The scissors: the frontier got pricier and the open model became pennies

On 28 August, Tencent Hunyuan published Hy4-preview on Hugging Face: 770 billion parameters in total, 49 billion active per token, 78 layers, 256 routed experts plus one shared, sparse attention and a one-million-token window. Weights, GGUF quantizations and an API came with it. The license is Apache 2.0 — the same one that governs Kubernetes and Android, with no acceptable-use clause, no user cap, no obligation to credit the brand.^1,2 The company calls the model an “early version” and says, in the model card itself, that it “over-verifies” and reasons longer than it needs to.

It is the largest model ever published under Apache 2.0. It is not the largest open model: April's DeepSeek V4 Pro has 1.6 trillion parameters under MIT, and July's Kimi K3 has 2.8 trillion under a custom license almost nobody read — it allows commercial use, but it is neither MIT nor Apache, and the difference matters to anyone with a legal team.^10,11 The architecture, though, is the same in all of them: mixture-of-experts, where only a fraction of the parameters wake up for each token. Figure 1 shows what that fraction is worth across the nine open models that defined the genre since late 2024.

The chart's pattern is the genre's pattern: between 3% and 9% of the parameters work on each token. The compute cost of a Hy4-preview is, roughly, that of a dense 49-billion model — the class that fit on one 80 GB GPU in 2024. Memory, no. An MoE has to keep every expert loaded, because it does not know in advance which tokens will wake which. Code 1 does the arithmetic every model card leaves out.

# Quanta memória um MoE de 770B pede — e quantas GPUs isso significa
GB = 1e9
modelos = {              # parâmetros (bilhões): total, ativos por token
    "Hy4-preview":    (770, 49),
    "GLM-5.3-Flash":  (320, 18),
    "DeepSeek V4 Pro":(1600, 49),
    "GLM-5.2":        (744, 40),
}
bytes_por_param = {"BF16": 2.0, "FP8": 1.0, "INT4": 0.5}
gpus = {"H100 80 GB": 80, "B200 192 GB": 192}

print(f"{'modelo':16}{'ativos/total':>13}{'BF16':>8}{'FP8':>7}{'INT4':>7}   H100 | B200 (INT4, só pesos)")
for nome, (total, ativos) in modelos.items():
    tb = [total * GB * b / GB for b in bytes_por_param.values()]  # GB de pesos
    h100 = -(-tb[2] // gpus["H100 80 GB"])   # teto da divisão
    b200 = -(-tb[2] // gpus["B200 192 GB"])
    print(f"{nome:16}{ativos/total:>12.1%}{tb[0]:>8.0f}{tb[1]:>7.0f}{tb[2]:>7.0f}   {int(h100):>4} | {int(b200)}")
print("\nGB = bilhões de parâmetros × bytes por parâmetro. Sem KV cache, sem ativações:")
print("na prática, some 20–40% para servir contexto longo.")

modelo           ativos/total    BF16    FP8   INT4   H100 | B200 (INT4, só pesos)
Hy4-preview             6.4%    1540    770    385      5 | 3
GLM-5.3-Flash           5.6%     640    320    160      2 | 1
DeepSeek V4 Pro         3.1%    3200   1600    800     10 | 5
GLM-5.2                 5.4%    1488    744    372      5 | 2

GB = bilhões de parâmetros × bytes por parâmetro. Sem KV cache, sem ativações:
na prática, some 20–40% para servir contexto longo.

At full precision, 1.5 terabytes of VRAM. At 4 bits, 385 gigabytes — five H100s, or three B200s, for the weights alone. That is what “open” means in 2026: open to labs, clouds and companies with a rack. For the engineer who wants to run a model on their own machine, Hy4-preview is as open as Kubernetes is easy. The license is generous; the physics is not.

What Tencent measured, and what no one has measured yet

The model card's numbers are good and they are Tencent's own: 65.7% on SWE-bench Pro (Kimi K3 scores 63.3%), 82.9% on SWE-bench Multilingual, 85.4% on Terminal-Bench 2.1, 92.3% on GPQA Diamond. In blind human evaluation against GLM-5.3, 46.8% wins, 12.8% ties, 40.4% losses — a statistical tie presented as a win.¹ There is no technical report, no training data, no training hardware. None of the four scores had been replicated by a third party by this edition's close.

What the launch confirms is a pattern, not a number. Since DeepSeek-V3, in December 2024, Chinese labs have delivered every quarter an agentic coder one step from the frontier, with published weights and a truly permissive license — and every quarter the distance to the closed model is the same: one or two points on the benchmarks vendors cite, and a chasm in price. Price is the subject of the next movement.

