The price list became the same; the bill did not

On 29 September, together with GPT-6.1 Sol, OpenAI added an Ultrafast mode to GPT-6 Astra: US$ 60 per million input tokens and US$ 300 for output.¹ With it, Astra now has four prices — batch or flex processing, at US$ 25 output; standard, at US$ 50; fast, at US$ 100 (called Priority until July); and Ultrafast.³ Figure 1 shows the ladder. OpenAI announced in the same post an Ultrafast for Sol, “in the coming days”.²

OpenAI is not the only one charging by the clock. Anthropic sells Opus 5.5's fast mode at US$ 40 output, twice the standard.⁵ DeepSeek, since August, charges half off peak — peak runs from 01:00 to 04:00 and from 06:00 to 10:00 (UTC) on weekdays, which in Brasília is 22:00 to 01:00 and 03:00 to 07:00.⁴ Nobody is charging for the model; they are charging for the queue.

For people running agents, this changes the design more than the bill. Most of an agent's work has nobody waiting: reviewing a repository overnight, classifying documents, generating reports. That work fits in batch, at half price. What has someone waiting — a conversation, a copilot in the editor — may be worth fast mode. Splitting the two queues in code, and sending each call to the right tier, is probably the cheapest cost optimization available today.

GPT-6.1 Sol (gpt-6.1-sol) came out on 29 September, for “complex coding and professional work at a lower cost than GPT-6 Astra”. It costs the same as GPT-6 Sol, US$ 2 input and US$ 10 output, with cache reads dropping to US$ 0.10, and it replaces the Sol launched a week earlier. It brings in beta a multi-agent mode, in which the API itself delegates parts of a request to subagents; the same day, the Agents API gained computer use.^1,2 Artificial Analysis measured it at 52, one point below Astra.⁷

This Tuesday, from 03:40 Brasília time, the official Terminal-Bench 4.0 leaderboard received a batch of submissions: 6.1 Sol, the previous Sol, Luna, Opus 5.5, Sonnet 5.5 and a few open models. 6.1 Sol solved 58.18% — Astra's number, to the hundredth. The leaderboard also publishes the cost of each full run, 66 tasks over five runs: US$ 634 for Sol, US$ 3.3 thousand for Astra.⁶ Code 1 divides the cost by the number of solved attempts.

# Placar oficial do Terminal-Bench 4.0: quanto custa cada tentativa resolvida (66 tarefas × 5 rodadas = 330)
placar = {  # modelo (esforço máximo): % resolvida, custo total da execução em US$, como o placar publica
    "Claude Opus 5.5":   (64.85, 4_700),
    "Claude Sonnet 5.5": (61.82, 7_300),
    "GPT-6 Astra":       (58.18, 3_300),
    "GPT-6.1 Sol":       (58.18,   634.24),
    "Claude Fable 5.1":  (57.88, 6_200),
    "GPT-6 Sol":         (49.39, 1_500),
    "GLM-5.3-Flash":     (35.76,   118.74),
    "GPT-6 Luna":        (16.36,    72.14),
}
print(f"{'modelo':18}{'resolvida':>10}{'execução':>10}{'US$/resolvida':>15}")
for m, (pct, custo) in sorted(placar.items(), key=lambda kv: kv[1][1] / (kv[1][0] / 100 * 330)):
    resolvidas = pct / 100 * 330
    print(f"{m:18}{pct:>9.1f}%{custo:>10,.0f}{custo / resolvidas:>15.2f}")

modelo             resolvida  execução  US$/resolvida
GLM-5.3-Flash          35.8%       119           1.01
GPT-6 Luna             16.4%        72           1.34
GPT-6.1 Sol            58.2%       634           3.30
GPT-6 Sol              49.4%     1,500           9.20
GPT-6 Astra            58.2%     3,300          17.19
Claude Opus 5.5        64.8%     4,700          21.96
Claude Fable 5.1       57.9%     6,200          32.46
Claude Sonnet 5.5      61.8%     7,300          35.78

Three readings. 6.1 Sol solves terminal tasks at the same hit rate as Astra for a fifth of the cost — and it improved 9 points over the Sol of a week earlier, spending less. Opus 5.5 solves the most and costs US$ 22 per hit, less than Fable 5.1, which solves fewer. And Sonnet 5.5, the cheapest Claude per token among the three, is the most expensive per hit on the board: the model that writes the most per task, as Artificial Analysis had already measured, turns a low price into a high bill. To choose an agent model, the right unit is neither the token nor the score: it is the cost per solved task, on your own work.

