The strongest open model costs 87 cents

The MiMo-V2.6 repositories appeared on Hugging Face on the afternoon of 21 September, Brasília time; Xiaomi's announcement is dated the 22nd, already in Beijing.^1,2 There are two multimodal models: Pro, with 1 trillion parameters and 42 billion active per token, and Flash, with about 310 billion. Each shipped in two training variants, under the MIT license. Xiaomi says it also published the technical report, the training environment and the reinforcement-learning code — and it tells how much the reinforcement phase cost: six days, about 750 thousand trajectories, US$ 2.62 million for Pro and US$ 850 thousand for Flash.¹ These are the company's numbers, unaudited, but rare: almost no lab says what it spent on any stage.

In Artificial Analysis's independent measurement, Pro scores 46 on index v4.3 and Flash 38. The 46 puts Pro ahead of every open-weight model — GLM-5.3-Flash, which the panel followed since no. 2, has 42.^3,5 On Xiaomi's own API, outside China, Pro costs US$ 0.435 per million input tokens and US$ 0.87 for output; Flash, US$ 0.14 and US$ 0.28. It is the same price list as MiMo-V2.5, in force since May.⁴ Figure 1 repeats no. 1's figure — score against price — on the new yardstick.

Within the 46-to-47-point band, the output price runs from US$ 0.87 to US$ 20 — 23 times. But the output price is the wrong number for people running agents. An agent rereads its history, files and tools at every step; it reads far more than it writes. Code 1 does the arithmetic for one month of use with a ratio typical of that load — 25 tokens read for each token written — and with no cache, which is left for the fourth movement.

# Um mês de agente: 1.000 tarefas por dia, 200 mil tokens de entrada e 8 mil de saída por tarefa, sem cache
tarefas = 1_000 * 30
entrada = tarefas * 200_000 / 1e6   # milhões de tokens
saida = tarefas * 8_000 / 1e6

modelos = {  # índice AA v4.3, US$ por milhão de tokens: entrada, saída (preço de lista em 21/9)
    "GPT-6 Astra":      (53, 10.00, 50.00),
    "Claude Fable 5.1": (53, 10.00, 50.00),
    "Claude Opus 5":    (51,  5.00, 25.00),
    "GPT-5.6 Sol":      (47,  4.00, 20.00),
    "Grok 4.7":         (46,  2.00,  6.00),
    "MiMo-V2.6-Pro":    (46,  0.435, 0.87),
    "MiMo-V2.6-Flash":  (38,  0.14,  0.28),
}
base = modelos["MiMo-V2.6-Pro"]
custo_base = entrada * base[1] + saida * base[2]
print(f"{entrada:,.0f} M tokens de entrada, {saida:,.0f} M de saída por mês\n")
print(f"{'modelo':17}{'índice':>7}{'US$/mês':>11}{'% entrada':>11}{'× MiMo Pro':>12}")
for nome, (idx, pin, pout) in modelos.items():
    custo = entrada * pin + saida * pout
    print(f"{nome:17}{idx:>7}{custo:>11,.0f}{entrada * pin / custo:>11.0%}{custo / custo_base:>12.1f}")

6,000 M tokens de entrada, 240 M de saída por mês

modelo            índice    US$/mês  % entrada  × MiMo Pro
GPT-6 Astra           53     72,000        83%        25.5
Claude Fable 5.1      53     72,000        83%        25.5
Claude Opus 5         51     36,000        83%        12.8
GPT-5.6 Sol           47     28,800        83%        10.2
Grok 4.7              46     13,440        89%         4.8
MiMo-V2.6-Pro         46      2,819        93%         1.0
MiMo-V2.6-Flash       38        907        93%         0.3

With this load, 83% to 93% of the bill is input, for every model. A month of Astra costs 25 times a month of MiMo Pro; Grok 4.7, with the same score as MiMo, almost five. The index says Astra is seven points better. Whether seven points are worth US$ 69 thousand a month depends on what the agent does — and that is exactly the question the index does not answer. The agent numbers Xiaomi published, such as Pro's 72.6% on DeepSWE, are its own and have not been replicated.¹

On 21 September xAI launched Grok 4.7 — the announcement page is already signed SpaceXAI LLC. The price starts at US$ 2 per million input tokens and US$ 6 for output, the same as Grok 4.6, and there is a fast variant with twice the speed for twice the price. The model shipped the same day on the API, in Cursor, in third-party harnesses and on clouds.⁶ On Artificial Analysis it scored 46 — the same as MiMo-V2.6-Pro, at seven times the output price.¹⁵