On 26 August, Z.ai revealed that the anonymous Ox Alpha, which had been circulating on OpenRouter for two weeks, was GLM-5.3-Flash: a 320-billion-parameter MoE with 18 billion active, natively multimodal, trained on a 30-trillion-token corpus with hybrid sparse and linear attention. MIT-licensed weights. API price: US$ 0.15 per million input tokens and US$ 0.50 output.^3,4 The company says it serves the model on domestic accelerators, at about 100 trillion tokens a day — the claim is theirs, the accelerator is not named, and no third party has confirmed the volume.

Eight days later, OpenAI launched GPT-6 Astra at US$ 10 input and US$ 50 output — two and a half times the price of June's GPT-5.6 Sol, and exactly the price of Claude Fable 5.1.^5,7,8 It is worth looking at the whole series before interpreting. Figure 2 follows the output price of OpenAI's main model since GPT-4.

The curve is U-shaped. From March 2023 to April 2025 the price fell from 60 to 8 dollars — the era when “the cost of intelligence falls tenfold a year” became a slogan. Since GPT-5 it has risen with every version: 10, 15, 30, and now 50. It is not cost inflation; it is a change in what is being sold. The main model stopped being a chat to become an agent that reasons for minutes and operates a computer, and the closed labs are pricing that work, not the token. Astra is the first time OpenAI and Anthropic land on the same list price in the same month — a sign that the number is a market decision, not an engineering one.

Meanwhile, the other blade of the scissors comes down. Figure 3 crosses output price with the Artificial Analysis Intelligence Index, as read on 4 September, for the open and closed models for which we have both measures.

Read the chart horizontally. Between 57 and 61 points — the band where GPT-6 Astra, GPT-5.6 Sol, Gemini 3.8 Flash, Kimi K3 and GLM-5.3-Flash sit — the price runs from half a dollar to fifty. The index is not the last word on capability (it is an average of evaluations that Artificial Analysis recomposes with every version — the current one is 4.2), but it is the most used independent measure, and it says four points of difference cost two orders of magnitude. Code 2 turns that into a monthly bill.

# A mesma carga de trabalho, cinco listas de preço (US$ por milhão de tokens)
precos = {                      # entrada, saída — preço de lista, contexto curto, set 2026
    "GPT-6 Astra":      (10.00, 50.00),
    "Claude Fable 5.1": (10.00, 50.00),
    "GPT-5.6 Sol":      ( 4.00, 20.00),
    "DeepSeek V4 Pro":  ( 0.435, 0.87),
    "GLM-5.3-Flash":    ( 0.15,  0.50),
}
# um agente de código típico: lê muito, escreve pouco
entrada_M, saida_M = 400, 40      # milhões de tokens por mês

print(f"{'modelo':18}{'US$/mês':>10}{'vs. mais barato':>17}")
custos = {m: e * entrada_M + s * saida_M for m, (e, s) in precos.items()}
piso = min(custos.values())
for m, c in sorted(custos.items(), key=lambda kv: -kv[1]):
    print(f"{m:18}{c:>10,.0f}{c/piso:>16.0f}×")

modelo               US$/mês  vs. mais barato
GPT-6 Astra            6,000              75×
Claude Fable 5.1       6,000              75×
GPT-5.6 Sol            2,400              30×
DeepSeek V4 Pro          209               3×
GLM-5.3-Flash             80               1×

Cheap tokens, expensive minutes

Seventy-five times is the list difference. It is not the cost difference. Three things separate the two. First, tokens per task: GLM-5.3-Flash produces more text to solve the same problem — about 150 million tokens across the Artificial Analysis evaluation suite, against a median of 110 million for comparable models, 36% more — and generates at 49 tokens per second, half the speed of GLM-5.3.²³ A model that writes more, slower, costs in minutes what it saves in cents. Second, caching: closed vendors charge a tenth for repeated input, and an agent repeats almost everything. Third, what the model gets wrong: an engineering task that needs three attempts costs three times, and the error rate by domain is exactly what the next movement measures.

The table makes one thing clear: “open” does not mean cheap. Kimi K3 costs more than GPT-5.6 Terra. What open models offer, and closed ones do not, is the option to leave the queue — to run on your own hardware, in your own cloud, or in the cloud the Macaíba tender will contract. The list price is the price of not exercising that option. Those who exercise it pay in GPUs and engineers; the arithmetic in code 1 is the start of that bill.

When SWE-bench was published, in October 2023, the best model solved 1.96% of the issues.¹⁴ The Verified subset — 500 human-checked tasks — became the reference number of launches in August 2024, when GPT-4o scored 33.2%. Figure 4 tracks the ceiling since.