On 30 September, Google announced Gemini 4 Argon, in a post signed by Google DeepMind's Koray Kavukcuoglu. The price is “introductory”: US$ 2 per million input tokens and US$ 10 for output, with cached input at US$ 0.10, until 30 November. From 1 December, US$ 4, US$ 20 and US$ 0.20.^8,9 At announcement, access was only for cybersecurity defenders, through the Fairwind program; opening to paying API customers came later. The post's performance numbers are Google's — 77.9% on DeepSWE v1.1, first place on AutomationBench at 51.3% — with no method and no public leaderboard linked.⁸ Artificial Analysis measured it at 53 at high effort.⁷

The price is what stands out. US$ 2, US$ 0.10 and US$ 10 is the price list of Claude Sonnet 5.5, launched two days earlier, and of GPT-6.1 Sol, launched the day before.^3,5 And US$ 4, US$ 0.20 and US$ 20, Argon's price after November, is Claude Opus 5.5's list. Code 2 runs the last editions' month of agent work on the five lists.

# Três laboratórios, uma tabela de preço: o mês de agente dos nº 3 e 4 (6 bi lidos, 240 mi escritos, 90% em cache)
entrada, saida, acerto = 6_000, 240, 0.90

precos = {  # US$ por milhão de tokens: entrada, leitura em cache, saída
    "Claude Sonnet 5.5 (28/9)":        (2.00, 0.10, 10.00),
    "GPT-6.1 Sol (29/9)":              (2.00, 0.10, 10.00),
    "Gemini 4 Argon até 30/11":        (2.00, 0.10, 10.00),
    "Gemini 4 Argon a partir de 1/12": (4.00, 0.20, 20.00),
    "Claude Opus 5.5 (22/9)":          (4.00, 0.20, 20.00),
}
for nome, (pin, pc, pout) in precos.items():
    mes = entrada * ((1 - acerto) * pin + acerto * pc) + saida * pout
    print(f"{nome:33} US$ {mes:>6,.0f}/mês")

Claude Sonnet 5.5 (28/9)          US$  4,140/mês
GPT-6.1 Sol (29/9)                US$  4,140/mês
Gemini 4 Argon até 30/11          US$  4,140/mês
Gemini 4 Argon a partir de 1/12   US$  8,280/mês
Claude Opus 5.5 (22/9)            US$  8,280/mês

Three labs reaching the same list, to the cent, in three days says that list price is no longer where they compete in this band. When the per-token price is equal, the difference lies in what code 1 measured: how many tokens each model spends to solve the task. By that criterion, the three US$ 10 models do not cost the same — on the Terminal-Bench leaderboard, Sonnet 5.5 comes out almost eleven times more expensive per hit than 6.1 Sol. Equal prices also take away the buyer's shortcut of choosing by price: what is left is measuring.

And there is the deadline. Whoever builds an agent on Argon now pays Sonnet's list until November and Opus's from December, without changing a line. It is the reverse of what OpenAI did with Sol the week before: there, the same score got cheaper; here, the same model will get pricier on a set date. In both cases, the agent's budget depends on a vendor decision that does not show up in the code.

On 30 September, Anthropic announced the deprecation of Claude Sonnet 4.5, with API retirement on 30 November, and recommended migrating to Sonnet 5.5. Haiku 4.5 goes “not sooner than” 15 October.^10,11 On 1 October, OpenAI deprecated GPT-5.1, GPT-5.3-Codex and GPT-5.4-Nano, with shutdown on 1 April 2027, and the tts-1 and tts-1-hd voice models, on 6 January.¹² Code 3 measures the notice each gave.

