September 2026: The AI Stack I Use To Build And Decide

🔍 Read the full analysis: September 2026: The AI Stack I Use To Build And Decide 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 said on September 29 that he uses Claude Opus 5.5 as his main model for building software and the newly released GPT-6.1 Sol for detailed review. His account compares model scores and estimated task costs, arguing that similar benchmark results can come with sharply different prices—a distinction that could shape how developers assign work and budget for AI use.

Meyer bases the comparison primarily on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a measure of general capability rather than a verdict on any particular workload. In his table, Opus 5.5 scores 58 at its top setting and costs an estimated $5.98 per task. GPT-6.1 Sol scores 51 at xhigh for $0.39 per task. The table also lists GPT-6 Astra at 53 for $3.26, Claude Fable 5.1 at 53 for $7.63, Sonnet 5.5 at 56 for $7.60, and GPT-6 Luna at 37 for $0.07.

Those figures lead Meyer to assign different roles rather than choose one model for every job. He uses Opus 5.5 at high for features, APIs and multi-file changes, and xhigh for harder work such as architecture and migrations. He assigns GPT-6.1 Sol at high or xhigh to focused investigations and reviews. Sonnet 5.5 and Luna handle scoped subtasks and routine classification, while Astra or Fable are alternatives when his tests favor them or when models disagree.

Meyer says the effort setting can change task costs substantially. In the index figures he cites, raising Opus 5.5 from medium to max increases its score from 51 to 58 while estimated cost per task rises from $1.34 to $5.98. For Sonnet 5.5, max costs $7.60 for a score of 56, compared with $2.74 and a score of 52 at xhigh. He favors high or xhigh Opus for development and describes medium as his default for documents and everyday work.

How Cost Shapes Model Assignments

Meyer’s account reflects a practical change in how he evaluates AI tools: the model with the highest score is not automatically the one he wants running every task. If his estimates hold for his workload, a less expensive model can make a routine second review affordable, while a stronger model is reserved for work where he believes its added capability justifies the cost.

He says a different model family reviewing Opus can provide a useful second perspective, and that GPT-6.1 Sol at high or xhigh costs $0.32 to $0.39 per task in the cited index. That is a claim about the potential value of his process, not evidence that cross-model review catches every problem. Meyer also warns that lower model charges do not automatically mean lower total costs: added human review time can outweigh token savings. His example is illustrative rather than measured.

For teams, the report offers a case for testing models by task and effort level before changing deployment choices. The index figures may help shortlist candidates, but they do not establish which model is most accurate, fast or economical on a reader’s own code, documents or decision process.

The Benchmark Behind Meyer’s Choices

Meyer’s comparison is dated September 29, 2026, and says GPT-6.1 Sol was released that day. It places that model among six options whose top-setting scores range from 37 to 58 on the named index. Meyer says the group sits within about 20 index points, while estimated per-task costs differ by roughly 100 times. Those are his summary figures, drawn from the cited benchmark and cost estimates.

The source gives token prices as well as estimated task costs: GPT-6.1 Sol at $2 per million input tokens and $10 per million output tokens; Opus 5.5 at $4 and $20, with cache reads at $0.20; and Luna at $0.10 and $0.50. Fable and Astra are listed at $10 and $50. Token prices alone do not determine the cost of a task, which can also depend on how many tokens a model uses and which setting is selected.

Meyer notes that GPT-6.1 Sol’s high and xhigh settings took 57 and 69 seconds, respectively, to produce a first token in the index. He also reports 25 million output tokens at high, compared with an 82 million median for comparable models. These are benchmark-specific observations; they do not establish response times for every prompt or service configuration.

“The index is a map of general capability, not a verdict on your workload.”

— Thorsten Meyer

What the Index Cannot Establish

The source does not provide enough detail to independently assess how its estimated cost per task was calculated or how closely the benchmark tasks resemble Meyer’s day-to-day work. The author advises readers to shadow-test before switching models. Results on a general index may not predict quality on a particular codebase, document set or workflow.

Meyer also says Artificial Analysis had not yet published GPT-6.1 Sol results for low or max effort. He cautions that a one-point index difference is within the noise, so small gaps should not be read as firm rankings. The source does not report a controlled comparison of review accuracy, human time saved or total project cost across the models.

His account is a first-person description of his own model choices. It does not establish that Opus 5.5, GPT-6.1 Sol or any other model will produce the same quality or savings for other users. The source’s final discussion of human review cost is cut off, so its full illustrative calculation is not available.

Testing Before a Model Switch

Meyer’s stated next step for anyone considering a change is to shadow-test the candidate model on their own work before replacing an existing system. A useful comparison would track task quality, model charges, waiting time and human review effort, since the benchmark’s per-task estimate alone does not capture all of those costs.

For Meyer’s workflow, the described plan is to keep Opus 5.5 on development work, use GPT-6.1 Sol for detailed investigation and review, and call on alternatives when a particular task or disagreement warrants them. Future Artificial Analysis results for Sol’s missing effort settings could add comparison points. Whether those results change Meyer’s assignments, or show similar advantages on other workloads, remains to be seen.

Key Questions

Which models does Meyer use for development and review?

He says Opus 5.5 is his main model for building, while GPT-6.1 Sol handles focused investigation and review. These are his reported choices, not a universal recommendation.

What benchmark does the comparison use?

The scores mainly come from the Artificial Analysis Intelligence Index v4.3.x. Meyer says it measures general capability and advises readers to test models on their own workloads.

Why does Meyer use different effort settings?

His cited figures show higher effort can raise scores while also increasing estimated task cost. He favors high or xhigh Opus for development and says medium remains his default for documents and everyday work.

Does GPT-6.1 Sol outperform Opus 5.5?

Not in Meyer’s cited top-setting scores: Opus 5.5 scores 58 and Sol xhigh scores 51. Sol’s estimated task cost is lower in the comparison, but Meyer assigns the models different jobs.

Should readers switch models based on these figures?

Meyer recommends shadow-testing before a switch. The index describes general capability, and the source does not establish which model will perform best on another user’s tasks.

Source: ThorstenMeyerAI.com