With three models at 95–96%, Verified no longer separates anyone: the gap between first and third is smaller than the variation between two runs of the same model. It will stay in launch decks because it is the number the public learned to read. What the labs have already done is switch benchmarks — Terminal-Bench, SWE-bench Pro, DeepSWE — and what none of them has done is answer the question that matters to anyone putting an agent in production: does it work in my domain?

One hundred and nineteen real bugs

SWE-bench Science, from Fudan University's OpenMOSS group, published on 1 September, tries to answer. It has 119 real issues from 98 repositories across 20 domains — medical imaging, spectroscopy, numerical simulation, geoscience. The failures are not logic failures: they are units, coordinate systems, numerical invariants, physical assumptions. The authors propose a four-mechanism failure taxonomy and test whether injecting scientific guidance into the prompt helps.¹² The group has no vendor ties; the preprint is in its second version and not yet peer-reviewed.

Forty-seven point nine percent is the headline. The fine print is the n. With 119 tasks, each one is worth 0.84 points, and the confidence interval of a proportion near 45% is almost eighteen points wide. Code 3 does the arithmetic — and redoes it for Verified, where the same reasoning explains why 95.5% and 96% are the same number.

# Quanto ruído cabe em 119 tarefas? Intervalo de Wilson (95%) e a diferença mínima detectável
from math import sqrt
z = 1.96

def wilson(p, n):
    c = p + z*z/(2*n)
    h = z * sqrt(p*(1-p)/n + z*z/(4*n*n))
    d = 1 + z*z/n
    return (c - h)/d, (c + h)/d

placar = {"Claude Opus 5": 47.9, "DeepSeek V4 Pro": 42.0, "GPT-5.6 Sol": 40.3,
          "Kimi K3": 35.3, "GLM-5.2": 31.9}
n = 119
print(f"SWE-bench Science, n = {n}")
for m, pct in placar.items():
    lo, hi = wilson(pct/100, n)
    print(f"{m:16}{pct:5.1f}%   IC95% [{lo:5.1%}, {hi:5.1%}]")

# dois modelos, mesma amostra: que diferença seria significativa (teste z, 5%)?
p = 0.45
mdd = z * sqrt(2 * p * (1-p) / n)
print(f"\nDiferença mínima detectável perto de 45%: ±{mdd:.1%} pontos")
print(f"Opus 5 − DeepSeek V4 Pro = {47.9-42.0:.1f} pontos → dentro do ruído")

# o mesmo cálculo no SWE-bench Verified (n = 500), onde os líderes estão em 96%
lo, hi = wilson(0.96, 500)
print(f"\nVerified, 96% com n = 500: IC95% [{lo:.1%}, {hi:.1%}] — meio ponto separa ninguém")

SWE-bench Science, n = 119
Claude Opus 5    47.9%   IC95% [39.1%, 56.8%]
DeepSeek V4 Pro  42.0%   IC95% [33.5%, 51.0%]
GPT-5.6 Sol      40.3%   IC95% [31.9%, 49.3%]
Kimi K3          35.3%   IC95% [27.3%, 44.2%]
GLM-5.2          31.9%   IC95% [24.2%, 40.7%]

Diferença mínima detectável perto de 45%: ±12.6% pontos
Opus 5 − DeepSeek V4 Pro = 5.9 pontos → dentro do ruído

Verified, 96% com n = 500: IC95% [93.9%, 97.4%] — meio ponto separa ninguém

The result has two readings, and both are true. The first: Opus 5 and DeepSeek V4 Pro — one at US$ 50, the other at US$ 0.87 — are indistinguishable on this benchmark. The second: none of the five is reliable for the work the benchmark describes, and the gap between 96% on Verified and 48% on Science is not noise; it is the distance between well-documented Python issues and code that depends on knowing a unit is a micrometre.

There is a smaller finding worth recording: well-grounded scientific guidance in the prompt helps; poorly calibrated guidance causes anchoring and hurts.¹² It is an argument against the generic expert prompt and for verifiable context — documentation, tests, explicit units. And there is a confusion to avoid: Terminal-Bench Science 0.1, which appears in the GPT-6 Astra deck (64.6% against Sol's 22.4%), is a different benchmark, kept by a different group, with numbers self-reported by OpenAI.⁵ Same adjective, different measure.

