Best AI chatbot

The layer architecture this table instruments

The unit of measurement in this table is the model, not the product. A model is a component the vendor publishes numbers about: a leaderboard score, a cost per million tokens, a context window capacity. The product is the layer above it: a client that swaps its underlying model without notice. Opus 5 simultaneously powers the infrastructure behind the Claude app, the API, and Claude Code; treat "Claude" as a single row and the same number gets counted again in the coding table. That is why the video category also ranks the model version rather than the tool.

The operational cost of that decision is explicit: ChatGPT, the Claude app, and the Gemini app carry no row in this table and are linked to their product pages instead. For the study category we made the inverse call, because Google documents NotebookLM operating limits as figures: fifty sources, five hundred thousand words. For a chat app, no vendor publishes a comparable specification.

Table inputs: five criteria and their weights

The weights are fixed and documented on this page: text leaderboard 45, score per dollar 20, context window 15, tooling maturity 10, access from Iran 10. Only ten of the hundred points sit on an ordinal scale. Any weight change logs a row in the changelog.

Two columns in this architecture depend on vendor data disclosure, and that dependency is what produces empty cells. Qwen3.8 Max and Gemini 3.1 Pro each carry two empty columns; that is a disclosure gap, not a model weakness, but it genuinely pulls the final score down.

Today output from the ranking architecture

Fable 5 holds first place on the Arena text leaderboard among 389 models at 1506, and it captures the entire forty-five-weight column. The same model loses the score-per-dollar column outright: output cost of $50 per million tokens, the most expensive row in the table, for a score-per-dollar ratio of 30.1, the floor of the pricing architecture.

Opus 5 sits twelve points lower and runs at half the operating cost. On a leaderboard whose entire top tier fits inside a roughly fifty-point band, a twelve-point gap is negligible. The optimal operating point for building on the API is therefore Opus 5; for a product subscription, the price column has no effect and the top of the table is the only criterion that matters.

One operational constraint has no column defined for it in this architecture: Fable 5 is the only model Anthropic itself flags as "slower" in its own table, because its adaptive thinking is always active with no lever to disable it. That latency is felt in real conversation, but the table does not measure it.

The boundary of this architecture: ChatGPT

ChatGPT is a product, so it was excluded from model rows under any configuration. The harder constraint is that it cannot even be documented: every path on openai.com returns a 403 status code to this server, re-tested on 12 August 2026, while that domain robots.txt does not forbid crawling.

One integration path stays open: developers.openai.com responds in full. GPT-5.6 Sol is listed there at $5 input and $30 output, a 1.05M-token context window, a 128K-token output ceiling, and a knowledge cutoff of 16 February 2026, and it carries a defined Arena row. On that basis, it occupies the third row of this table today.

Its WebDev row is registered exactly as gpt-5.6-sol-xhigh (codex-harness): the score belongs to the model inside OpenAI own dedicated coding agent, not the bare model. Cost and context window attach to the base model, and the score attaches to whatever configuration was actually tested. What remains unknown is which model powers the ChatGPT infrastructure today, what its plan pricing is, and whether it is reachable from Iran; all three live on the domain that does not respond. The OpenAI page documents this operational boundary in more detail.

Tolerance of the measurement method

The nine rows in this table fit inside a 71-point Elo band, from 1506 down to 1435. Min-max normalization expands that same 71-point band into a range from 78.6 to 29.1. As a result the score column overstates the real distance, and the last row is not "weak" — it is the last position in an architecture composed entirely of top-tier models.

A sharper example: GPT-5.6 Sol captures the context-window column at 1,050,000 tokens, with Kimi K3 Max close behind at 1,048,576. The 1,424-token gap, about a tenth of a percent, has no practical effect. Both absorb nearly the full weight of that column; but provisioning a model with a two-million-token window would drop both out of that column simultaneously, with nothing about their own specification having changed.

Two rows previously recorded as "unverified" were on this leaderboard from the start: GLM-5.2 Max at 33rd place with 1470, and DeepSeek V4 Flash at 83rd place with 1435. That was a gap in our measurement process, not in vendor disclosure. The deepseek-v4-pro row at 1458 on the same leaderboard belongs to a different model in the same family, not a configuration variant, so its figure was not borrowed.