Porter’s 1996 exhibit, redrawn with 100 real model settings from 23 labs
In “What Is Strategy?”, Michael Porter asks us to “imagine for a moment a productivity frontier that constitutes the sum of all existing best practices at any given time” (Porter, 1996). The market for AI models lets us draw that frontier with real numbers. Every point on the charts below is one model at one reasoning setting. Its height is what it can do, measured by the Artificial Analysis Intelligence Index. Its position from left to right is what it costs to run one task.
What emerges is Porter’s argument, almost line for line. Most models sit behind the frontier. The few on it face a hard trade-off between capability and cost. The frontier moves outward every few months. Rivals converge on similar capability, and competition slides toward price. And only three labs are shifting the curve.
The frontier today
The short version
1 · Behind the frontier
89 of 100
settings are beaten by a rival, or a sibling, that scores at least as high for less. Qwen3.8 Max scores 45 at $5.41 a task; MiMo-V2.6-Pro scores 46 for $0.13.
2 · At the frontier
+16 points, 11× the cost
On the frontier, capability has a price. Claude Opus 5.5’s last two points cost $1.26 each, against about 9 cents a point for its first nine.
3 · The frontier moves
13× a year
The cost of a given level of performance has fallen about 47% a quarter since 2023. New releases routinely make a lab’s own older models obsolete.
4 · Convergence
8 labs, 2 points
Eight labs score between 44 and 46, at costs from $0.13 to $5.41 a task. Between rivals this close, the competition shifts to price.
5 · Who shifts the curve
3 labs, 5 models
OpenAI’s GPT-6 Luna holds the cheapest end, Xiaomi’s MiMo-V2.6 the middle, and Anthropic’s Claude Opus 5.5 everything from $1.34 a task up.
The strategic point
Porter’s warning applies with unusual force. When rivals converge on the same best practice, “competition becomes a series of races down identical paths that no one can win” (Porter, 1996). A lasting advantage in AI, for a lab or for a company that uses AI, will have to come from strategic positioning, not from the model alone.
The map
Read it like Porter’s exhibit. Capability rises up the page and cost falls to the right, so the best place to be is the top-right corner. The black staircase is the productivity frontier: for any budget, the highest score anyone offers at that cost or less. The shaded area beyond it is what no model can do yet. Hover over or tap any point, or focus the chart and use the arrow keys.
The frontier is a menu, not a league table. A buyer who needs a score of 46 can pay $0.13 a task for it, or $5.41.
| Setting (as Artificial Analysis lists it) | Lab | Index | Cost per task | Frontier |
|---|
Lens 1 of 5 · Behind the frontier
A setting is behind the frontier when another one scores at least as high for less. By that test, 89 of the 100 settings are behind. That is normal in an industry where best practice moves this fast, but the distances are large. The arrows run from five well-known models to the cheapest rival that matches or beats them.
Being behind is not only a problem for smaller labs. Anthropic’s own Claude Fable 5.1 (max) scores 53 at $7.63, while Claude Opus 5.5 (high) scores 54 for $1.82. Meta’s Muse Spark 1.3 (max) is close: GPT-6 Sol (max) matches its 48 for $1.06 instead of $1.60.
Behind the frontier, Porter’s trade-off does not yet bind. A lab that improves how it trains and serves its models can raise capability and cut cost at the same time. That is operational effectiveness: necessary, valuable, and open to everyone.
“Simultaneous improvement of cost and differentiation is possible only when a company begins far behind the productivity frontier or when the frontier shifts outward.”Porter (1996)
Lens 2 of 5 · At the frontier
Most models can run at several reasoning settings. A higher setting lets the model think longer, which uses more tokens and costs more per task. Each step up the dial is a movement along the frontier, not beyond it: capability bought with cost.
Claude Opus 5.5 shows the trade-off cleanly. From its low setting to max it gains 16 points, 42 to 58, while its cost per task rises almost elevenfold, from $0.55 to $5.98. The price of each extra point climbs as you go: about 9 cents a point from low to medium, 16 cents from medium to high, 82 cents from high to xhigh, and $1.26 from xhigh to max.
Not every trade-off is real. Three models are listed at both a low-reasoning and a non-reasoning setting, all from OpenAI, and for each of them the low setting scores higher at the same or lower cost. GPT-6 Sol scores 34 at $0.13 with low reasoning, and 28 at $0.33 without. Porter calls these false trade-offs, which “occur primarily when there is redundant or wasted effort, poor control or accuracy, or weak coordination.”
For a buyer, the right setting is a position, not a ranking. A task that needs a score of 51 should not pay for 58.
“At the frontier, where companies have achieved current best practice, the trade-off between cost and differentiation is very real indeed.”Porter (1996)
Lens 3 of 5 · The frontier moves
Porter wrote that “the productivity frontier is constantly shifting outward as new technologies and management approaches are developed and as new inputs become available.” In AI the shift is measured in months. Epoch AI estimates that “the cost of achieving a given level of AI performance has fallen about 47% per quarter since 2023, or 13× per year” (Emberson & Roodman, 2026). Across individual tasks the range is wider still, from 9× to 900× a year (Cottier et al., 2025). The Stanford AI Index reports that the cost of a system performing at the level of GPT-3.5 “dropped over 280-fold between November 2022 and October 2024” (Stanford Institute for Human-Centered Artificial Intelligence [Stanford HAI], 2025).