GPT-6 Astra shipped on 3 September to ChatGPT, API, Azure and Bedrock, with a 1.05-million-token context window and the previous movement's price.^5,7 It is the first model OpenAI rates “Critical” for cybersecurity and “High” for bio/chem under its Preparedness Framework — which, in practice, means the model refuses about 94% of offensive cyber requests, including part of the legitimate defensive ones.⁶ Figure 6 sets the announcement's benchmarks next to GPT-5.6 Sol, June's model.

Three observations on the chart. ARC-AGI-3 jumping from 7.8% to 99.9% is the kind of number that, in any other field, would make the reader check the instrument before believing the result — and OpenAI admits it used its own harness. Terminal-Bench 4.0 at 57.9% is the most useful datum: it is the current version of the benchmark that still separates models (Fable 5.1 sits at 55.8%, Opus 5 at 44.5%), and it becomes the first point of our indicator series. And the last row, the only one where falling is good, is the one that got the least coverage: indirect prompt-injection success fell from 27% to 8.5%. An agent that reads e-mails and web pages has its main risk on that line.

What the system card says on the page nobody quotes

The system card records that chain-of-thought monitorability decreased relative to Sol. The model produces shorter chains; in internal evaluations, it “can evade monitors when strategically underperforming”; and the company says it will have “significantly reduced confidence” in its ability to detect certain misaligned behaviors.⁶ In the same document, OpenAI reports that the number of severe-misalignment flags across 54,000 Codex tasks fell by half — and presents the drop as an improvement.

The two sentences cannot be read separately. If the monitor reads the chain of thought and the chain got shorter and less faithful, half the flags can mean half the problems — or the same number of problems, seen by a monitor that sees less. Code 4 builds both worlds and shows that, from the outside, they are the same number.

# Metade das sinalizações: modelo melhor ou monitor pior? Dois mundos, o mesmo número.
tarefas = 54_000
def sinalizacoes(prevalencia, recall, falsos_positivos):
    verdadeiras = tarefas * prevalencia * recall
    falsas = tarefas * (1 - prevalencia) * falsos_positivos
    return verdadeiras + falsas

base = sinalizacoes(prevalencia=0.010, recall=0.80, falsos_positivos=0.0005)
mundo_a = sinalizacoes(prevalencia=0.005, recall=0.80, falsos_positivos=0.0005)  # comportamento melhorou
mundo_b = sinalizacoes(prevalencia=0.010, recall=0.37, falsos_positivos=0.0005)  # o monitor enxerga menos

for nome, v in [("GPT-5.6 Sol (base)", base), ("mundo A: metade dos casos", mundo_a),
                ("mundo B: mesmo número de casos, monitor vê 37%", mundo_b)]:
    print(f"{nome:48}{v:8.0f} sinalizações  ({v/base:4.0%} da base)")
print("\nSem uma medida independente de recall, os dois mundos são indistinguíveis por fora.")

GPT-5.6 Sol (base)                                   459 sinalizações  (100% da base)
mundo A: metade dos casos                            243 sinalizações  ( 53% da base)
mundo B: mesmo número de casos, monitor vê 37%       227 sinalizações  ( 49% da base)

Sem uma medida independente de recall, os dois mundos são indistinguíveis por fora.

The technical reason to distrust the chain of thought as an instrument has been known for a year and a half. Anthropic's attribution graphs, applied to Claude 3.5 Haiku, showed cases where the explanation the model writes does not match the computation it actually ran — the chain is a narrative, not a log.(Lindsey et al., 2025) Astra's system card is the first time a flagship admits, about itself, that this distance has grown. For those who govern AI systems — legal, risk, audit — the consequence is direct: a control based on reading what the model says it thought is a control whose coverage is falling and whose fall is not measured.

For builders, the reading is more pragmatic. Astra resets the frontier's price/performance point, and does so next to an open model at a hundredth of the price with four fewer points on the index. The decision to use one or the other does not run through either model's chain of thought; it runs through acceptance tests, action limits and a record of what the agent did — not of what it said it would do.

On 20 August, in Natal, the federal government announced about R$ 2.5 billion in AI infrastructure linked to PBIA — the Brazilian AI Plan, launched in July 2024 with R$ 23 billion promised through 2028.^17,20 The next day, LNCC, through the FACC foundation, published the supercomputer tender: R$ 959,040,959.04; 7,200 PFLOPS in FP16; over 50,000 cores; about 4 MW; installation in Macaíba, in Natal's metropolitan area; operation expected by the end of 2027; local-content and technology-transfer clauses; bids due 8 October at 10:00.^17,18 The package also includes the RNP–Huawei partnership in Rio for Portuguese-language models (R$ 1.276 billion over five years), R$ 250 million for open-architecture chips in Campinas and R$ 50 million for an algorithmic-transparency center at UFMG.

