The claim, in his own framing
Krishna did not present 90 percent as a benchmark result. He presented it as a purchasing question on IBM's third-quarter earnings call on October 21, reported by The Register the next day. "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?" he asked investors, before narrowing the claim himself: some tasks do require the frontier model, he said, but the vast majority of routine enterprise work does not.
That narrowing matters, and it is the strongest part of his case. A claim about summarizing email and reading contracts does not need a model that wins a composite intelligence index. It needs a model that is good enough at those jobs, priced low enough that running it inside your own building beats renting the best. IBM sells the building: on-premises infrastructure, its Spyre accelerators, and the software layer that runs other people's models. The pitch and the product line point the same way, which does not make the numbers wrong. It makes them worth checking independently.
What the price-and-score board actually shows
OpenRouter is a useful place to check, because it lists prices and benchmark scores for the same models in the same feed. Pulled on October 26, its data gives 214 models with an Artificial Analysis Intelligence Index score once duplicate variants are removed. Counting as open only the models whose listings link to weights a buyer can download today, the top of the open board is Xiaomi's MiMo V2.6 Pro at 46.3. The top of the whole board is Claude Opus 5.5 at 57.6. The index-score ratio between them is 80.4 percent.
That is a ratio between two scores on one composite index. It is not a measure of how well either model summarises an email or reads a contract, and Krishna's 90 percent was not a benchmark result either. What the board can do is test the general shape of his claim on one defined snapshot, with the rules stated. On that test, the quality half of his sentence runs ahead of the best score a buyer can actually download today.
The exception sits one line above the open board, and it matters. Moonshot's Kimi K3.1 Pro scores 55.1 in the same pull, a 95.7 percent index-score ratio against Opus 5.5, and Moonshot describes the model as open-weight. Its OpenRouter listing, however, links to no downloadable weights as of October 26, so under the rule this check uses — currently downloadable open weights — it does not count yet. The earlier Kimi K3, at 43.6, does qualify and sits in the open top ten. The two are different models with different availability, and the distinction is the whole ballgame for an on-prem buyer: a model you can only rent from an API is not yet an on-prem option, whatever its licence promises.
On price, the listed API numbers run wider than he claimed. MiMo V2.6 Pro's blended price in the same pull is $0.54 per million tokens against Opus 5.5's $8.00, which is 6.8 percent, below the bottom of his 10 to 25 percent range. The medians point the same way: $0.38 for open models against $2.61 for closed ones across the scored, paid listings. These are unadjusted medians across scored, paid listings in the October 26 snapshot, not quality-matched prices or on-prem inference costs. And the 6.8 percent itself is a ratio of listed API prices. It is not the cost of running MiMo on your own racks, which is the purchase Krishna was actually pitching; on-prem costs depend on hardware, utilisation and the people running it.
The gap that remains is a quality gap, and its size depends on the work. An 11-point index difference is not a rounding error on hard, multi-step tasks, where a model that loses the thread loses the whole job. On the routine tasks Krishna listed, most buyers will never see the difference. Both halves of that sentence are in the data. Only the first half was in the headline number.
How we checked
We pulled OpenRouter's public models API (openrouter.ai/api/v1/models) on October 26, 2026, at about 14:00 UTC and archived the raw response. The pull lists 466 models; 214 carry an Artificial Analysis Intelligence Index after removing :free and :batch duplicate variants of the same models. A model counts as open here only if its listing links to weights downloadable from Hugging Face — currently downloadable open weights. That is an availability test, not a definition of open source, and downloadable weights do not by themselves establish unrestricted licensing; buyers must read each licence separately. Prices are an illustrative blend of three input tokens to one output token, recomputed from each listing's prompt and completion prices. Duplicate model identities are counted once, and the price medians cover the scored, paid listings (open n=112, closed n=102).
The top ten are all Chinese
Rank the open models by score and a second fact appears, one the earnings call did not mention. The top ten are Xiaomi (twice), Z.ai (twice), StepFun, Moonshot, MiniMax, Alibaba's Qwen and DeepSeek (twice). Ten out of ten are from Chinese labs. The highest-scoring Western qualifying listing in our snapshot was Nvidia's Nemotron 3.5 Super at 26.5, although its listed vision-language configuration complicates a strictly text-only comparison.
IBM's own entry is further down still. Granite 4.2 8B scores 11.1. That comparison needs a fairness note: Granite 4.2 8B is a small model built for narrow, license-clean enterprise work, not for topping a general leaderboard, and IBM's actual offer is the deployment stack and the indemnity paperwork around it. But that is precisely the point worth keeping. The sentence "open models are now 90 percent as good" is made true, to the extent the board supports it at all, by models IBM does not make and cannot sell you. What IBM sells is a place to run them, inside your firewall, under your data rules. For an American enterprise weighing Chinese weights in production, that distinction is the decision — and origin alone does not disqualify a model. It raises the questions that have to be answered before deployment, not after.
The one listing that could change the quality half
Return to the caveat, because it cuts in Krishna's favour. Kimi K3.1 Pro's 55.1 is the second-highest score in the entire pull, behind only Opus 5.5. If its weights ship in downloadable form, the best open model's index-score ratio would sit near 96 percent of the top score — closer to Krishna's figure, though a score ratio still would not by itself validate a claim about task effectiveness — and the argument would move entirely to price, where K3.1 Pro at $4.00 blended is half the frontier price rather than a tenth of it, and to data control. Until the weights ship, an 80.4 percent score ratio at 6.8 percent of the listed API price is the trade that actually exists today, dated October 26, 2026.
Cheap is no longer an open-weights exclusive
One more line in the same data complicates the framing for both sides. Claude Haiku 5.5, a closed model sold only through Anthropic's API, scores 43.4 at a blended $0.20 per million tokens, ahead of six of the open top ten. OpenAI's GPT-6 Luna scores 38.1 at the same $0.20. The frontier labs have already cut their small-model prices to open-model levels, which we have covered as those cuts landed.
So the durable advantage of open weights is no longer the token price. It is the two things an API cannot offer at any price: the data never leaves your building, and nobody can reprice, restrict or retire the model underneath your workflow. Those advantages are conditional, not automatic: they hold only where the model's licence permits the use, the deployment is genuinely local, and the organisation's own infrastructure is run well enough to keep them true. Krishna's 90 percent figure does not survive as a score ratio. The strategy underneath the claim does not need the figure to be exact, and the board suggests he could have made a cleaner case with the real numbers: an 80.4 percent score ratio, 6.8 percent of the listed API price, and a top ten that makes provenance a procurement question. Buyers should examine model provenance, licensing and deployment controls alongside benchmark scores and total operating cost — and test the models on the actual work, the email, the contract and the customer call, before believing any percentage, his or ours.
Sources
- [1] The Register, Simon Sharwood, IBM CEO says open weight models are now irresistibly good and cheap on-prem, October 22, 2026 — Krishna's earnings-call remarks and IBM's Q3 figuresRead source
- [2] IBM Q3 2026 earnings call transcript, October 21, 2026 (as reproduced by MarketBeat) — Krishna's remarks in full, including the 90 percent claim and the two or three racks estimateRead source
- [3] OpenRouter public models API — live prices, Hugging Face weights links and Artificial Analysis Intelligence Index; pulled by ALFW on October 26, 2026, at about 14:00 UTC, with the raw response archived in this story's working filesRead source
- [4] Tensorust, Open weight models vs closed: IBM says 90%, the data says 80%, October 26, 2026 — an independent pull of the same data with a different recency filter; its medians differ from ours and are not used hereRead source