What you call is not what answers

On 10 September DeepSeek published V4.1-Flash, with weights on Hugging Face under the MIT license. It is a 552-billion-parameter mixture-of-experts with an architecture the company calls Causal Encoder–Decoder: 40 layers, 8 billion active parameters when reading the prompt (prefill) and 16 billion when generating (decode). The attention cache takes a quarter of the GPU memory it would otherwise need, and an eighth when it goes to SSD.^1,2 All numbers are DeepSeek's; none had been replicated by a third party at closing.

Separating reading from writing is the week's technical idea: the prompt, which in agents is usually much larger than the answer, goes through half the parameters that writing uses. But the line that changes the bill for anyone already using DeepSeek is in the same notice, three paragraphs down: from 04:00 UTC on 14 September, calls to deepseek-v4-pro are served by V4.1-Flash and billed at its rate, “until V4.1-Pro launches”. deepseek-v4-flash and the experimental vision model were retired the same day, and the recommended name is now deepseek-flash, with no version.¹ Table 1 sums up who answers each name.

Pointing an old name at a new model is nothing new — every provider does it with aliases like latest. What changes here is that deepseek-v4-pro looked like a fixed name, with a version in the name itself, and now points to a model of another family, with a different size and a different price. For anyone with an agent in production, the consequences are concrete: the evaluation that approved V4 Pro no longer applies to what is answering; the invoice changes with no code change; and a log that stores only the name called does not say which model produced each answer.

On 10 September OpenAI opened the Agents API in public beta. The company describes it as a managed Codex harness: it handles session orchestration, context compaction when the window fills up, and recovery when something fails. A session is “a durable instance of an agent” that spans turns; progress arrives by streaming or webhook; the agent uses the customer's own tools and MCP servers; and it runs in a sandbox hosted by OpenAI, on the customer's infrastructure (type: self_hosted) or with a partner provider.^4,5 The beta label is in the code itself: the client.beta.agents namespace and the OpenAI-Beta: agents=v1 header.⁵

Until now, the harness — the loop that calls the model, runs the tool, keeps state, summarizes the context when it overflows and restarts after an error — was your team's code, written by hand or built on a framework. It is the part of the agent that shows least in demos and breaks most in production. Table 2 shows what moves to OpenAI and what stays with you.

The trade is less code for more dependency, and the dependency sits exactly on the hardest part to migrate: state. An agent whose history lives in a vendor session, compacted by the vendor's rules, does not move to another model by changing one line. Running the sandbox on your own server reduces the sensitive-data problem, not the state problem. Before adopting it, test two things with fake data: whether you can export the whole session in a format you can read, and what the agent still knows after compaction runs.

The operational week

The rest of the week's changelog is for people already in production. On 8 September, OpenAI shipped prompt-cache diagnostics — to find out why a call missed the cache you expected. On 10 September, project API keys started accepting an expiration date, and GPT-Live 1, the real-time voice model, left preview.⁴ None of it makes headlines. All of it cuts the bill or the risk for whoever already operates: a key that expires on its own is one less leaked key, and a cache that explains itself is money that stops leaking unseen.

On Monday, 7 September, Artificial Analysis published version 4.3 of its Intelligence Index — the number almost every model comparison cites, this one included. Three changes: Terminal-Bench 2.1 gave way to Terminal-Bench 4.0, weighted at 10%; τ³-Banking gave way to AutomationBench-AA, at 5%; and the private sets, which no lab can see before the test, went from 40% to 45% of the index.⁶ The ranking barely moved: GPT-6 Astra and Claude Fable 5.1 tied at 53, Claude Opus 5 at 51, GPT-5.6 Sol at 47. The scores did move. Figure 2 shows the same six models on both yardsticks.

Opus 5 and Astra swapped places (63.0 against 61.2 on v4.2; 51 against 53 on v4.3) and Fable 5.1 dropped into a tie with Astra — two-point differences, within the noise of an index of this kind. What the figure really shows is the scale: a model's absolute number on the index depends on the index version as much as on the model. Code 1 does the arithmetic of the drops.