The announcement's table puts its own model behind. On Terminal-Bench 4.0, Grok 4.7 shows 37.6%, tied with GPT-5.6 Sol (37.3%) and twenty points below Claude Fable 5.1 (57.9%).⁶ The page does not say where the numbers come from, but they match, one by one, the benchmark's official leaderboard, where each lab submits its own agent with five runs per task: Fable 5.1 by Anthropic, Sol by OpenAI, Grok 4.7 by xAI — the last in a row created on launch day, with 324 of 330 runs complete.¹⁶ That is the right way to cite a competitor: with the number it submitted itself, on a protocol anyone can check.

Not everyone did it this way. GPT-6 Astra's launch table, on 3 September, gave Fable 5.1 55.8% — a number that is not on the leaderboard, that OpenAI does not source, and that is the same one Anthropic publishes in its own tables, measured on its internal setup.^7,17 And Artificial Analysis, which runs everything on its own harness, measured 52.0%.⁵ Figure 2 brings the three measures together.

Six points between the highest and lowest measure of the same model are four tasks out of 66, and the leaderboard's own confidence interval covers almost all of it. Nobody is cheating: each measure comes from a different harness and number of runs. The lesson is about reading. When a launch table shows competitors, the question is who ran each row. When the answer is “the competitor itself, on the public leaderboard”, the comparison can be checked; when the table does not answer, the row is worth less than it looks. For the panel, nothing changes: the Terminal-Bench 4.0 ceiling stays at Astra's 59.1% measured by Artificial Analysis, above everything the official leaderboard had that week.

On 15 September Google announced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, real-time voice conversation models — the second one thinks before answering. For developers, both went “into rollout” on the Gemini API and Google AI Studio; for companies, into private preview in Gemini Enterprise. The post gives no API model identifier, no price and no benchmark of its own.⁸ It gives one sentence: Extended Thinking would have taken first place overall on Artificial Analysis's voice index, at 82.6.

Artificial Analysis confirms the 82.6. First place cannot be confirmed: the index page does not say when each model entered. OpenAI's GPT-Live 1 left preview five days before Google's announcement and today shows ahead, at 83.2 — on a single run, as the page itself warns. And on 14 October, after this week, xAI launched Grok Voice 3.0, which passed both.⁹ Figure 3 shows the board as it stood on this edition's publication date.

Two things hold for people building with voice. First: Extended Thinking scores 8.1 points above the base model, which says reasoning before speaking already pays off in conversation — but the post does not say how long the user waits for it, and in voice the delay is the product. Second: a model “in rollout”, with no published price, does not go into a budget. Until then, the number that matters is the one for the model you can buy today, not the one on the board.

Neither lab launched a model between 15 and 21 September — the launches came on the 22nd, and are left for no. 4. What shipped was platform. On 15 September, OpenAI gave administrators control over who creates API keys: service accounts only, keys in one's own project only, or no new keys; the organization's rule overrides the project's, and existing keys do not change.¹⁰ On 18 September, Anthropic made the diagnostics field, from the beta cache diagnostics, always present in the response — null when not requested —, and the Compliance API began returning Claude in Chrome session transcripts to Enterprise customers.¹¹

Together with the key expiration and the cache diagnostics OpenAI shipped a week earlier, the message is the same: both vendors are selling to people who already have an agent in production and need to know who spent what. Cache diagnostics is the more important of the two, for the reason code 1 already showed: an agent reads far more than it writes. And what it reads is, almost all of it, the same text as the previous call — the history, the instructions, the tools. If that prefix hits the cache, input comes at a discount. Code 2 shows how much.

# Preço efetivo de entrada conforme a taxa de acerto do cache (US$ por milhão de tokens)
precos = {  # entrada sem cache, entrada com cache (leitura)
    "GPT-6 Astra":   (10.00, 1.00),
    "MiMo-V2.6-Pro": (0.435, 0.0036),
}
print(f"{'acerto':>7}" + "".join(f"{m:>16}" for m in precos) + f"{'razão':>9}")
for acerto in (0.0, 0.5, 0.8, 0.9, 0.95):
    ef = {m: (1 - acerto) * cheio + acerto * cache for m, (cheio, cache) in precos.items()}
    a, b = ef.values()
    print(f"{acerto:>7.0%}" + "".join(f"{v:>16.4f}" for v in ef.values()) + f"{a / b:>8.0f}×")
for m, (cheio, cache) in precos.items():
    print(f"\n{m}: desconto do cache {1 - cache / cheio:.1%}; um erro de cache custa {cheio / cache:.0f}× um acerto", end="")
print()

 acerto     GPT-6 Astra   MiMo-V2.6-Pro    razão
     0%         10.0000          0.4350      23×
    50%          5.5000          0.2193      25×
    80%          2.8000          0.0899      31×
    90%          1.9000          0.0467      41×
    95%          1.4500          0.0252      58×

