Three Roles, One Workflow: My September 2026 AI Stack

🔍 Read the full analysis: Three Roles, One Workflow: My September 2026 AI Stack 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

Meyer says he uses Opus 5.5 at high effort for features, APIs and multi-file work, where the source lists a score of 54 and a cost of $1.82 per task. He raises it to xhigh effort for architecture, migrations and trust boundaries; that setting scores 56 and costs $3.46 per task in the cited index. He says max effort is rarely worth the added cost for his work.

For specific-file analysis and independent review, Meyer assigns GPT-6.1 Sol at high or xhigh effort. The index figures in his report put those settings at $0.32 and $0.39 per task, with scores of 50 and 51. He says Sol is very concise on the benchmark, but high and xhigh take 57 to 69 seconds to produce a first token, making those settings less suited to interactive use.

The report assigns Sonnet 5.5 to scoped subtasks and documents, and Luna to classification, extraction and routing. Meyer lists Astra and Fable as second opinions when Sol and Opus disagree. He also describes Jev as a decision model for high-volume yes-or-no and routing judgments, but the supplied material gives no benchmark figures for Jev.

Why Meyer Splits Building and Review

The workflow shows how a buyer might use benchmark and price data to decide which model handles each task, rather than selecting a single model for every job. In Meyer’s account, the relatively low reported cost of Sol makes a separate review pass practical for more work, while Opus remains his choice for building and difficult engineering decisions.

That division is a personal operating choice, not proof that the same allocation will improve results elsewhere. The source itself cautions that the index measures general capability rather than performance on an individual workload. A team would need to compare quality, latency and total review time on its own tasks before changing its model assignments.

The Benchmark Behind the Model Choices

The model scores and task costs cited in the report come from the Artificial Analysis Intelligence Index v4.3.x, unless otherwise stated. In the top-setting comparison, Opus 5.5 scores 58 at $5.98 per task, while Sonnet 5.5 scores 56 at $7.60. Fable 5.1 scores 53 at $7.63, Astra scores 53 at $3.26, GPT-6.1 Sol scores 51 at $0.39, and Luna scores 37 at $0.07.

Meyer says GPT-6.1 Sol was released September 29 at the same listed token prices as GPT-6 Sol: $2 per million input tokens and $10 per million output tokens. He reports that Sol’s medium setting scored 48 for $0.21 per task, matching the earlier model’s score at one-fifth of its reported $1.06 task cost. The source also gives token prices for the other models, but these rates and index task costs are separate measurements.

Effort settings change the comparison. For Opus 5.5, the cited score rises from 51 at medium to 58 at max, while reported task cost increases from $1.34 to $5.98. Meyer’s preferred high setting scores 54 for $1.82. He says Sonnet 5.5’s max setting costs $7.60 for a score of 56, compared with $2.74 and a score of 52 at xhigh.

““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, in the September 29 report

What the Benchmark Cannot Establish

The reported scores and costs do not establish how the models perform on Meyer’s full workload or on other users’ tasks. The source advises shadow-testing before switching, but provides no results from such a test. It also says one index point falls within measurement noise and that Artificial Analysis had not yet published Sol’s low- or max-effort results.

The report does not define the full methodology behind its task-cost estimates or provide a measured comparison of human review time and total project cost. It says a minute of human review could erase savings from cheaper model pricing, but labels its example illustrative, not measured. The supplied source text ends partway through that example, so it does not give a complete calculation.

How Teams Could Test the Split

Meyer recommends shadow-testing models on the intended workload before changing a production workflow. A meaningful comparison would need to track output quality, model cost, response time and the human effort required to catch or correct errors. His report does not specify a date for further benchmark updates or say whether he plans to publish results from a workload-specific test.

For now, the described stack is a snapshot dated September 29, 2026, based on the index data available to Meyer at publication. The roles he assigns may change as model settings, prices and benchmark measurements change; the source gives no confirmed schedule for a future revision.

Key Questions

What model does Meyer use for building?

He says he uses Claude Opus 5.5 at high effort for features, APIs, multi-file work and refactors, and xhigh for harder engineering problems.

Why does he use GPT-6.1 Sol for review?

Meyer says Sol can examine specific files and review changes at a reported $0.32 to $0.39 per task at high or xhigh effort. That is his workflow choice, not a guarantee of review quality on other projects.

Are the model scores proof of which model is best?

No. The scores are from the Artificial Analysis Intelligence Index v4.3.x, which Meyer describes as a general capability measure. The report recommends testing models against the workload where they would be used.

What remains unknown about GPT-6.1 Sol?

The source says the index had not yet published Sol’s low- or max-effort scores. It also does not provide results from a shadow test on Meyer’s own tasks.

Source: ThorstenMeyerAI.com