GPT-6 versus Claude Opus 5.5: ten days after the call to slow down, three new models in one afternoon

On Saturday 12 September 2026, Anthropic CEO Dario Amodei published a roughly 3,800-word essay on his personal site titled *We Must Pace the Frontier*. Its core argument: the AI industry should deliberately slow the rate at which model capabilities improve, so that safety work can keep up. Within hours, OpenAI CEO Sam Altman said he agreed. Elon Musk wrote that Dario was right.

On Tuesday 22 September, Anthropic released Claude Opus 5.5. A few minutes later, OpenAI released two new GPT-6 models, Sol and Luna.

Ten days.

This piece has two parts for two kinds of reader. The first is for anyone choosing a model: what the GPT-6 family contains, where Opus 5.5 stands, and how to read the comparison tables without being steered. The second is for anyone wanting to understand what the slowdown call actually commits to, and why ten days later it still has not been broken — even though three new models have shipped.

A note on sources. Most of the head-to-head numbers below were published by Anthropic itself in its Opus 5.5 announcement. We mark which figures are the vendor's and which come from independent measurement. And because the second half of this piece criticises both companies — with Anthropic as the one that made the call — we apply the same yardstick to both.

What the GPT-6 family contains

As of today, GPT-6 has three models built with the same training methods, differing in size and price:

• GPT-6 Astra — released 3 September, the flagship. $10 per million input tokens, $50 per million output tokens. It is the first model OpenAI itself rated Critical for cybersecurity capability; we covered it separately in GPT-6 Astra: skip the AGI claim, these three numbers are the ones that matter.
• GPT-6 Sol — released 22 September, the middle tier. $2 input, $10 output.
• GPT-6 Luna — released 22 September, the cheapest tier. $0.10 input, $0.50 output.

All three have a context window — the amount of text the model reads in one pass — of about 1.05 million tokens. Cached input is billed at a tenth of the normal rate. One clause is easy to miss: any request with more than 272,000 input tokens has the whole request billed at twice the input rate and one and a half times the output rate, not just the excess.

The previous generation, GPT-5.6, had three tiers named Sol, Terra and Luna. GPT-6 currently has no Terra and adds Astra at the top. So the name Sol does not mean the same position across generations: GPT-5.6 Sol was the flagship, GPT-6 Sol is the middle tier.

On capability, the independent benchmarker Artificial Analysis found the new Sol and Luna hold the composite intelligence score of their GPT-5.6 counterparts at roughly half the price. In other words, OpenAI's news on 22 September was not a smarter model but the same intelligence for half the money.

What Claude Opus 5.5 is

Opus 5.5 is the first model in the Claude 5.5 family, released 22 September. It costs $4 input and $20 output per million tokens, with cached input at $0.20. Against Opus 5 at $5 and $25, the unit price is 20% lower; Anthropic says total cost on a typical workload is about 40% lower because the model uses fewer tokens for the same work.

Anthropic describes Opus 5.5 as performing at the level of Claude Fable 5.1 — its own most expensive model — at a much lower cost. It also says Opus 5.5 scored best to date on its automated misalignment audit of nearly 2,000 scenarios, and that external evaluators including METR and Frontier Design tested it before release.

On Artificial Analysis's composite intelligence index, Opus 5.5 ranks first at 58 on maximum reasoning effort. GPT-6 Astra and Claude Fable 5.1 are tied at 53. GPT-6 Sol at maximum effort scores 48.

Head to head: how to read the numbers

The comparison table in Anthropic's announcement — that is, figures published by the vendor — gives these results for Opus 5.5 against GPT-6 Astra:

• Agentic coding at the command line (Terminal-Bench 4.0): Opus 5.5 66.4%, Astra 57.9%.
• Real-world office work across 44 occupations (GDPval-AA, a chess-style rating): Opus 5.5 1846, Astra 1542.
• Expert multidisciplinary questions, with tools (Humanity's Last Exam): Opus 5.5 67.7%, Astra 57.2%.
• Coding on real repositories (FrontierCode): Opus 5.5 54.4%, Astra 53.3% — a 1.1-point gap, within measurement error.
• Business workflow automation (AutomationBench, run and reported by Zapier): Astra 41.4%, Opus 5.5 40.0%.
• Agentic scientific research (Terminal-Bench-Science): Astra 64.6%, Opus 5.5 58.7%.