# Índice de Inteligência da Artificial Analysis: os mesmos modelos, duas réguas
v42 = {  # leitura de 4 de setembro (v4.2)
    "Claude Fable 5.1": 65.7, "Claude Opus 5": 63.0, "GPT-6 Astra": 61.2,
    "GPT-5.6 Sol": 58.9, "GLM-5.3-Flash": 57.5, "GPT-5.6 Terra": 55.0,
}
v43 = {  # artigo de 7 de setembro (v4.3)
    "Claude Fable 5.1": 53, "Claude Opus 5": 51, "GPT-6 Astra": 53,
    "GPT-5.6 Sol": 47, "GLM-5.3-Flash": 42, "GPT-5.6 Terra": 42,
}
abertos = {"GLM-5.3-Flash"}

print(f"{'modelo':18}{'v4.2':>6}{'v4.3':>6}{'queda':>7}")
for m in v42:
    print(f"{m:18}{v42[m]:>6.1f}{v43[m]:>6}{v42[m] - v43[m]:>7.1f}")

fechados = [v42[m] - v43[m] for m in v42 if m not in abertos]
print(f"\nqueda média dos fechados:     {sum(fechados) / len(fechados):.1f}")
print(f"queda do aberto (GLM-5.3-Flash): {v42['GLM-5.3-Flash'] - v43['GLM-5.3-Flash']:.1f}")
print(f"modelos com índice >= 55 na v4.3: {sum(v >= 55 for v in v43.values())}")

modelo              v4.2  v4.3  queda
Claude Fable 5.1    65.7    53   12.7
Claude Opus 5       63.0    51   12.0
GPT-6 Astra         61.2    53    8.2
GPT-5.6 Sol         58.9    47   11.9
GLM-5.3-Flash       57.5    42   15.5
GPT-5.6 Terra       55.0    42   13.0

queda média dos fechados:     11.6
queda do aberto (GLM-5.3-Flash): 15.5
modelos com índice >= 55 na v4.3: 0

Closed models fell 11.6 points on average; the only open model in the sample, GLM-5.3-Flash, fell 15.5. One model supports no conclusion, but the direction is what you would expect from an index that gives more weight to what nobody saw before: a private test is less exposed to contamination, and whoever would have benefited from leaked test data loses more when it weighs less. It is a hypothesis, and the next version of the index will test it on its own.

The cut the panel used

The output's last line is the one that forces us to correct the panel. The “permissive open price” indicator read the lowest price among MIT or Apache 2.0 models with an index ≥ 55. On v4.3, no model, open or closed, reaches 55. Instead of lowering the cut to where the numbers fell — and lowering it again at the next version —, the indicator now follows the highest-ranked permissive open model on the index, whatever its score. Today it is GLM-5.3-Flash itself, at index 42, at US$ 0.50 per million output tokens. The value did not change; the method did. The series restarts in this edition, and the panel says so next to the number.

At GPT-6 Astra's launch, OpenAI published a table with its own model and competitors on Terminal-Bench 4.0: 57.9% for Astra, 55.8% for Claude Fable 5.1, 44.5% for Claude Opus 5 and 37.3% for GPT-5.6 Sol.⁸ That was the number no. 1 used. On the same 7 September when we closed that edition, Artificial Analysis published its own measurement, with a stated protocol: all 66 tasks, three runs, average pass@1.⁶ Figure 3 puts the two side by side.

None of the gaps is a scandal: with 66 tasks, each one is worth 1.5 points, and three of them cover the figure's largest distance. The point is another. A launch table puts together numbers from different origins without saying which is which: OpenAI's gave Fable 5.1 almost four points more and Opus 5 four and a half fewer than the independent measurement, and it does not say who ran those rows, or how. The order among the four did not change; the distance between them did.