GPT-6 Astra: desconto do cache 90.0%; um erro de cache custa 10× um acerto
MiMo-V2.6-Pro: desconto do cache 99.2%; um erro de cache custa 121× um acerto

Without cache, Astra's input costs 23 times MiMo Pro's. With a 90% hit rate — reachable in an agent whose prefix does not change between calls —, 41 times. The gap between closed and open grows exactly where the agent spends, because Xiaomi charges 99.2% less for a cached token and OpenAI 90%. The other side of the same arithmetic is risk: on MiMo, a token that misses the cache costs 121 times one that hits. A small change at the start of the prompt — a timestamp, a session identifier, the order of the tools — knocks the hit rate down and multiplies the bill without anything in the code looking wrong. That is what cache diagnostics are for, and why they are worth more than they look.

Since no. 2, the open-price indicator follows the MIT or Apache 2.0 model ranked highest on the Artificial Analysis index. In this edition it moves from GLM-5.3-Flash, at US$ 0.50, to MiMo-V2.6-Pro, at US$ 0.87. The frontier price did not change — US$ 50, for Astra and Fable 5.1 —, so the ratio between the two falls from 100 to 57.5. Whoever reads only the number concludes the scissors closed. They did not: the open model went up in score and in price at once. Code 3 shows how much the reading depends on the rule.

# A largura da tesoura em 22/9 depende de qual aberto entra na conta
fronteira = 50.00  # US$ por milhão de tokens de saída: GPT-6 Astra e Claude Fable 5.1
abertos = {  # MIT ou Apache 2.0: índice AA v4.3, preço de saída na API do laboratório
    "MiMo-V2.6-Pro":   (46, 0.87),
    "GLM-5.3-Flash":   (42, 0.50),
    "MiMo-V2.6-Flash": (38, 0.28),
}
metodos = {
    "até o nº 1: o mais barato com índice >= 55":
        [m for m, (i, p) in abertos.items() if i >= 55],
    "o mais barato com índice >= 42":
        [m for m, (i, p) in abertos.items() if i >= 42],
    "o mais barato, sem corte":
        list(abertos),
    "desde o nº 2: o mais bem colocado":
        [max(abertos, key=lambda m: abertos[m][0])],
}
for nome, candidatos in metodos.items():
    if not candidatos:
        print(f"{nome:44} sem modelo")
        continue
    m = min(candidatos, key=lambda m: abertos[m][1])
    print(f"{nome:44} {m:16} US$ {abertos[m][1]:.2f}  razão {fronteira / abertos[m][1]:6.1f}")

até o nº 1: o mais barato com índice >= 55   sem modelo
o mais barato com índice >= 42               GLM-5.3-Flash    US$ 0.50  razão  100.0
o mais barato, sem corte                     MiMo-V2.6-Flash  US$ 0.28  razão  178.6
desde o nº 2: o mais bem colocado            MiMo-V2.6-Pro    US$ 0.87  razão   57.5

With the same three models and the same day, the ratio runs from 57.5 to 178.6 depending on the rule. None is the right one; each answers a question. The panel's rule answers this one: how much does the best permissive open model cost, relative to the frontier? The answer, this week, is that the best got pricier and closer. We keep the rule and show the arithmetic, so that whoever disagrees with it can redo it with their own.

Brazil: nothing new, two deadlines ahead

The PBIA indicator stays at R$ 1.009 billion: no new call with a fixed value and a deadline appeared in the week. The two deadlines that count were set for October — RHAE IA proposals by 9 October and the opening of the Macaíba bids on 8 October, later postponed. PL 2338/2023 remains stalled in the Chamber, awaiting an opinion; the record shows no movement since 2 September.¹⁴ The other two agent indicators do not change either: the Terminal-Bench 4.0 ceiling stays at 59.1% and the SWE-bench Science one at 50.42%.