the cheapest model with any score and the single highest scorer regardless of cost. Being on it only means no alternative is better on both axes. The right frontier point depends on the minimum quality your task needs and how much you will pay for each extra point. ## Sources Artificial Analysis (2026). _Models leaderboard_, Intelligence Index v5 and cost per index task by reasoning effort. Independent benchmarking service. Read 2026-10-28 and 2026-10-29, matching values. Artificial Analysis (2026). _Announcing Artificial Analysis Intelligence Index v5_. Artificial Analysis, October 27, 2026\. Artificial Analysis (2026). _GPT-6.1 Terra pushes the cost efficiency frontier_. Artificial Analysis, October 26, 2026\. Artificial Analysis (2026). _Only two labs occupy the Time per Task Pareto frontier_. Post on X, August 2026\. Kapoor, S., Stroebl, B., Siegel, Z. S., Nadgir, N. and Narayanan, A. (2024). _AI Agents That Matter_. arXiv (preprint). Wikipedia (2026). _Pareto efficiency_. Encyclopedia entry (secondary).