# The half-life of the best AI model > Last year's best model is this year's cheap one. Two years of OpenAI and Anthropic releases, prices and usage show how fast. By Mayank Kumar Poddar ยท Published: 2026-10-02T12:10:00.000Z Canonical: https://www.mynkpdr.in/posts/half-life-of-the-best-ai-model/ **TL;DR** - The best model in the world changes hands every few weeks, and Anthropic currently holds it. - A score that costs a fortune today costs one tenth as much in about five months. - The cheapest tier trails the frontier by roughly six months, and that gap is the number to watch. - The labs are not timing launches against each other. They are just shipping faster. - Anthropic overtook OpenAI among US businesses paying for AI, while ChatGPT is still far bigger with the general public. I collected two years of data (October 2024 to the end of September 2026) on releases, list prices, benchmark scores and usage for OpenAI and Anthropic, with Google, xAI, DeepSeek, Qwen, Meta and Mistral as reference points. The question: how quickly does a top model become a cheap one? ## The treadmill I use the Artificial Analysis Intelligence Index (v4.3) as a single score. It is imperfect, but it is the only public score that covers every model on one scale. ![Step chart of the best model so far for OpenAI, Anthropic and other labs](./c1_frontier.webp) Two things stand out. The frontier rose from about 11 to nearly 58 in two years, so it is steady rather than stalled. And the lead keeps swapping. OpenAI's o1 opened the window ahead; Anthropic and OpenAI then leapfrogged each other repeatedly, and today Claude Opus 5.5 (57.6) sits above GPT-6 Astra (52.7). ## What a price tier buys Now ask a different question: not "what is the best model?" but "what is the best model I can get for under $X per million tokens?" ![Best score at each price ceiling over time](./c2_price_tiers.webp) The cheapest tier ($0.20 or less) climbs in steps and always trails the frontier. Today's best model at that price, GPT-6 Luna (38.1), matches what the frontier first reached with GPT-5.4 in March 2026. That is a lag of about **six months** (5.0 months if I count only scores Artificial Analysis measured rather than estimated). The $1 tier lags about five months and the $5 tier under two. ## A tenfold price drop in five months For each flagship I counted the months until any lab offered a model with at least the same score at one tenth of the price or less. ![Months until each flagship's score was available at a tenth of the price](./c3_ten_x.webp) The median is **5.3 months**, and it is trending down: flagships from 2025 took 2 to 8 months (mostly 5 to 8), the 2026 ones 3 to 6. The open circles are recent models still waiting for their cheap twin, so the true figures for them are higher than shown. The cheap twin is usually from another lab, which means the "price of a fixed capability" is falling for everyone at once. ## Same model, different bill A model's list price is only half of the cost. These models "think" before answering, and the effort setting controls how long. ![Score against cost to run the benchmark for each model and effort setting](./c4_effort.webp) Claude Sonnet 5.5 scores about 36 on low effort and 56 on max, at roughly 18 times the cost per task. Opus 5.5 gets 53.6 at high effort for $1.82 and 57.6 at max for $5.98: a 4-point gain for more than three times the price. The dial moves cost far more than it moves score, so "which model is cheapest?" has no answer until you say how hard you want it to think. ## Release rhythm The two labs have each launched about a flagship every ten weeks on average, and both are speeding up. Anthropic's median gap between flagships fell from about 106 days in early 2025 to 39 in late 2026. OpenAI's fell from 132 to 66. ![Flagship launch dates and gaps for each lab](./c5_cadence.webp) I expected to find the labs reacting to each other, so I tested it. In 21 flagship launches, the nearest rival launch was a median of 17 days away. Randomly shuffled launch dates give 24. A shuffle that lucky happens about one time in four (p = 0.23), so I cannot call it timing games. Both labs just got faster. ## Who actually uses what This is where the picture splits, because it depends on whom you ask. ![Three audiences: US businesses, search interest, developers](./c6_audiences.webp) - **US businesses.** Per Ramp's card data, OpenAI led 32.3% to 9.6% in October 2024. Anthropic passed it in May 2026 and was ahead 43.8% to 39.8% by August. - **The general public.** Google searches for ChatGPT outnumbered Claude about 44 to 1 at the start of the window and about 6 to 1 now. OpenAI said ChatGPT had 900 million weekly users in February 2026 and, according to press coverage, 1.2 billion by late September. - **Developers.** On OpenRouter, the Anthropic share of tokens fell from about 40% to under 4% of tokens in the latest week, and OpenAI sits around 13%. Cheaper rivals took most of it. So "who is winning?" has three answers. Paying businesses lean Anthropic, the public leans OpenAI, and price-sensitive developers lean towards whatever is cheapest. ## Caveats - **Benchmarks are a proxy.** One index cannot capture what you care about. Before 2026, most scores are Artificial Analysis estimates, and I repeated the key results using measured scores only. - **List prices, not bills.** Discounts, caching and batch tiers change real costs. - **Usage data is partial.** Ramp is US card spend, Trends is search interest, and OpenRouter skews towards developers who switch models on purpose. - **Small samples.** Twenty-odd flagship launches is enough for a trend, not for fine detail. ## Sources and checks - **Scores and cost per task:** [Artificial Analysis leaderboard](https://artificialanalysis.ai/leaderboards/models), Intelligence Index v4.3, retrieved 2 October 2026. - **Release dates and prices:** Anthropic's [release notes](https://platform.claude.com/docs/en/release-notes/overview), OpenAI's [API docs](https://developers.openai.com/api/docs) and [pricing page](https://openai.com/api/pricing), plus Google, xAI, DeepSeek and Mistral changelogs. Older prices come from [Wayback Machine](https://web.archive.org/) copies of those pages. - **Usage:** [Ramp AI Index](https://ramp.com/data/ai-index), [Google Trends](https://trends.google.com/trends/explore), [OpenRouter rankings](https://openrouter.ai/rankings) and [Similarweb](https://www.similarweb.com/website/claude.ai). I had 20 of the headline claims (prices, launch dates, user counts, usage shares) re-checked against primary sources by a separate pass. Two were wrong and are fixed: the OpenRouter token shares. One date was off by a month (ChatGPT's 900 million weekly users, first stated in February 2026).