Slide forward to see what the 13-fold rate would do to today’s frontier if it simply continued. After a year, the score of 58 that costs $5.98 a task today would cost about $0.46, and every setting on today’s map would sit behind the new frontier.
The same force shows up inside each lab, where new releases make older ones obsolete:
The releases themselves come quickly. Google shipped Gemini 3.8 Flash on 2 September 2026, three weeks after its previous model release (Li, 2026).
“OE competition shifts the productivity frontier outward, effectively raising the bar for everyone. But although such competition produces absolute improvement in operational effectiveness, it leads to relative improvement for no one.”Porter (1996)
The moved frontier illustrates a historical rate. It is not a forecast.
Lens 4 of 5 · Convergence
Porter’s second reason why operational effectiveness is not enough is, in his words, “more subtle and insidious”: competitive convergence. “The more benchmarking companies do, the more they look alike.” AI labs benchmark one another in public, on shared leaderboards.
Move the band and count the labs inside it. Between 44 and 46 there are 8 labs, from the United States and China, with 12 settings between them. Their costs run from $0.13 to $5.41 a task, 42× apart. Between rivals this close, the thing left to compete on is price.
The very top has not converged. Anthropic’s best score is 58, OpenAI’s 53 and Meta’s 48. But the pack behind them is dense, and the gaps that once separated kinds of model have nearly closed. On the Arena leaderboard, the best closed-weight model led the best open-weight model by 15.2% in May 2023, by 0.5% in August 2024 and by 3.3% in March 2026. The top US model led the top Chinese model by 2.7% (Stanford HAI, 2026). On broad knowledge tasks, “the overall spread between the top-ranked and 15th-ranked model is just over 4 percentage points.”
“With capability no longer a clear differentiator, competitive pressure is shifting toward cost, reliability, and real-world usefulness.”Stanford HAI (2026), AI Index, Chapter 2
“As rivals imitate one another’s improvements in quality, cycle times, or supplier partnerships, strategies converge and competition becomes a series of races down identical paths that no one can win.”Porter (1996)
Lens 5 of 5 · Who is shifting the curve
Colour the frontier by who owns it. All 11 frontier points come from 5 models and 3 labs. Every other lab, including Google, Meta, SpaceXAI, Alibaba and DeepSeek, is behind the frontier at every price.
These are three different positions: the cheapest capable model, the best value in the middle, and the highest capability at a premium. Whether each is a strategy or merely a lead is the question Porter would ask. At a 13-fold annual shift, a lead in capability or cost lasts only until the next release. The durable question is what each lab does that others cannot easily copy.
“A company can outperform rivals only if it can establish a difference that it can preserve.”Porter (1996)
| Budget per task | Lab | Frontier model | Scores |
|---|
Positions, not just scores
A score is operational effectiveness. A position is a choice about which customers, needs and price points to serve, backed by activities that rivals cannot copy without giving something up. Each card gives a lab’s place on the map and the question Porter would ask of it. They are prompts for discussion, not verdicts.
On the map. The highest score measured, 58, from Claude Opus 5.5 at max, and every frontier point above $1.34 a task. Its other models sit behind. Claude Fable 5.1 scores 53 at up to $7.63. Its cheapest listed model, Claude 4.5 Haiku, scores 17 at $0.28, where GPT-6 Luna (medium) scores 29 for $0.02.
Porter’s question. A premium position lasts only while the top is scarce. What would a rival have to give up to occupy it?
On the map. The broadest line-up, with 26 settings. It holds the cheapest end of the frontier with GPT-6 Luna and a band in the middle with GPT-6 Sol. Its most capable model, GPT-6 Astra, scores 53 at max but is behind the frontier at every setting.
Porter’s question. Is serving every price point a coherent position, or straddling, the trap Porter describes at Continental Lite?
On the map. MiMo-V2.6-Pro scores 46 for $0.13 a task, the best capability available for any budget from $0.13 to $1.06. MiMo-V2.6-Flash holds the frontier at $0.06. Xiaomi is the only lab outside the United States on the frontier.
Porter’s question. Is its low cost a position, backed by activities rivals cannot copy, or a price others can match next quarter?
On the map. Its best setting, Gemini 3.8 Flash (high), scores 41 at $1.24. MiMo-V2.6-Pro scores 46 for about a tenth of that. The newest Pro model on the leaderboard is Gemini 3.1 Pro Preview, 30 at $0.67. Google shipped 3.8 Flash three weeks after its previous release.
Porter’s question. If Google’s advantage is not on this chart, where is it? Consider access-based positioning: reaching customers through channels others lack.
