The most expensive stopped being the best

On 22 September Anthropic launched Claude Opus 5.5 (claude-opus-5-5): US$ 4 per million input tokens and US$ 20 for output, against Opus 5's US$ 5 and US$ 25. There is a fast mode, up to 2.5 times quicker, at US$ 8 and US$ 40. The window is one million tokens and output reaches 128 thousand.^1,3 The same day, Artificial Analysis measured it at 58 on the index, at maximum effort — “the highest score we have measured, by several points”. At high effort, 54; medium, 51; low, 42.²

Six days later came Claude Sonnet 5.5, at the same price as Sonnet 5 — US$ 2 and US$ 10 —, at 56 on the index: second place, two points behind Opus. Artificial Analysis also recorded that it spends more output tokens per task than any model it has ever measured.^5,6 Figure 1 redoes the score-against-price chart with the week's launches.

In no. 1's chart, the highest dot sat at the right end, at US$ 50. Now it sits in the middle. The two US$ 50 models remain at 53; Opus 5.5 charges 40% of that and scores 58, and Sonnet 5.5 charges 20% and scores 56. The frontier-price indicator stays at US$ 50, because it measures the most expensive model, not the best one — and the gap between the two, which did not exist until now, becomes the panel's most interesting datum.

Two caveats change the reading. The first is effort: the same Opus 5.5 runs from 42 to 58 depending on how much it is allowed to think, and each extra point costs tokens. At high effort, 54, it already passes Astra at maximum. The second is consumption: the same per-token price as Sonnet 5 does not mean the same cost per task, and Sonnet 5.5 is the model that writes the most per task among all Artificial Analysis has measured. Anthropic says it costs “up to 30% less per task”; that is its own number, with no published method.^5,6

Also on 22 September, OpenAI launched GPT-6 Sol and GPT-6 Luna. Sol costs US$ 2 per million input tokens, US$ 0.20 cached and US$ 10 for output; GPT-5.6 Sol, which it replaces, cost US$ 4 and US$ 20. Luna costs US$ 0.10, US$ 0.01 and US$ 0.50, against the predecessor's US$ 0.20 and US$ 1.20.^7,9 Artificial Analysis measured both at the 5.6 generation's level: 48 for Sol (5.6 had 47) and 38 for Luna (as did 5.6).⁸

Price per token tells half the story. The other half is how many tokens the model spends to solve the same task. Artificial Analysis publishes the cost of running each model through the whole index, and figure 2 shows the bill per task: Sol fell 47%, Luna 61%.⁸

It is the most useful kind of price cut for anyone who already has a product: the same agent, the same measured quality, half the bill, with nothing rewritten but the model name. GPT-5.6 had come out in June; in three months, OpenAI delivered the same capability for half. If the pace holds, the cost of keeping an agent still — no vendor change, no prompt optimization — falls on its own, and the decision to migrate to a cheaper open model is less urgent than it looked in no. 3.

With a caveat that came three days later. On 25 September, OpenAI fixed an image-encoding bug that degraded Sol and Luna and recommended rerunning evaluations that use images.⁷ Whoever tested the two models with scanned documents, screens or charts between 22 and 25 September measured the bug, not the model. The lesson holds for any launch: a model's first week is also the first week of the service that serves it, and the two show up mixed in the evaluation.

At Opus 5.5's launch, Anthropic published 66.4% on Terminal-Bench 4.0, at xhigh effort; at max, per the system card, the model scores 64.8%, “within noise”.^1,10 Six days later, Sonnet 5.5 came with 70.6% — Sonnet 5 had 10.3% in the same table.⁵ Both numbers are self-reported, but they come with what most lack: five runs per task, a standard error of 2.5 to 2.6 points, the agent used (Claude Code), and a calibration — on Anthropic's setup Opus 5 scores 52.3%, and on the benchmark's official leaderboard, 51.8%.¹⁰ Figure 3 shows what was known about the test at the end of the week.

The calibration is the strongest argument for the numbers: if the internal setup reproduces the public leaderboard on an old model, it is reasonable to expect it to reproduce it on new ones. Reasonable, not guaranteed — and the official leaderboard did not yet have the new models to check. The panel records the ceiling as 70.6%, self-reported, and says so next to the number. It is an eleven-point jump over the previous week's independent 59.1%.

When another model answers

Sonnet 5.5's system card has a detail that does not appear in the table: the 70.6% was measured with safeguards on, and the requests they blocked were answered by a fallback model — 1.2% of requests, in 1.5% of runs. On Opus 5.5, 2.5% of requests were flagged, affecting 10% of runs.¹⁰ The rules are written down: on Opus 5.5, blocks for biological risk and for frontier-model development fall back to Opus 5; for cybersecurity misuse, to Opus 4.8. The swap applies in Anthropic's own products and to whoever opts into it on the API, and the company says the blocks are transparent — they do not covertly change responses.¹²

It is no. 2's theme by another road: the model you call is not always the one that answers. Here the swap is declared, optional on the API and rare. But it exists inside the benchmark number — and, for an agent working on security or writing kernels for AI accelerators, “rare” may be the whole task. Whoever builds in that area needs to test with the fallback on and off, and log when it kicked in.

