My September AI Stack: A Place For Building, Digging, And Decisions

🔍 Read the full analysis: My September AI Stack: A Place For Building, Digging, And Decisions on ThorstenMeyerAI.com

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

Get the latest gadgets 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 a September 29 account of his AI model stack, assigning Claude Opus 5.5 to building and newly released GPT-6.1 Sol to detailed analysis and review. The comparison says six models cluster within about 20 points on Artificial Analysis’ Intelligence Index, while their listed cost per task varies by roughly 100 times, shifting the author’s choice from highest score alone to the lowest cost that meets a task’s quality bar.

Meyer’s table draws on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a general capability measure rather than a verdict on any individual workload. At each model’s listed top setting, Opus 5.5 scores 58 and costs $5.98 per task; GPT-6.1 Sol at xhigh scores 51 and costs $0.39. GPT-6 Luna scores 37 and costs $0.07. The source recommends testing models against a team’s own work before switching.

The proposed division of work gives Opus 5.5 high as the default for features, APIs, refactors and multi-file changes, at 54 index points and $1.82 per task. Meyer reserves xhigh, at 56 points and $3.46, for work such as architecture and migrations. He assigns Sol high or xhigh to investigating files and diffs and reviewing changes, at $0.32 to $0.39 per task. Astra or Fable serve as alternative opinions when models disagree; Sonnet 5.5 and Luna cover scoped or routine jobs.

The source’s effort-level comparison says the setting can change costs substantially. Opus 5.5 rises from $1.34 per task at medium to $5.98 at max, while its index score moves from 51 to 58. Sonnet 5.5 at max is listed at $7.60 per task for a score of 56; its high setting scores 47 at $1.08. Those figures describe the benchmark’s measured tasks and settings, not guaranteed costs for a reader’s own prompts.

Why Cost Shapes Model Roles

The comparison offers a practical way to think about model selection as a cost and quality trade-off. If a lower-priced model clears a task’s quality bar, it may be affordable to use it more often, including for an additional review pass. Meyer says his use of a separate model family to check Opus output is affordable at Sol’s listed price. That is his workflow rationale, not evidence that a second model will catch every error.

The stakes are clearest for teams that run large numbers of model tasks. A difference in per-task price can compound across repeated coding, extraction or routing jobs. But the index score and benchmark cost do not show whether the models meet a particular organization’s requirements, how much human checking its work needs, or the full cost of operating a workflow. Meyer also cautions that saving on model tokens can be outweighed by additional human review time; the example he introduces is illustrative rather than measured.

The review advice also sets limits on what model choice can accomplish. Meyer says increased effort cannot supply missing requirements, and a different reviewer may share the same flawed specification. He treats passing tests as insufficient on its own for approval to ship. Those points make the proposed stack a workflow recommendation as well as a price comparison: model review is one check, with evidence and human judgment still needed.

The September Model Comparison

The article frames its comparison around a shift over roughly four weeks: six models, it says, now sit within about 20 index points of one another, while cost per task differs by roughly 100 times. The supplied table lists Opus 5.5, Sonnet 5.5, Fable 5.1, GPT-6 Astra, GPT-6.1 Sol and GPT-6 Luna. Their scores and prices are tied to specific effort settings and to the index’s tasks.

According to Meyer, GPT-6.1 Sol launched September 29 at the same listed token rates as its predecessor: $2 per million input tokens and $10 per million output tokens. The source says Artificial Analysis listed three effort settings for Sol on launch day. Its medium setting scored 48 at $0.21 per task; high scored 50 at $0.32; xhigh scored 51 at $0.39. Meyer reports that medium matches the earlier GPT-6 Sol’s score of 48 at one-fifth of its $1.06 per-task cost.

The listed token prices differ by model. Opus 5.5 is shown at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. Fable and Astra are listed at $10 and $50; Luna at $0.10 and $0.50. Per-task prices in the source are separate benchmark observations, so token rates alone do not determine the cost of a task.

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

— Thorsten Meyer, in the September 29 article

What the Index Cannot Establish

The index is not a direct test of a reader’s own workload. The source provides no independent assessment of its methodology, uncertainty ranges or how much the scores vary between runs. Meyer says a one-point difference is within the noise, so small score gaps should not be treated as definitive rankings. He also says low and max settings for GPT-6.1 Sol had not yet been published when he wrote.

Sol’s reported high and xhigh settings take 57 to 69 seconds to produce a first token in the index, according to Meyer, which may limit their suitability for interactive use. The source does not specify how these timings will compare in other systems or under different conditions. Its claim that Sonnet 5.5 at max generated about 193,000 output tokens per task is attributed to Artificial Analysis; the comparison does not establish that the same output volume will occur in ordinary use.

The article does not report a controlled comparison of the author’s full workflow, error rates, review accuracy or total human time. Nor does it establish that any model is the cheapest choice for every task. The cost and score figures should be read as reported benchmark results and the author’s current assessment, rather than settled conclusions about all users.

Testing the Stack on Real Work

The next step for readers considering a similar setup is a shadow test on their own tasks, as Meyer advises. Teams can compare outputs at the settings they would actually use, check whether results meet their quality requirements, and record both model and human review costs. The source does not provide a date for a follow-up comparison or report results from such a test.

Further Artificial Analysis entries could change the comparison as settings become available or model scores are updated. In particular, Meyer notes that Sol’s low and max settings were not listed on September 29. Until those figures and workload-specific results are available, the article’s stack remains a dated snapshot: Opus for building, Sol for digging and review, and alternatives for selected tasks.

Key Questions

What is the main development in the article?

Thorsten Meyer published his September 29 model stack, pairing Opus 5.5 for development with GPT-6.1 Sol for detail work and review, based partly on a comparison of benchmark scores and cost per task.

Why does Meyer use GPT-6.1 Sol for review?

He says its listed high and xhigh costs—$0.32 and $0.39 per task—make a separate review pass affordable in his workflow. The article does not establish how accurately Sol reviews work across other users’ tasks.

Does the index prove which model is best?

No. Meyer describes the Intelligence Index as a general capability measure and advises testing against the work at hand. A score does not establish that a model will meet every organization’s quality or cost requirements.

What remains unknown about GPT-6.1 Sol?

At publication, the source says Artificial Analysis had listed medium, high and xhigh settings, but not low or max. It also reports long first-token times at high and xhigh; performance and timing on other workloads are not established in the article.

Source: ThorstenMeyerAI.com