An Everyday Guide To My September AI Stack

🔍 Read the full analysis: An Everyday Guide To My September AI Stack on ThorstenMeyerAI.com

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

Get hardware and tech essentials 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 guide on September 29, 2026, describing how he assigns six AI models to building, review, and routine tasks as capability scores cluster while reported costs per task vary widely. His current setup uses Claude Opus 5.5 for most development and newly released GPT-6.1 Sol for detailed analysis and independent review.

Meyer bases the comparison mainly on the Artificial Analysis Intelligence Index v4.3.x, which he describes as a general capability measure rather than a verdict on any specific workload. In his table, Opus 5.5 scores 58 at its top setting, GPT-6.1 Sol scores 51 at xhigh, and GPT-6 Luna scores 37. The listed cost per task ranges from $0.07 for Luna to $7.63 for Fable 5.1; these are the guide’s estimates, not prices that every user will necessarily pay.

For development, Meyer selects Opus 5.5 at high effort for routine feature work and xhigh for more demanding problems such as architecture, migrations, and trust boundaries. At those settings, the guide reports scores of 54 and 56, respectively, and costs of $1.82 and $3.46 per task. Meyer says max effort reaches a score of 58 at $5.98 per task, which he considers rarely worthwhile for his work.

GPT-6.1 Sol’s role is narrower: examining details and reviewing work produced by another model. The guide lists its high setting at $0.32 per task and xhigh at $0.39. Meyer says that price makes routine review practical in his workflow. He reserves Astra or Fable for second opinions when Sol and Opus disagree, and names Sonnet 5.5 and Luna as options for scoped subtasks and routine checks.

Why Meyer Splits Building and Review

The guide’s central point is that model selection can depend on the cost of reaching an acceptable result, not just a leaderboard position. Meyer reports that six models sit within about 20 index points, while their estimated task costs differ substantially. In his account, GPT-6.1 Sol at xhigh scores one to two points below Astra at xhigh or max and Fable 5.1, while costing $0.39 per task compared with $3.26 and $7.63.

That price difference shapes his review process. Meyer argues that a model from a different family can provide a useful second look at code or other work, and says Sol’s cost lets him apply that check to each meaningful change. He also cautions that a second model cannot correct missing requirements in the original specification, and that passing tests alone does not amount to approval to ship. Those are his workflow recommendations, not independently verified findings about review accuracy.

The cost figures also need to be read alongside latency and human effort. Meyer reports that Sol’s high and xhigh settings took 57 to 69 seconds to produce a first token in the index evaluation, which makes them less suited to interactive exchanges. He further argues that human review time can outweigh a modest reduction in model charges. The source gives an illustrative example, not a measured general result, and its final cost discussion is incomplete.

The Index Behind the Model Choices

The guide compares model scores and costs using Artificial Analysis Intelligence Index v4.3.x unless it says otherwise. It presents the index as a map of general capability. Meyer explicitly advises readers to shadow-test models on their own tasks before switching a workflow, since the ranking may not predict performance on a particular team’s work.

Its cost comparisons cover both model choice and effort setting. For Opus 5.5, the table gives scores and estimated costs from low through max: low is 42 at $0.55 per task, medium 51 at $1.34, high 54 at $1.82, xhigh 56 at $3.46, and max 58 at $5.98. Meyer says moving from medium to max raises the estimated cost 4.46 times for seven additional index points. For Sonnet 5.5, the table puts high at 47 and $1.08, while max reaches 56 at $7.60.

The guide also lists token prices, which are distinct from its estimated cost-per-task figures. It gives Opus 5.5 at $4 per million input tokens and $20 per million output tokens, with cache reads at $0.20. GPT-6.1 Sol is listed at $2 and $10 per million input and output tokens, respectively. These rates alone do not establish what a given task will cost; usage and task length affect the bill.

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

— Thorsten Meyer

Limits of the Cost and Score Data

The source does not provide enough information to independently reproduce every cost-per-task estimate or assess how closely the index tasks match a reader’s workload. Meyer says one index point is within the noise, and the guide notes that Artificial Analysis had not yet published GPT-6.1 Sol’s low or max settings when he wrote it. It also does not establish that the listed model choices will perform best for other users.

The guide’s final section on the cost of human review ends mid-sentence in the supplied material. It introduces an illustrative example saying that halving model prices saves 12.5% of “the real cost” and that an extra minute of human review can erase the saving, but the remaining calculation and its assumptions are not available. That claim cannot be assessed from the material provided. The source also does not say whether the listed task costs include all usage, subscription, or operational expenses.

Test the Stack on Real Tasks

Meyer’s stated next step for readers is to run candidate models against their own work before switching. A practical comparison would record quality, cost, and response time for the same tasks, then check whether the result meets the team’s requirements. The guide does not announce a formal benchmark update or a scheduled follow-up.

Further comparisons may change as index results are added, including the missing GPT-6.1 Sol effort settings noted in the article. Until those results appear and teams test the models directly, Meyer’s September 29 stack is a snapshot of his own choices and the cost and capability figures available to him at publication.

Key Questions

Which model does Meyer use for most building work?

He names Claude Opus 5.5 at high effort for routine development and xhigh for harder problems such as architecture and migrations.

What does GPT-6.1 Sol do in his setup?

Meyer uses GPT-6.1 Sol at high or xhigh to investigate details and review work produced by another model. The guide lists estimated costs of $0.32 to $0.39 per task at those settings.

Are the article’s task costs guaranteed prices?

No. They are cost-per-task estimates reported in Meyer’s comparison. The source does not establish that they will match every user’s workload, usage, or total operating costs.

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

No. Meyer says the index measures general capability and advises readers to shadow-test models on their own tasks before making a switch.

Source: ThorstenMeyerAI.com