The floor fell to ten cents, and Macaíba slipped to November

On 7 October Anthropic launched Claude Haiku 5.5 (claude-haiku-5-5), with a one-million-token window and output of up to 128 thousand. The price has two bands: up to 100 thousand input tokens, US$ 0.10 per million for input and US$ 0.50 for output; above that, US$ 0.50 and US$ 2.50. Cache reads cost US$ 0.01 and US$ 0.05, and the 100-thousand limit counts every input token, cached ones included.^1,2 Haiku 4.5 cost US$ 1 and US$ 5. Anthropic says the new model comes out, on average, about 75% cheaper — the arithmetic already accounts for the new tokenizer, which spends slightly more tokens per task, and for 90% of Haiku 4.5 calls staying under 100 thousand tokens.¹

In the lower band, Haiku 5.5's list is identical, to the cent, to GPT-6 Luna's: US$ 0.10, US$ 0.01 and US$ 0.50.³ It is the second time in two weeks that two labs reach the same list. The difference is in the upper band, which Luna does not have up to 272 thousand tokens. Code 1 does the arithmetic per call.

# O degrau dos 100 mil tokens: custo de uma chamada sem cache, com 2 mil tokens de resposta (US$)
def haiku55(entrada, saida):
    pin, pout = (0.10, 0.50) if entrada <= 100_000 else (0.50, 2.50)  # a faixa vale para a chamada inteira
    return (entrada * pin + saida * pout) / 1e6

def haiku45(entrada, saida):
    return (entrada * 1.00 + saida * 5.00) / 1e6

def luna(entrada, saida):  # GPT-6 Luna, até 272 mil tokens de entrada
    return (entrada * 0.10 + saida * 0.50) / 1e6

saida = 2_000
print(f"{'entrada':>9}{'Haiku 5.5':>12}{'Haiku 4.5':>12}{'GPT-6 Luna':>12}{'5.5 / 4.5':>11}")
for entrada in (20_000, 60_000, 100_000, 100_001, 150_000, 250_000):
    a, b, c = haiku55(entrada, saida), haiku45(entrada, saida), luna(entrada, saida)
    print(f"{entrada:>9,}{a:>12.4f}{b:>12.4f}{c:>12.4f}{a / b:>11.0%}")
print(f"\num token a mais depois de 100 mil multiplica a chamada por {haiku55(100_001, saida) / haiku55(100_000, saida):.1f}")

  entrada   Haiku 5.5   Haiku 4.5  GPT-6 Luna  5.5 / 4.5
   20,000      0.0030      0.0300      0.0030        10%
   60,000      0.0070      0.0700      0.0070        10%
  100,000      0.0110      0.1100      0.0110        10%
  100,001      0.0550      0.1100      0.0110        50%
  150,000      0.0800      0.1600      0.0160        50%
  250,000      0.1300      0.2600      0.0260        50%

um token a mais depois de 100 mil multiplica a chamada por 5.0

Below 100 thousand tokens, Haiku 5.5 costs a tenth of its predecessor and exactly the same as Luna. One token above, the call becomes five times pricier, and between 100 and 272 thousand tokens Luna comes out five times cheaper by the list. Coding and document agents cross that limit easily: the history grows at every step, and the cache, which in no. 3 looked like the answer to the bill, counts toward the limit. Whoever migrates to Haiku 5.5 needs a simple control in the harness — compact or summarize the context before 100 thousand — or will pay the upper list without noticing.

Haiku 5.5's launch table shows the model at 39.2% on Terminal-Bench 4.0, against 0.0% for Haiku 4.5 and 16.4% for GPT-6 Luna, the latter run by Anthropic itself. Other numbers in the same table: 72.4% on the OSWorld 2.1 offline subset, 46.4% on FrontierCode 1.1 and 45.9% on Humanity's Last Exam without tools. All self-reported, with the method referred to the system card.¹ Figure 1 sets the Terminal-Bench numbers beside the official leaderboard.

The Luna row is the figure's most informative. Anthropic ran OpenAI's model on its own setup and got 16.4%; OpenAI submitted the same model to the leaderboard, with its own agent, and got 16.36%. When a lab's internal setup reproduces the number a competitor published, the calibration gains a point that does not depend on trusting whoever did it. It is an argument for Haiku 5.5's 39.2% — with the same caveat as no. 5, where calibration held for three models and failed on the fourth.

If the 39.2% holds, the bottom floor has changed function. A model at US$ 0.50 output that solves four in ten terminal tasks can do an agent's triage — read the error, try the obvious fix, run the test — and leave to Opus or Sol only what it cannot solve. By the figure's arithmetic, Haiku 4.5 was no good for that; Luna, at 16%, barely is. GLM-5.3-Flash, open and with no reasoning, comes close, at US$ 119 for the leaderboard's full run.⁴

On 6 October Mistral launched Mistral Large 4 as a public preview on its own API. By the company's numbers, it is a mixture-of-experts of about a trillion parameters (1.05 trillion, in the documentation) with 52 billion active, multimodal from training, with an optional reasoning mode and a one-million-token window. It costs US$ 1.36 per million input tokens, US$ 0.14 cached and US$ 4.18 output. The weights, the announcement says, come out “by the end of the month”. The text calls the model open-weight and names no license.^5,6

Artificial Analysis measured it the same day: 38 on the index, with 49 billion active parameters by its count, and a 50% launch discount for two weeks that only its article mentions.⁷ Thirty-eight is GPT-6 Luna's and MiMo-V2.6-Flash's score. Figure 2 shows what that costs.

