IBM says open AI is 90% as good. The top 10 are all Chinese

Short answer: IBM CEO Arvind Krishna told investors that open-weight AI models are now "90 percent as effective" as frontier models at 10 to 25 percent of the token cost. We pulled 215 priced models with benchmark scores. The best open model hits 80% of the best closed one at 6.8% of the price, so he is a little generous on quality and way too modest on price. The part he left out: the 10 smartest open-weight models on the board are all Chinese, and IBM's own Granite scores 11.1 out of a top score of 57.6.

TL;DR: Claude Opus 5.5 leads the Artificial Analysis Intelligence Index at 57.6 for $8.00 per million blended tokens; the best open model, Xiaomi's MiMo V2.6 Pro, scores 46.3 for $0.54. Across every paid model listed in the past year, the median open-weight model costs $0.32 per million tokens against $2.00 for closed, which is 16 cents on the dollar. The best open model from any American lab, Nvidia's Nemotron 3.5 Super, scores 26.5, and the IBM Granite model on the same board scores 11.1, so the "open wins the enterprise" pitch Krishna sold on the Q3 call runs on Beijing's weights, not Armonk's.

Key numbers

What	Number	Source, checked Oct 26 2026
Krishna's claim: open vs frontier quality	"90 percent as effective"	The Register, Q3 call
Krishna's claim: open vs frontier cost	10% to 25% of token cost	The Register
Best closed model (Claude Opus 5.5)	57.6 index, $8.00 per M tokens	OpenRouter models API
Best open model (MiMo V2.6 Pro)	46.3 index (80%), $0.54 (6.8%)	OpenRouter models API
Chinese models in the open top 10	10 of 10	OpenRouter models API
Best American open model (Nvidia Nemotron 3.5 Super)	26.5 index	OpenRouter models API
IBM Granite 4.2 8B	11.1 index, $0.11 per M tokens	OpenRouter models API
Median price, paid models listed since Oct 2025	open $0.32 vs closed $2.00	OpenRouter models API
IBM Q3 2026 revenue	$17.10B	SEC EDGAR XBRL

Krishna is right about the price and soft on the gap

On IBM's Q3 2026 earnings call, Arvind Krishna laid out the pitch in one line: "Would you like something that's 10 percent of the cost if it's 90 percent as good as the best, or do you want to pay that 10 times higher cost?" Simon Sharwood at The Register wrote it up on Oct 22, and it hit the Hacker News front page this weekend.

We checked it against the market instead of a slide. OpenRouter's public models API lists every model it sells with live prices and, for 215 of them, the Artificial Analysis Intelligence Index (a composite benchmark score; higher is smarter). OpenRouter also links each open-weight model to its Hugging Face page, which is how we split open from closed.

The top of the board is Claude Opus 5.5 at 57.6. The top open model is MiMo V2.6 Pro at 46.3. That is 80%, not 90%. Price is the other way round: $0.54 against $8.00 per million blended tokens (three input tokens to one output, the usual chat mix) is 6.8%, cheaper than Krishna's best case.

So the real trade is 80% of the brains for one fifteenth of the bill. That is still a no-brainer for summarizing email and reading contracts, the exact jobs Krishna listed. It is a worse deal for the hard agentic work where a 10-point gap means the job fails.

The open-weight top 10 is a Chinese leaderboard

Here is the part that did not make the earnings call. Rank every open-weight model by that index and the top 10 are Xiaomi, Z.ai (twice), StepFun, Moonshot, MiniMax, Alibaba's Qwen, DeepSeek (twice) and Xiaomi again. Ten for ten, all from Chinese labs.

We already showed that 58% of new Hugging Face fine-tunes run on Chinese bases. This is the same story at the top end: when an American CEO says "open weights are irresistibly good," the models that make that sentence true are made in China.

Our read: Krishna knows it, and that is why the pitch is "on-prem" and "IBM software that eases deployment." He is not selling you IBM's model. He is selling you the racks, the mainframe Spyre cards and the Red Hat layer to run someone else's.

