MiMo-V2.6-Pro vs GPT-5.6 Sol: One Index Point Costs Fifteen Times More Per Task

On the Artificial Analysis Intelligence Index v4.3.2, read on 25 September 2026, GPT-5.6 Sol at maximum effort scores 47 and Xiaomi’s MiMo-V2.6-Pro scores 46. One point. The number that actually separates them is not the score and it is not the per-token rate — it is the cost of a single index task, which Artificial Analysis publishes as $1.99 for GPT-5.6 Sol and $0.13 for MiMo-V2.6-Pro. Fifteen times the money for one point. And as of this reading, the evaluator’s own page for that model carries a deprecation notice, because GPT-6 Sol has replaced it.

Xiaomi published MiMo-V2.6-Pro’s weights and technical report on 21 September 2026, four days before this reading; OpenAI shipped GPT-5.6 Sol on 9 July 2026. Both of the facts above are new since this pairing was last written up, and they pull in opposite directions. The deprecation notice says the comparison is against a model OpenAI has already moved past. The per-task figure says the thing that looked like a rounding error when both models were quoted at list price is, on a real workload, the whole decision. Which of the two matters depends on whether you are buying a model or replacing one.

What the two pages now say

GPT-5.6 Sol was released on 9 July 2026 as the flagship of the GPT-5.6 family. Its list price is $4.00 per million input tokens and $20.00 per million output on OpenAI’s first-party API, with a 90% cache discount, and the page records 90 million output tokens generated across the index run against a 88 million median. It measured 73 output tokens per second, which the same page calls faster than average against a 72 t/s median. Its index entry sits at rank 19 of 210 models in its class.

Xiaomi MiMo-V2.6-Pro is a sparse mixture-of-experts checkpoint at 1.02 trillion total parameters and 42 billion active, MIT-licensed, with a 1M-token context window and text, image, speech and video input against text output. It lists at $0.43 and $0.87 per million tokens with a 99% cache discount — a tenth of Sol’s input rate and a twenty-third of its output rate. It generated 140 million output tokens on the index run, and it measured 43.6 output tokens per second, which its page describes as notably slow against a 67.1 t/s median.

• Index task cost — MiMo-V2.6-Pro $0.13; GPT-5.6 Sol $1.99 — the gap is 15× • List price — MiMo-V2.6-Pro $0.43 / $0.87 per 1M; GPT-5.6 Sol $4.00 / $20.00 • Cache discount — MiMo-V2.6-Pro 99%; GPT-5.6 Sol 90% • Output tokens on the index run — MiMo-V2.6-Pro 140M; GPT-5.6 Sol 90M • Output speed — MiMo-V2.6-Pro 43.6 t/s, against a 67.1 median; GPT-5.6 Sol 73 t/s, against a 72 median • Context and weights — MiMo-V2.6-Pro 1M and MIT-licensed; GPT-5.6 Sol 1M and proprietary

The deprecation notice is the more consequential line

Artificial Analysis now opens the GPT-5.6 Sol page with a boxed notice: “This model is deprecated. We only continue performance benchmarking for the default 10k input token workload”, followed by the sentence that OpenAI has launched a newer model, GPT-6 Sol (max), and that it should be considered instead. The 47 on the index is therefore a reading of a model that is no longer the thing to buy. That does not make the reading wrong — it was measured on the same index, on the same day, at the same effort setting as the 46 — but it changes what the comparison is for. It is no longer “which flagship should I use”. It is “what does the top of the proprietary board cost, at the moment it was the top”.

That reframing matters more than it sounds, because deprecation notices on evaluator pages are one of the few places where a vendor’s product cadence shows up as a data change rather than a press release. The score is frozen; the model is not on offer. Anything you build against the 47 is building against a snapshot.

Why the per-task number is the one to model against

Per-token rates flatter cheap models and mislead on workloads that are not token-neutral. The index task cost is a different measurement: it is what the evaluator actually spent to get one answer out of each model, which folds in both the rate and the number of tokens the model chose to emit. MiMo-V2.6-Pro emits more tokens than Sol on the index run — 140 million against 90 million, about 1.6 times as many — and still costs $0.13 a task against $1.99. The rate advantage is roughly twenty-threefold on output; the verbosity penalty is 1.6×; the product is the fifteenfold per-task gap.

That is the arithmetic a cost model should be built on, and it is the arithmetic that is invisible if you compare $0.87 against $20.00 and stop. It also tells you what would break the conclusion. If your workload is dominated by very short outputs, Sol’s token count collapses toward yours and the twenty-threefold rate gap does most of the work, so the real multiple moves toward the rate gap rather than away from it. If your workload is dominated by long reasoning traces, MiMo-V2.6-Pro’s 1.6× verbosity compounds and $0.13 becomes the ceiling rather than the estimate. The published figure is a measurement of one benchmark’s shape, not a quote you can hand to finance.

The latency line is the honest counterweight

At 43.6 output tokens per second, MiMo-V2.6-Pro is slower than GPT-5.6 Sol measured on the same day at 73 — and slower than the models its own page compares it to, whose median is 67.1. Its page says so in as many words: “notably slow”. For a batch pipeline this is a throughput cost you can price. For an interactive agent it is a product decision, and fifteen times cheaper per task does not make a slow model fast. If your constraint is a per-turn budget rather than a monthly bill, the ratio that matters is 43.6 against 73, and MiMo-V2.6-Pro loses it.

Where this lands on a router

The useful shape of this pairing is not “pick one”. It is that both models answer the same index and the evaluator publishes a per-task cost for each, so the split can be made on measured cost rather than on vendor tiers. On OrcaRouter the catalogue runs to 200+ models behind one key at each provider’s list price with 0% markup — a vendor price cut reaches your invoice the same day, with no second contract to renegotiate — and the routing DSL lets a workload send the cheap, high-volume calls to one model and the hard ones to another by task rather than by brand. Automatic failover covers the case where the cheaper route is degraded. Xiaomi MiMo-V2.6-Pro is not on our routes — we do not list a model before a provider has onboarded it — so for that side of the comparison the honest answer is the vendor’s own API or whichever hosted platform you already have a relationship with.

What to do with this by next week

Two things changed on 25 September 2026 and only one of them is about price. GPT-5.6 Sol carries a deprecation notice on the index where it was measured at 47, which retires it as a purchasing decision and leaves it as a benchmark reference. MiMo-V2.6-Pro is measured at $0.13 per index task against Sol’s $1.99, which is fifteen times the money for one index point — and it is 43.6 tokens per second, which is slower than the model it is being compared to and slower than its own class median. Run the two figures against your own traffic before either number becomes a decision: the per-task cost is a measurement, not a rate, and the latency is a measurement, not an opinion. If your outputs are short and non-interactive, the case for the open checkpoint is stronger than the index gap suggests. If your outputs are long or your users are waiting, the fifteenfold saving buys you a slower product, and that is a trade only your workload can price.