What Each AI Does In My September 2026 Stack

🔍 Read the full analysis: What Each AI Does In My September 2026 Stack on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on tech for your team

• Business-only prices and quantity discounts
• Tax-exempt purchasing
• Multiple users, one account, clear invoices

Create a free Business account →Free to register

As an affiliate, we earn on qualifying purchases.Hide

Thorsten Meyer said on September 29, 2026, that he uses Claude Opus 5.5 as his main model for building and the newly released Claude Opus 5.5 for detail work and review. His account, based on Artificial Analysis Intelligence Index v4.3.x results, argues that reported costs per task can differ sharply even among models with relatively close AI model scores.

Meyer assigns Opus 5.5 to features, APIs, multi-file work and refactors, typically at high effort. He reserves xhigh for architecture, migrations and trust boundaries. In the cited index, those settings score 54 and 56 respectively, at reported costs of $1.82 and $3.46 per task. Max effort scores 58 but costs $5.98 per task. Meyer says the two additional points from xhigh to max do not usually justify the higher cost for his work.

He uses GPT-6.1 Sol at high or xhigh to examine specific files or diffs and to review Opus’s output. The index lists scores of 50 at high and 51 at xhigh, costing $0.32 and $0.39 per task. Meyer says a second model from a different family is useful as a review pass, while cautioning that reviewers can still miss problems when they share a flawed specification.

The remaining models have narrower roles in his setup. He describes Astra and Fable as alternatives when his own tests favor them, Sonnet 5.5 as an option for scoped subtasks and documents, and Luna for classification, extraction and routing. A separate decision model, Jev, handles high-volume yes-or-no and routing judgments; Meyer says it cannot write sentences. The source does not provide Jev’s benchmark score or per-task price.

Cost Shapes the Model Assignment

The account illustrates a shift in how one developer chooses models: instead of relying on a single overall ranking, Meyer matches model and effort level to the task and its budget. In his cited figures, GPT-6.1 Sol xhigh costs $0.39 per task, compared with $3.26 for Astra and $7.63 for Fable 5.1. Those figures make frequent review passes more affordable in his workflow, though they do not establish that Sol performs as well on every kind of review.

The effort setting can also change the bill substantially. Meyer reports that Opus 5.5’s cost rises from $1.34 per task at medium to $5.98 at max, while its index score moves from 51 to 58. He presents high and xhigh as a practical balance for his development tasks. Readers considering a similar split would need to measure quality, latency and human review time on their own work before applying the same choices.

From Rankings to Task Costs

Meyer frames the account around a comparison of six models in Artificial Analysis Intelligence Index v4.3.x. The index is a general capability measure, he says, rather than a verdict on any particular workload. The reported top-setting scores range from 37 for GPT-6 Luna to 58 for Opus 5.5; costs per task range from $0.07 to $7.63. These are figures from the source account, not a guarantee of equivalent quality for a reader’s own tasks.

The source says Opus 5.5 was released on September 22, Sonnet 5.5 on September 28, and GPT-6.1 Sol on September 29. Fable 5.1 is listed with a September 1 release date, while Astra and Luna are dated September 3 and September 22. Meyer recommends shadow-testing before switching models. He also says one index point is within measurement noise, limiting the meaning of small score gaps.

“which model clears my quality bar at the lowest cost per task?”

— Thorsten Meyer

Benchmark Limits and Open Questions

The reported costs and scores come from a general benchmark and may not transfer to different prompts, tools or workloads. The source urges readers to shadow-test before switching and says a one-point score gap is within the noise. It does not provide independent validation of Meyer’s role assignments or results from his own task-by-task tests.

The source also reports first-token delays of 57 seconds at GPT-6.1 Sol high and 69 seconds at xhigh, which could limit its suitability for interactive use. Artificial Analysis had not published low or max settings for Sol at the time of the account. The material provided ends during an illustrative discussion of how human review time affects total cost, so it does not establish a measured estimate of that effect.

Test Before Changing the Stack

Meyer recommends running candidate models alongside the current setup on representative work before switching. That would let a team compare output quality, cost and response time on its own tasks, rather than treating index results as a direct forecast. The account gives no timetable for additional benchmark settings or a later update to the stack.

For his own workflow, Meyer says GPT-6.1 Sol is the routine second reviewer, while Astra or Fable serve as second opinions if Sol and Opus disagree. He says review failures should be returned to Opus with the failing case and supporting evidence. That describes his process; the account does not report an independently measured error rate or show how often disagreements change a decision.

Key Questions

What models does Meyer use for building and review?

He uses Claude Opus 5.5 for building, generally at high effort, and GPT-6.1 Sol at high or xhigh for detail work and review.

How much does GPT-6.1 Sol cost per task in the cited index?

The account lists $0.32 per task at high and $0.39 at xhigh. These are benchmark figures and may not match a particular user’s workload.

Does the index show which model is best for every task?

No. Meyer describes the Artificial Analysis Intelligence Index as a map of general capability, not a verdict on a specific workload. He recommends shadow-testing before a switch.

Why does Meyer not use maximum effort for every task?

He says higher effort can raise costs sharply for a limited score gain. In his figures, Opus 5.5 at max scores 58 and costs $5.98 per task, compared with 54 and $1.82 at high.

Source: ThorstenMeyerAI.com