# Quanto tempo cada fornecedor dá entre o aviso e o desligamento de um modelo
from datetime import date
avisos = [  # fornecedor, o que sai, aviso, desligamento
    ("Anthropic", "Claude Sonnet 4.5",               date(2026, 9, 30), date(2026, 11, 30)),
    ("OpenAI",    "o1, o1-pro, GPT-4, GPT-4 Turbo…", date(2026, 4, 22), date(2026, 10, 23)),
    ("OpenAI",    "GPT-5, o3, o3-pro",               date(2026, 6, 11), date(2026, 12, 11)),
    ("OpenAI",    "tts-1, tts-1-hd",                 date(2026, 10, 1), date(2027, 1, 6)),
    ("OpenAI",    "GPT-5.1, GPT-5.3-Codex",          date(2026, 10, 1), date(2027, 4, 1)),
]
for quem, o_que, aviso, fim in avisos:
    print(f"{quem:10}{o_que:33}{aviso:%d/%m/%Y} → {fim:%d/%m/%Y}{(fim - aviso).days:>5} dias")

Anthropic Claude Sonnet 4.5                30/09/2026 → 30/11/2026   61 dias
OpenAI    o1, o1-pro, GPT-4, GPT-4 Turbo…  22/04/2026 → 23/10/2026  184 dias
OpenAI    GPT-5, o3, o3-pro                11/06/2026 → 11/12/2026  183 dias
OpenAI    tts-1, tts-1-hd                  01/10/2026 → 06/01/2027   97 dias
OpenAI    GPT-5.1, GPT-5.3-Codex           01/10/2026 → 01/04/2027  182 dias

OpenAI gives about six months for text models and three for voice. Anthropic gave Sonnet 4.5 two months — a model from a year ago, launched in September 2025. Sixty days are enough to change a model name in a call. They are not necessarily enough to redo an agent's evaluation, adjust prompts and go through an approval cycle at a regulated company. Whoever builds on Claude needs to treat each model's launch date as the start of a countdown — and keep the evaluation ready to run on the successor.

On the bright side, the API began answering on its own a question that used to require reading documentation. On 1 October, Anthropic's Models API gained a new field, line; on 5 October, capabilities.thinking.types.disabled, which reports whether the model accepts thinking switched off.¹⁰ That is what no. 4's problem lacked: a harness can ask the API whether the new model accepts what the old one did, before breaking in production.

In no. 4, the argument for Sonnet 5.5's 70.6% on Terminal-Bench 4.0 was calibration: on Anthropic's internal setup Opus 5 scored 52.3%, and on the official leaderboard, 51.8%. Now the leaderboard has the new models, submitted by Anthropic itself, with the same agent (Claude Code), the same maximum effort and the same five runs per task.^6,13 Figure 2 puts both sides of each model together.

On Opus 5, Fable 5.1 and Opus 5.5, the two measures sit within 2.1 points of each other — inside each one's noise. On Sonnet 5.5, the gap is 8.8 points, larger than both intervals combined. Both measures are Anthropic's, and neither document explains the difference. The system card describes the internal setup: Claude Code in --bare mode, no internet access, each task's resources preloaded, and safeguards on with a fallback model.¹³ The leaderboard does not describe the submission's configuration.

The panel follows its method: the ceiling is the highest published number, with who measured it alongside. It stays at 70.6%, self-reported, and the indicator's note now shows the leaderboard's 61.8%. Calibration was the reason to take 70.6% seriously last week; it still holds for three Anthropic models and stopped holding for the fourth. For anyone choosing between Opus 5.5 and Sonnet 5.5 for a terminal agent, the difference matters: on the leaderboard, Opus solves more and costs less per hit.

Brazil: deadline week

No new tender with a value and a deadline appeared in the week, and the PBIA indicator stays at R$ 1.009 billion. The two deadlines that count fall in the coming days: Thursday, 8 October, was set for the opening of the Macaíba supercomputer bids — which ended up postponed, as no. 6 reports; on Friday the 9th, RHAE IA applications close. PL 2338/2023 remains without movement in the Chamber since 2 September.¹⁴