The 22nd's release notes carry two lines that matter more than the benchmark table for anyone migrating. On Opus 5.5, adaptive thinking cannot be switched off. And a call with tool_choice set to any or tool — which forces the model to use a tool — returns a 400 error.⁴ Forcing a tool is a common pattern in structured extraction and in harnesses that chain steps. Whoever uses that pattern needs to rewrite the call before switching models, and whoever paid for short answers with no thinking will pay for whatever thinking the model decides to do.

On the other side, the cache. Opus 5.5's cache read costs US$ 0.20 per million tokens — 5% of the input price; on Opus 5 it was 10%. Sonnet 5.5 follows the same rule: US$ 0.10.³ Code 1 repeats no. 3's month of agent work — six billion tokens read, 240 million written — now with a 90% cache hit rate.

# O mesmo mês de agente do nº 3 (6 bilhões de tokens lidos, 240 milhões escritos), agora com 90% de acerto no cache
entrada, saida, acerto = 6_000, 240, 0.90   # milhões de tokens

modelos = {  # US$ por milhão de tokens: entrada, leitura em cache, saída (lista em 28/9)
    "Claude Opus 5":   (5.00, 0.50,   25.00),
    "Claude Opus 5.5": (4.00, 0.20,   20.00),
    "Claude Sonnet 5.5": (2.00, 0.10, 10.00),
    "GPT-6 Astra":     (10.00, 1.00,  50.00),
    "GPT-6 Sol":       (2.00, 0.20,   10.00),
    "MiMo-V2.6-Pro":   (0.435, 0.0036, 0.87),
}
def mes(pin, pcache, pout, h=acerto):
    return entrada * ((1 - h) * pin + h * pcache) + saida * pout

print(f"{'modelo':18}{'sem cache':>11}{'90% cache':>11}{'economia':>10}")
for nome, p in modelos.items():
    sem, com = mes(*p, h=0), mes(*p)
    print(f"{nome:18}{sem:>11,.0f}{com:>11,.0f}{1 - com / sem:>10.0%}")

o5, o55 = mes(*modelos["Claude Opus 5"]), mes(*modelos["Claude Opus 5.5"])
print(f"\nOpus 5.5 contra Opus 5, só pelo preço de lista: {1 - o55 / o5:.0%} mais barato")

modelo              sem cache  90% cache  economia
Claude Opus 5          36,000     11,700       68%
Claude Opus 5.5        28,800      8,280       71%
Claude Sonnet 5.5      14,400      4,140       71%
GPT-6 Astra            72,000     23,400       68%
GPT-6 Sol              14,400      4,680       68%
MiMo-V2.6-Pro           2,819        489       83%

Opus 5.5 contra Opus 5, só pelo preço de lista: 29% mais barato

With cache, Opus 5.5 comes out 29% cheaper than Opus 5 on the same load, against 20% without cache. Anthropic says the model costs 40% less “on typical workloads”.¹ The gap between the arithmetic's 29% and the announcement's 40% would have to come from the model spending fewer tokens to do the same work — which is possible, but is not measured anywhere public. And the opposite holds for Sonnet 5.5: the per-token price is Sonnet 5's, but it is the model that writes the most per task on the index. For both, the number that decides is the cost per task in your own agent, and only you measure that.

No lab published new top-tier weights between 22 and 28 September, and the top of open models on the Artificial Analysis index is still the previous week's: MiMo-V2.6-Pro at 46, GLM-5.3 at 45, Kimi K3 at 44, GLM-5.3-Flash at 42.¹⁵ Only two of them are under MIT. GLM-5.3, released in August, uses Z.ai's own license, which allows commercial use with one exception: whoever offers the model as a service and has combined revenue above US$ 10 billion over twelve months must go through a security review by the company.¹³ Kimi K3 also has its own license, as no. 1 reported. Table 1 sums up.

The GLM-5.3 clause reaches only those who sell the model as a service and earn above US$ 10 billion a year — a handful of companies in the world. Even so, the model stays out of the indicator, and the reason is practical: a legal department reads licenses in binary. MIT and Apache 2.0 skip the question “are we in the exception?”; any custom license forces someone to answer it, in writing, before the first deploy. The indicator measures the price of what can be used without that conversation. Whoever can have the conversation finds, in GLM-5.3, one point less than MiMo at five times the output price — which, in the first movement's chart, leaves it alone halfway.

Brazil

No new tender with a value and a deadline appeared in the week, and the PBIA indicator stays at R$ 1.009 billion. PL 2338/2023 remains without movement in the Chamber since 2 September.¹⁷ The two October deadlines were still standing at the week's end: RHAE IA proposals by the 9th and the opening of the Macaíba bids on the 8th.