By yardstick and by price, Large 4 does not compete with the Chinese open models: MiMo-V2.6-Pro scores eight points more for a fifth of the price. Mistral's argument is another, and it does not show on the index: a European lab, subject to European rules, with weights that a bank or a government will be able to run on its own server. For that to work, the two things the announcement left for later are missing — the weights and the license. Until then the model stays off the panel, and the question that matters to anyone considering it is not technical: under exactly what terms it will ship.

On 9 October, the SWE-bench Science maintainers published a correction: tasks 079, 104 and 107 of DeepSeek V4 Pro were rerun, and the model fell from 42.02% to 39.50% overall, from 73.16% to 71.50% on the private set and from 20.00% to 17.14% on the hard subset, according to the repository history.^11,12 With that, DeepSeek moved from third to fourth place, behind GPT-5.6 Sol. The top did not change: GPT-6 Astra at 50.42%, since 10 September.

# SWE-bench Science: o placar em tarefas, não em porcentagem (119 tarefas)
N = 119
placar = {  # % geral, leitura de 9 de outubro
    "GPT-6 Astra (max)":         50.42,
    "Claude Opus 5 (max)":       47.90,
    "GPT-5.6 Sol (max)":         40.34,
    "DeepSeek V4 Pro, corrigido": 39.50,
    "DeepSeek V4 Pro, antes":     42.02,
    "Kimi K3":                   35.29,
}
for m, p in placar.items():
    print(f"{m:28}{p:6.2f}%  {p / 100 * N:5.1f} tarefas")
print(f"\ncada tarefa vale {100 / N:.2f} ponto")
print(f"correção do DeepSeek: {(42.02 - 39.50) / 100 * N:.0f} tarefas; distância Astra–Opus 5: {(50.42 - 47.90) / 100 * N:.0f} tarefas")

GPT-6 Astra (max)            50.42%   60.0 tarefas
Claude Opus 5 (max)          47.90%   57.0 tarefas
GPT-5.6 Sol (max)            40.34%   48.0 tarefas
DeepSeek V4 Pro, corrigido   39.50%   47.0 tarefas
DeepSeek V4 Pro, antes       42.02%   50.0 tarefas
Kimi K3                      35.29%   42.0 tarefas

cada tarefa vale 0.84 ponto
correção do DeepSeek: 3 tarefas; distância Astra–Opus 5: 3 tarefas

The correction touched three tasks — as many as separate Astra from Opus 5 at the top. It is the clearest illustration of what no. 1 calculated with confidence intervals: with 119 tasks, rereading three of them changes the leaderboard's order. No. 1's SWE-bench Science chart showed DeepSeek at 42.0%, the correct number on that date; today it is another. That is why every number in this newsletter comes with the date it was read, and why a leaderboard that corrects backward, with its history open, deserves more trust than one that never changes.

The session for Public Selection no. 27/2026, the Macaíba AI supercomputer, was set for Thursday, 8 October. It did not happen: the FACC page began showing the session on 4 November, at 10:00, and an Errata 01 among the documents.¹⁴ According to Convergência Digital, which read the errata, the change went beyond the date. The technical proposal's weight stayed at 70%; price's fell from 30% to 10%; and the remaining 20% now rewards national production, under the Basic Productive Process. The report notes that the original tender's formulas still give 30% to price, and the rule is not clear.¹³

The changes point in one direction. Taking 20 points from price and giving them to national production favors whoever assembles servers in Brazil under the PPB, even if more expensive. Requiring offline operation and confidential computing defines a use — sensitive state data, running isolated — that the original tender did not spell out. And paying a third in advance, with a full guarantee, filters who can compete. None of these choices is wrong in itself; all make the contest pricier or narrower, and the price formula still contradicts the new weight. It is the kind of inconsistency that leaves room for challenges.

On the same 9th, ANPD published the second partial report of its AI regulatory sandbox, which follows three companies — Metatext, Synapse AI and Prevvine — with a supervision methodology designed by USP's Ciaam. The part on explainability is the most useful for builders: for two companies, the agency recorded that concrete evidence of the transparency mechanisms they described is still pending; for the third, it left for later tests the clarity of the alerts sent to security teams.^16,17 Describing a mechanism was not accepted as proof that it works. It is Brazil's rule on explaining automated decisions being written in practice, case by case, ahead of the law.

The PBIA indicator does not change: the R$ 959 million remain in a tender, only with a new date. RHAE IA's application deadline was 9 October. PL 2338/2023 had no movement in the Chamber in the week; the last action is still from 2 September.¹⁵