On the map. Muse Spark 1.3 (max) scores 48 at $1.60, just behind GPT-6 Sol (max), which matches it for $1.06. Its small Muse Glimmer (high) scores 17 at $0.06.
Porter’s question. How close to the frontier is close enough, when the frontier moves 13-fold a year?
On the map. Grok 4.7 scores 46 at $2.73 (high) and $3.74 (xhigh). MiMo-V2.6-Pro reaches 46 for $0.13. SpaceXAI is the former xAI, which SpaceX acquired in February 2026 and renamed in July (“SpaceXAI,” 2026).
Porter’s question. Belonging to a rocket and satellite company is a distinctive set of activities. Does it create fit that a buyer of a model would pay for?
On the map. 12 settings, the third-largest line-up. Qwen3.8 Max scores 45 at $5.41, the most expensive setting at that level, while Qwen3.8-Flash-Next scores 40 for $0.37.
Porter’s question. Many models for many needs is variety-based positioning. Is each one distinct enough to win its own segment?
On the map. V4.1 Flash (max) scores 39 at $0.27 and beats DeepSeek’s own V4 Pro. In February 2025, DeepSeek-R1 briefly matched the top US model (Stanford HAI, 2026).
Porter’s question. A breakthrough moves the frontier for everyone. What does it take to keep the lead it creates?
On the map. GLM-5.3 (max) scores 45 at $2.01 and GLM-5.3-Flash 42 at $0.25. Kimi K3 (max) scores 44 at $2.00 and Step 5 Preview 44 at $0.72. MiniMax-M3 scores 29 at $0.51, and Tencent’s Hy3 25 at $0.07.
Porter’s question. In the densest band of the map, what makes any one of them the obvious choice?
“Closest to the frontier” is the lab’s best-placed setting: how many times the cheapest matching cost it pays for its score. 1× means it is on the frontier.
| Lab | Settings | On frontier | Best score | Lowest cost | Closest to the frontier |
|---|
What it means
For AI labs
The frontier moves about 13-fold a year, and eight labs crowd into a two-point band. A lead built only on a better model is a lead in operational effectiveness, and such leads diffuse: “Competitors can quickly imitate management techniques, new technologies, input improvements, and superior ways of meeting customers’ needs” (Porter, 1996). Durable advantage needs choices with trade-offs: which customers and tasks to serve, through which channels, with which kinds of model, and what to refuse.
For companies that buy AI
The right model is the cheapest setting that clears your quality bar, and it changes every few months. A task that needs a score of 46 can be served for $0.13 a task today rather than $5.41. The last two points of the top model cost $1.26 each; pay for them only if the task needs them. Re-run the choice every quarter.
For strategists
If every firm in an industry uses the same frontier model, AI becomes operational effectiveness: necessary, quickly copied and competed away. Advantage comes from what a firm does differently with it: its data, its workflows, and the fit between AI and the rest of its activities. “The essence of strategy is choosing what not to do” (Porter, 1996).
Method, caveats and sources
Data. The Artificial Analysis LLM leaderboard, read on 28 September 2026. Its Intelligence Index v4.3.2 combines ten evaluations (AA-Briefcase, GDPval-AA, AutomationBench-AA, Terminal-Bench 4.0, SciCode, AA-LCR, AA-Omniscience, Humanity’s Last Exam, GDP.pdf and CritPt) as “a weighted average across four categories: Agents (30%), Coding (20%), Scientific Reasoning (20%) and General (30%)” (Artificial Analysis, 2026a). Cost per task is the leaderboard’s figure in US dollars. Artificial Analysis computes costs from “the token counts reported by each model’s API provider.”
Coverage. The leaderboard listed 249 settings, and 102 showed a cost per task. Two of those, Cohere’s Command A+ and North Mini Code, show $0.00, which cannot sit on a logarithmic axis. That leaves 100 settings from 23 labs. A setting is a model at a named reasoning level (non-reasoning, low, medium, high, xhigh or max). Each is plotted separately, as Artificial Analysis lists them.
The frontier. A setting is on the frontier when no other setting scores at least as high for less, or higher for the same cost. The staircase gives the highest score available at or below each cost. Every chart computes it in your browser from the table on this page.
Checks. The table was extracted from the leaderboard twice, independently, and the two extractions agreed on every row they shared. Every figure quoted in the text was computed by script from the same rows and checked against the charts.
Names. Artificial Analysis lists Moonshot AI’s models under “Kimi” and labels Anthropic’s settings “with fallback”. The table keeps its names; the charts shorten them. SpaceXAI is the former xAI.
Limits. This is one index on one day. Scores are whole numbers, so ties are common. Costs are shown to the cent, so ratios involving costs under ten cents are approximate. The index says nothing about speed, latency, context length, reliability in production, data protection or licensing, all of which buyers weigh. Epoch AI’s rate describes the past; the projection in Lens 3 illustrates it and is not a forecast.
Independence. All scores and costs are Artificial Analysis’s, unaltered. This page adds only arithmetic and Porter.