There is a third source, which no. 1 did not read either: the official leaderboard, on the Terminal-Bench site, where each lab submits its own agent — Codex, Claude Code — with five runs per task. There, Fable 5.1 at maximum effort shows 57.9% since 3 September; Astra, 58.2%, in a row created the same day and updated on 10 September; Opus 5, 51.8%.¹⁶ They are the labs' own numbers, but on a common protocol and with a published confidence interval — about three points either way, which covers almost every gap in figure 3.

Three bugs ahead

On SWE-bench Science — the 119 real scientific-software bugs no. 1 introduced —, GPT-6 Astra entered the leaderboard on 10 September at 50.42%, running in Codex at maximum effort. The previous leader, Claude Opus 5, had 47.90%.⁹ In tasks: 60 against 57. By the noise arithmetic we did in no. 1, whose confidence interval is almost eighteen points wide in this range, three tasks do not separate two models. The leaderboard has a new first place; the measure has no new champion.

The lesson of both measures is the same as the previous movement's, by another road: a benchmark number compares only with another measured by the same group, with the same harness, on the same version. That is why the panel says, next to each reading, who measured it. In this edition the Terminal-Bench 4.0 ceiling moves to 59.1%, independent, and the SWE-bench Science one to 50.42%.

On 19 August, the Official Gazette published CNPq Call no. 29/2026, RHAE IA — Researcher in the Company. It is R$ 50 million from FNDCT for grants that bring master's and PhD researchers into research projects at micro and small companies and AI startups, up to R$ 300 thousand per proposal. Its object declares Brazil's AI Plan (Axis 4, AI for business innovation) and Nova Indústria Brasil; at least 30% of the money goes to companies in the North, Northeast and Center-West. Proposals closed on 9 October; the final result is scheduled for December.¹¹

The next day, 20 August, FACC published Selection no. 27/2026, the Macaíba AI supercomputer for LNCC: R$ 959,040,959.04.¹² That was the number no. 1 read as “PBIA tendered”. The CNPq call, published a day earlier with the same plan in its object, was left out. Code 2 redoes the sum.

# PBIA: o que já virou edital com valor fixo e prazo, em R$
editais = [
    ("2026-08-19", 50_000_000.00, "RHAE IA, CNPq/FNDCT, Chamada 29/2026"),
    ("2026-08-20", 959_040_959.04, "Supercomputador de IA, FACC/LNCC, Seleção 27/2026"),
]
total = sum(valor for _, valor, _ in editais)
for data, valor, nome in editais:
    print(f"{data}  R$ {valor / 1e6:>9,.2f} mi  {nome}")
print(f"{'total':10}  R$ {total / 1e6:>9,.2f} mi")
print(f"fração dos R$ 23 bi anunciados para 2024-2028: {total / 23e9:.1%}")

2026-08-19  R$     50.00 mi  RHAE IA, CNPq/FNDCT, Chamada 29/2026
2026-08-20  R$    959.04 mi  Supercomputador de IA, FACC/LNCC, Seleção 27/2026
total       R$  1,009.04 mi
fração dos R$ 23 bi anunciados para 2024-2028: 4.4%

R$ 1.009 billion, 4.4% of the R$ 23 billion announced for 2024–2028. Table 3 shows what goes into the sum and what does not. The BNDES and Finep AI investment fund stays out under the indicator's rule, even though it moved this week: on 14 September, the two picked Monashees as its manager, out of 17 proposals.¹⁵ The fund is expected to raise up to R$ 500 million, but the public commitment is a ceiling — up to R$ 125 million from BNDESPar and up to R$ 80 million from FNDCT/Finep —, and the money only exists once the fund is raised.

We corrected no. 1's reading from R$ 959 million to R$ 1.009 billion, and the panel shows the correction next to the number. It is also in the Method's corrections log. The error was one of search, not of rule: the call was on CNPq's site, published before that edition closed, and we did not find it.

In Congress, nothing. PL 2338/2023 remains in the Chamber “awaiting the opinion” of its rapporteur, deputy Aguinaldo Ribeiro (PP-PB), appointed in May 2025. The last movements were on 1 and 2 September, two attachments of other bills — PL 3060/2026 among them.¹³ In the week of 8 to 14 September, the record showed no movement.