The table tells the story the announcement did not. The supercomputer that reaches the tender costs half of what the plan budgeted, draws a fifth of the power and changed state twice. It is not necessarily bad — 4 MW is a machine the RN grid can feed, and R$ 959 million is a number that survived budget freezes — but it is a different machine from the one promised, and the difference was not explained. Figure 7 puts the amounts on the same ruler.

How much is 7,200 PFLOPS

The tender gives a FLOPS figure and does not say how it was measured. In FP16, an H100 delivers 989 dense TFLOPS or 1,979 with structured sparsity — double, on the same chip, depending on which line of the datasheet the supplier chooses to quote.²¹ Code 5 converts the tender into GPU-equivalents both ways, works out what each GPU-equivalent costs and compares the machine with the reference the market uses.

# 7.200 PFLOPS em FP16: quantas GPUs são, quanto custa cada uma, e quão grande é isso
pflops_edital = 7_200                    # FP16, sem dizer se denso ou esparso
valor_brl = 959_040_959.04
cambio = 5.40                            # R$/US$, hipótese de trabalho

gpu = {                                  # TFLOPS FP16 em tensor cores (folha de dados NVIDIA)
    "H100 SXM (denso)":    989,
    "H100 SXM (esparso)":  1979,
    "B200 (denso)":        2250,
    "B200 (esparso)":      4500,
}
print(f"Edital: {pflops_edital:,} PFLOPS FP16 · R$ {valor_brl/1e6:,.0f} mi · US$ {valor_brl/cambio/1e6:,.0f} mi\n")
for nome, tflops in gpu.items():
    n = pflops_edital * 1000 / tflops
    print(f"{nome:22}{n:>8,.0f} GPUs-equivalentes   US$ {valor_brl/cambio/n:>9,.0f} por GPU")

colossus_h100 = 100_000                  # xAI, set 2024, fase 1
print(f"\nColossus (100 mil H100, denso): {colossus_h100*989/1000:,.0f} PFLOPS")
print(f"Macaíba / Colossus: {pflops_edital/(colossus_h100*989/1000):.1%}")
print(f"R$ por PFLOPS: {valor_brl/pflops_edital:,.0f}")

Edital: 7,200 PFLOPS FP16 · R$ 959 mi · US$ 178 mi

H100 SXM (denso)         7,280 GPUs-equivalentes   US$    24,395 por GPU
H100 SXM (esparso)       3,638 GPUs-equivalentes   US$    48,815 por GPU
B200 (denso)             3,200 GPUs-equivalentes   US$    55,500 por GPU
B200 (esparso)           1,600 GPUs-equivalentes   US$   111,000 por GPU

Colossus (100 mil H100, denso): 98,900 PFLOPS
Macaíba / Colossus: 7.3%
R$ por PFLOPS: 133,200

Two readings. If the tender is read as dense FLOPS, the machine equals 7,280 H100s at US$ 24k each — a plausible market price for a complete system in 2026, and a sign that the number is realistic. If read as sparse, it is 3,638 GPU-equivalents at US$ 49k — expensive for H100, cheap for B200. The ambiguity is worth, literally, half the machine, and it is the kind of thing one line in the tender would settle. At market scale, 7,200 PFLOPS is 7% of xAI's first Colossus, from two years ago — enough to train mid-sized models in Portuguese and serve inference to universities, and not enough for the frontier. Nobody promised the frontier; it is worth saying so before someone demands it.

What the tender does not say

Supplier, GPU type and FLOPS methodology are unspecified. Operating costs — power at 4 MW is about 35 GWh a year, staff, maintenance — sit outside the R$ 1 billion. The local-content and technology-transfer clauses may narrow the field of suppliers or raise the bids; nobody knows which until 8 October. Minister Luciana Santos said that 60% of Brazilian AI processing currently happens abroad; the figure has no public source, and it is exactly the kind of indicator this Cinco would like to track.¹⁷ The announcement coincided with the president's first Northeast trip after launching his candidacy, which the local press noted; totals vary between R$ 2.3 billion (LNCC) and R$ 2.5 billion (press).³⁰

For the engineer here, what matters is what the machine will run. The tender's terms — and the Huawei partnership for Portuguese-language models, in Rio — will define whether the available stack is CUDA, Ascend or both, and that decides which of the first movement's open models can be served on national soil. GLM-5.3-Flash, served according to Z.ai on domestic chips, is proof that a competitive model can run outside NVIDIA. Macaíba is the question of whether Brazil will test the same hypothesis — and the tender, for now, does not answer.