Three things follow from this table, in order of importance.

First, whoever publishes chooses the tests. This is Anthropic's table, so it contains the tests Anthropic chose. That does not make the figures wrong, but it does mean that anything Astra is good at which is absent from the table will not appear here. The independent 58 against 53 is the figure to lean on, and it still favours Opus 5.5 — just by less than the top rows of the vendor table suggest.

Second, nobody wins outright. Astra wins on business automation and scientific research; Opus 5.5 wins on command-line coding and office work. If your work sits where Astra is stronger, the aggregate tells you nothing useful.

Third, and most important for Vietnamese businesses: the contest between the two flagships is not the contest you are in. Look at price per million output tokens:

• GPT-6 Astra: $50
• Claude Opus 5.5: $20
• GPT-6 Sol: $10
• GPT-6 Luna: $0.50

Opus 5.5 is 2.5 times cheaper than Astra. But Luna is 40 times cheaper than Opus 5.5 and 100 times cheaper than Astra. Most of what small and mid-sized firms hand to a language model — sorting email, extracting fields from invoices, summarising meetings, answering questions from internal documents — does not need a flagship. Picking the wrong tier costs far more than picking the wrong vendor.

And as we wrote in our piece on the real cost of a model being the tokens it consumes, comparing price sheets is not enough. A cheap model that needs three attempts to get it right is no longer cheap. The only way to know is to run it on your own work — the end of this piece shows how.

The slowdown call: what it actually says

Back to the 12 September essay, because the way it was summarised in the press and what it actually proposes are some distance apart.

Amodei proposed three steps:

• One: embed independent evaluators inside the company, with permanent, employee-level access, to check adherence to safety commitments, report incidents and assess model alignment during training itself. Anthropic committed unilaterally to doing this immediately.
• Two: AI companies in democratic countries agree on common safety standards and limits on unchecked capability growth.
• Three: global coordination, including with China, towards international agreements. Amodei himself acknowledged this is unlikely soon, because countries have strong incentives to break such an agreement.

And there is a clarification every summary should keep: pacing does not mean halting training or halting technical progress. It means leaving enough time for alignment work, third-party verification and operational discipline to keep up with model capability.

Altman agreed the same day and committed OpenAI to giving independent evaluators employee-level access as well. Musk agreed. Meta CEO Mark Zuckerberg rejected the call.

Ten days later: words set beside actions

Fairness means starting with what was actually done:

• In August, following an incident during internal evaluation, OpenAI paused for two weeks reinforcement-learning training on models nearing deployment to harden its research environments, and put its largest planned training run on hold indefinitely pending further safeguards.
• On 18 September, Anthropic named Accenture as its first embedded evaluator, led by Faculty, Accenture's specialist AI business. Evaluators get access comparable to an employee's: observing training, following operational decisions, speaking directly with staff. Each company expects to invest at least $1 billion in this area over five years.
• Opus 5.5 was tested by METR and Frontier Design before release.

Those are real steps, and evaluators with standing access are more than this industry had before September.

Now set beside them what happened in the same ten days:

• On 18 September — the same day as the Accenture announcement — it emerged that Anthropic and OpenAI were seeking additional small data centres of 20 to 30 megawatts, because demand is outrunning construction of the gigawatt campuses. That is, looking for more compute sooner.
• On 22 September, three new models shipped in one afternoon.
• Opus 5.5 topped the independent index, five points above Fable 5.1 — Anthropic's own previous best.
• Both companies cut prices: Opus 5.5 by 20% per unit, Sol and Luna by roughly half against the previous generation.

Anthropic addresses the contradiction directly. Its announcement calls Opus 5.5 the first release since the call, and the company's argument is that a unilateral slowdown does not work, because pace is set by competition rather than by any one actor's wishes; pacing concerns the manner and rhythm of capability gains paired with proportional safety investment, not a ceiling.

The first part of that argument holds up. The part about Opus 5.5 being "at the level of Fable 5.1" should be checked against the numbers: in Anthropic's own table, Opus 5.5 beats Fable 5.1 on all nine benchmarks listed, and the independent index puts it five points higher. Some margins are small — under a point on chart reading, just over a point on computer use — but the direction is consistent: the capability frontier moved up, and the price of reaching it came down.

