My September 2026 AI Lineup: Opus, Sol, And Jev

🔍 Read the full analysis: My September 2026 AI Lineup: Opus, Sol, And Jev on ThorstenMeyerAI.com

Prime Big Deal Days · Oct 6–7Offer from Amazon

Get smart everyday buys delivered free — and shop member deals

• Fast, free delivery on millions of items
• Access to Prime Big Deal Days deals on October 6–7
• Prime Video, Amazon Music and more included

Start your free Prime trial →Free trial for eligible customers · Cancel anytime

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer published his September AI lineup on Sept. 29, naming Claude Opus 5.5 as his main model for building and newly released GPT-6.1 Sol for detailed investigation and review. The account matters as a practical comparison of model costs and assignments, but its rankings and prices are based on a third-party benchmark and Meyer’s own reported workflow.

Meyer says the lineup is shaped by a gap between benchmark scores and reported task costs. He cites the Artificial Analysis Intelligence Index v4.3.x, on which Opus 5.5 scores 58 at its top setting, compared with 51 for GPT-6.1 Sol at xhigh. The figures describe results on that index, which Meyer says is a map of general capability rather than a verdict on any reader’s workload.

In Meyer’s table, a task costs $5.98 with Opus 5.5 at max and $0.39 with GPT-6.1 Sol at xhigh. The same table lists Luna at $0.07 per task and a score of 37, and GPT-6 Astra at $3.26 per task and a score of 53. These cost-per-task figures are attributed to the benchmark data Meyer cites; the source does not specify a universal definition of a task that would make the figures directly transferable to every use.

Meyer assigns Opus 5.5 at high effort to routine development and xhigh to difficult work such as architecture, migrations and trust boundaries. He uses Sol at high or xhigh to examine details and review changes. He lists Sonnet 5.5 and Luna for narrower jobs, including documents, routine checks and classification, while Astra or Fable may provide another opinion when his preferred models disagree.

How Meyer Splits Building and Review

The account describes a workflow built around different models for creation and review. Meyer argues that using a separate model family to review Opus output gives him a second perspective, and says Sol’s reported cost makes it practical to run that review on each meaningful change. That is his rationale, not evidence that independent model review will catch every defect.

His comparison also shows how effort settings affect his reported costs. On the index figures he cites, Opus at high scores 54 for $1.82 per task, while xhigh scores 56 for $3.46. Max reaches 58 for $5.98. Meyer says the higher settings are for harder problems, while medium remains his everyday setting for documents and routine work.

For readers choosing models, the practical point is to compare results on their own tasks, alongside cost and speed. Meyer explicitly recommends shadow-testing before switching. His figures do not establish that the same model or setting will be the best choice for other users, or that benchmark cost estimates capture all human review time.

Benchmark Scores and Effort Settings

Meyer frames the change as a shift from choosing a single “smartest” model to asking which option meets a quality threshold for the lowest cost on a particular task. He reports that six models in his comparison fall within roughly 20 index points of one another, while their costs per task differ by roughly 100 times. Both are summary comparisons in his article, not findings independently verified here.

His table lists Opus 5.5 as released Sept. 22, Sonnet 5.5 on Sept. 28, Fable 5.1 on Sept. 1, Astra on Sept. 3 and Luna on Sept. 22. Meyer says GPT-6.1 Sol launched Sept. 29 at token prices of $2 per million input tokens and $10 per million output tokens. He reports that the index then showed three Sol effort levels: medium, high and xhigh.

For Sol, Meyer lists scores of 48 at medium, 50 at high and 51 at xhigh, with costs per task of $0.21, $0.32 and $0.39, respectively. He also reports first-token times of 5.3 seconds at medium, 57 seconds at high and 69 seconds at xhigh. The source says high and xhigh used fewer output tokens than the median for comparable models in the index, but does not provide a separate methodology for applying that comparison to ordinary use.

“The practical reading: Sol is not the model I ask to build. It is the model I can afford to run on everything.”

— Thorsten Meyer

What the Cost Figures Leave Open

The article does not establish how these benchmark results translate to a specific company’s codebase, documents or review process. Meyer says the index measures general capability and recommends testing models against the reader’s own workload. He also notes that one index point is within the noise, and that the index had not yet published low or max settings for GPT-6.1 Sol.

The reported latency at Sol’s high and xhigh settings may matter to users who need a quick interactive exchange. Meyer’s figures show waits of 57 and 69 seconds to the first token, respectively. The source does not give latency results across a broader range of real-world tasks or explain how often those delays would occur.

The supplied source text ends during a section about human review time and model costs. It begins an illustrative example involving a $1 model task but cuts off before completing the calculation. No conclusion from that unfinished example can be established. The article also does not provide full details of benchmark methodology, workload composition or later updates to its prices and scores.

Testing the Lineup on Real Work

Meyer advises readers to shadow-test models before switching, comparing their outputs on the work they actually need done. That recommendation points to the next useful step for anyone considering the lineup: check quality, cost and response time under the same requirements, then judge whether a separate review pass adds value.

The next benchmark developments will include any new Sol settings or revised index data, which could change the comparisons. The source gives no timetable for those updates. Meyer also does not announce a formal evaluation program or a forthcoming change to his lineup, so his account should be read as his current practice on Sept. 29.

Key Questions

Which models does Meyer use most?

He identifies Opus 5.5 for building and GPT-6.1 Sol for investigation and review. Jev handles high-volume yes-or-no and routing judgments, according to his account.

What does Meyer say GPT-6.1 Sol costs?

His cited index figures put Sol at $0.21 per task at medium, $0.32 at high and $0.39 at xhigh. Those estimates are tied to the benchmark and may not match a reader’s own usage.

Does the comparison prove which model is best?

No. Meyer says the index measures general capability, not performance on every workload, and recommends testing models on the work they will be asked to do.

Why does Meyer use more than one model?

He assigns models by task and cost, and says a different model family can provide a separate review of work produced by Opus. He cautions that a different reviewer can still share flaws in the same requirements.

Source: ThorstenMeyerAI.com