The best Western open model scores less than half

The best open model from an American lab is Nvidia's Nemotron 3.5 Super at 26.5. Thinking Machines' Inkling Small is 25.7, Meta's open Muse Glimmer 30B is 17.5, Google's Gemma 4 26B is 16.7 and OpenAI's gpt-oss-120b is 11.6.

IBM's own Granite 4.2 8B, released Aug 31, scores 11.1. That is 19% of Opus 5.5 and 24% of the Chinese leader. It is cheap at $0.11 per million tokens, but cheap and small is not the "90 percent as good" Krishna described to investors.

Lowkey the funniest receipt on the board: the company pitching open models to the Fortune 500 ships one that scores below Google's free Gemma 4 and OpenAI's year-old gpt-oss. To be fair, an 8B model is meant for narrow tasks, and Granite's whole sell is license cleanliness and indemnity, not leaderboard rank. That is exactly our point: what IBM sells is the paperwork, not the brains.

Follow the money: who wins, who loses

IBM booked $17.10 billion of revenue in Q3 2026, up from $16.33 billion a year earlier, per its SEC filings. The Register reports software grew 9% to $7.8 billion and infrastructure grew 5% to $3.7 billion, with mainframe revenue down 20% as the z17 cycle ages. An on-prem AI wave is how that mainframe line gets a second wind.

Winners if Krishna is right: Chinese labs, whose weights become the default enterprise brain for free; Nvidia, which sells the GPUs either way (our tracker has the cheapest on-demand H100 at $1.99 an hour across 44 listings, median $3.95); and IBM, Dell and HPE, who sell the racks.

Losers: the API margin at OpenAI, Anthropic and Google. Across every paid model OpenRouter listed in the past 12 months, the median open model costs $0.32 per million tokens and the median closed one $2.00. Every enterprise that moves its email and contract work on-prem takes that 84% spread with it.

The fight nobody wants to say out loud: Washington has spent two years trying to slow Chinese AI with chip rules, and the CEO of America's oldest enterprise IT brand just told Wall Street the cheapest way to run AI is on open models, the best of which are Chinese. We wrote about how "safety" gets used as a moat against open weights. Expect that argument to be aimed at exactly this setup next, with "data security" as the excuse. Our bet: within a year a US agency or a big bank publishes a ban on Chinese open weights in production, and IBM quietly starts benchmarking a Granite that can compete.

The cheap closed models are the real threat to Krishna's math

One more thing in the data cuts against the open-vs-closed framing. Anthropic's Claude Haiku 5.5 scores 43.4 at $0.20 per million tokens, and OpenAI's GPT-6 Luna scores 38.1 at $0.20. Those are closed models priced like open ones, and Haiku beats six of the open top 10.

So "10% of the cost" is not an open-weights exclusive anymore. The labs saw the same chart and cut prices; we covered that price war here. What open weights still buy you is the one thing a closed API never will: your data never leaves your building, and nobody can raise your price or switch your model off.

What this means for you

If you pay for AI by the token, check your bill against this table today. Most teams run every request through one frontier model when 80% of the work (summaries, extraction, classification, drafts) would come back the same from a model that costs 5% as much. Route the easy work to a cheap model, keep the frontier model for the hard 20%, and the bill drops by more than half. The full model list with prices and scores is in the data download on this page, and our LLM hub has the rest.

If you would rather we do it: the AI Cost Audit reads your usage and shows exactly which calls can move to a cheaper model and what you save. If it does not do what it promises, we keep fixing it at no extra cost until it does. More AI money pieces live under AI industry and on Tensorust News.

Sources

• The Register, Simon Sharwood, "IBM CEO says open weight models are now irresistibly good and cheap on-prem", Oct 22 2026
• OpenRouter models API (prices, Hugging Face links, Artificial Analysis Intelligence Index), pulled Oct 26 2026
• Artificial Analysis, Intelligence Index, via OpenRouter, Oct 26 2026
• SEC EDGAR, IBM quarterly revenue (XBRL), pulled Oct 26 2026
• Tensorust GPU price tracker, H100 on-demand listings, Oct 25 2026
• Hacker News discussion, Oct 26 2026