Why the call cannot yet be broken: it contains no number

This is the most important point, and it is not about anyone's hypocrisy.

Read the commitments closely and they fall into two kinds. The first kind concerns verification: letting evaluators in, giving them access, letting them report incidents. These have concrete actions attached — Accenture, a billion dollars, METR. The second kind concerns speed: slowing the rate of capability gains. These contain no number at all.

No minimum interval between releases. No maximum capability gain per quarter. No compute cap. No threshold which, once crossed, requires a halt and a public statement. A commitment without a number cannot be broken — and for the same reason cannot be verified. Three models in ten days did not violate the call, because the call never defines what counts as too fast.

Add to that its conditional structure. Step one is unilateral, but it is a verification step. Steps two and three — the ones that would actually slow things — both require others to join. And the argument that "a unilateral slowdown does not work" means each party slows only when the others do. When everyone waits for someone else to go first, nobody goes.

Even the verification step raises questions. On the same 18 September, more than 100 researchers including Geoffrey Hinton signed a public letter demanding evaluators be genuinely independent: not owned or governed by the AI company, without significant other business with it, and not paid in any way contingent on their findings.

Measured against those three criteria, the Accenture arrangement trips at least two: Accenture and Anthropic have run a joint business group since December 2025 to bring Claude to enterprise customers, and by Anthropic's own announcement the evaluation work is initially funded directly by Anthropic — longer term it would prefer pooled or government funding. The announcement also does not say whether evaluator reports will be published. None of this means the evaluations will be wrong, but it does mean outsiders do not yet have grounds to trust they will be right.

Put briefly: the slowdown call is not just words — it has produced action on verification. But on speed it contains no number to act on, so there it remains words. Until there is a public number outsiders can measure, the right planning assumption is that the pace will not slow.

What would make it a real constraint? Nothing elaborate: a published minimum interval between flagship releases; evaluator reports made public, including unfavourable ones; evaluators with no commercial ties to the company they assess; and a capability threshold above which release stops until the evaluators sign off. If those appear, we will rewrite this section.

What a business should do

The practical consequence of all this is simple: plan as if the model will change every month, because that is what is happening. Four concrete steps.

One — build your own test set. Take thirty to fifty real requests that went through your system in the last month, including ones it previously got wrong. For each, write down what an acceptable result looks like. When a new model appears, rerun the whole set and record the pass count and tokens consumed. It takes an afternoon to build the first time, then tens of minutes per rerun. It is the only thing that answers "should we switch", because every score table above measures someone else's work.

Two — start at the cheapest tier. Run the set on Luna, or another vendor's equivalent, first. Keep whatever passes there. Push only the failures up to the middle tier, then to a flagship. With a hundredfold price gap between the ends, routing to the right tier alone usually saves more than any Opus-or-Astra debate.

Three — keep the model swappable within an afternoon. Put every model call behind a thin interface of your own. Keep prompts in version-controlled files, not scattered through code. Use a vendor-specific feature only when the test set shows it is worth it. This is also the no-vendor-lock-in principle we apply to every system we build for clients — not because one vendor is more trustworthy than another, but because price and capability move too fast to bet on one.

Four — on security, do not wait for anyone to slow down. Flagship models can now find previously unknown vulnerabilities, and what holds that capability back is each vendor's own monitoring. The basic defensive work — closing ports open to the internet, patching edge devices, a callback rule before moving money — we still worth doing now. None of it depends on whether the slowdown call comes true.

Closing

On the head-to-head: Opus 5.5 leads on the independent index, Astra wins in specific categories, and for most businesses the most consequential model released on 22 September is probably Luna — because it makes a great deal of work cheap enough to automate.

On the slowdown call: it has created something real — evaluators present inside the company. But that is for knowing, not yet for slowing. The part about speed still has no number, so there is nothing yet to keep a promise against.

The question that works for every similar commitment in future — from whichever vendor — is: *which number would tell me this has been broken?* If the answer is none, plan as though it had not been said.

If you would like someone to build a test set on your own data, or design routing across model tiers so costs do not rise with every new release, get in touch or see KonexForge's AI & ML capability.