The AI Stack I’m Working With In September 2026

🔍 Read the full analysis: The AI Stack I’m Working With In September 2026 on ThorstenMeyerAI.com

Buying for a business?Offer from Amazon

Get business pricing on monitors, keyboards and dev gear

• 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 for building software and GPT-6.1 Sol for detailed investigation and review, as his comparison puts substantial cost differences alongside closely grouped benchmark scores. The account matters to developers choosing models because it makes cost per measured task, rather than a single top score, the basis for assigning work.

Meyer’s comparison draws its general capability scores from the Artificial Analysis Intelligence Index v4.3.x, which he describes as a map of broad capability rather than a verdict on any individual workload. In the table he supplies, Opus 5.5 scores 58 at its top setting and costs an estimated $5.98 per task. GPT-6.1 Sol at xhigh scores 51 for $0.39 per task; GPT-6 Astra scores 53 at max for $3.26. These are index results and task-cost estimates, not guaranteed prices for every user or project.

Meyer assigns Opus 5.5 high to routine development, citing an index score of 54 for $1.82 per task, and xhigh to difficult work such as architecture, migrations and trust boundaries. He reserves max for occasional cases: in his figures, that setting raises Opus’s score from 56 at xhigh to 58 while increasing estimated cost from $3.46 to $5.98. He places Sonnet 5.5 at high for scoped tasks and documents, and says he uses Luna for classification, extraction and routing.

GPT-6.1 Sol, released on the publication date according to Meyer, has three settings in the index data he cites. Medium scores 48 at $0.21 per task, high scores 50 at $0.32, and xhigh scores 51 at $0.39. Meyer uses Sol high or xhigh to inspect specific files and diffs and to review Opus’s work. He says Astra or Fable may serve as another opinion when Sol and Opus disagree, while Jev, described as a decision model that cannot write sentences, handles high-volume yes-or-no and routing judgments.

How Task Costs Shape Model Choice

The comparison illustrates why model selection can affect both software costs and review practices. Meyer’s index table places several models within a narrow score range while listing much lower estimated costs for Sol than for Astra or Fable. If those estimates carry over to a team’s own tasks, a low-cost review pass may be practical to run more often. That is a possible operational benefit, not an independently measured outcome in the source.

The same figures show that a model’s effort setting can change the bill substantially. For Opus 5.5, Meyer reports a rise from $1.34 per task at medium to $5.98 at max, with the index score moving from 51 to 58. His stated choice of high or xhigh for development is a personal allocation based on that trade-off. Other teams may reach different results if their prompts, task lengths, tools or quality requirements differ.

Meyer also distinguishes a different model family from a fully independent check. He cautions that a reviewer can share a flawed specification with the model that produced the work, and says passing tests alone do not amount to approval to ship. Those are practical cautions in his account, rather than benchmark findings. They matter because low-cost review only helps if teams examine its evidence and resolve failures carefully.

The September Model Lineup

Meyer’s comparison covers models released between September 1 and September 29, 2026: Claude Fable 5.1 on September 1, GPT-6 Astra on September 3, Claude Opus 5.5 and GPT-6 Luna on September 22, Claude Sonnet 5.5 on September 28, and GPT-6.1 Sol on September 29. Those release dates and model descriptions come from the supplied source; the article does not independently verify them.

In the cited top-setting table, Fable 5.1 scores 53 at an estimated $7.63 per task, while Sonnet 5.5 scores 56 at $7.60. Astra is listed at 53 and $3.26. Meyer argues these costs make some models harder to justify at their highest settings for his purposes. The table also gives token prices: Opus at $4 per million input tokens and $20 per million output tokens, Sol at $2 and $10, and Luna at $0.10 and $0.50. Token prices and task estimates are different measures.

The source notes that the index has not published low or max settings for Sol. It also says one-point score differences fall within the noise. Meyer therefore advises readers to shadow-test models against their own workloads before switching. No results from such an external test are included in the material provided.

“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

What the Index Cannot Settle

The source does not establish whether Meyer’s task-cost estimates will match costs for other users, or how the six models perform on any particular company’s code and documents. It provides no independent replication of the index results and no controlled comparison of completed work, error rates or human review time. Meyer’s recommendation to shadow-test reflects that gap.

Sol’s high and xhigh settings also come with a speed trade-off in the cited data: time to first token is listed as 57 and 69 seconds, respectively. The source does not say how those timings vary across users or workloads. It gives output-token totals from the index, but does not establish what a particular user will generate on a given task.

The account’s cost argument has a further limit: it says a minute of additional human review can erase a model-price saving, but the supplied material ends before completing its illustrative dollar example. No measured human-review study or calculation is provided. Whether lower model costs reduce total work therefore remains unclear.

Testing the Stack on Real Work

Meyer’s immediate recommendation is to shadow-test before switching: run candidate models against the same real tasks and compare their results with the team’s quality bar. A useful follow-up would record output quality, latency, model charges and the human time needed to catch or correct errors. The supplied account does not report results from a broader test or name a next publication date.

Readers considering the same assignments can treat Meyer’s roles as a starting point: Opus for building, Sol for detailed review, and lower-cost or specialist models for defined tasks. The evidence presented supports a comparison of the cited index scores and estimates; whether this allocation saves money or improves work in another setting remains to be tested.

Key Questions

Which model does Meyer use for building?

He says he uses Claude Opus 5.5, usually at high effort and at xhigh for harder work such as architecture or migrations.

What role does GPT-6.1 Sol have in his stack?

Meyer uses Sol at high or xhigh to investigate specific files and diffs and to review work produced with Opus. He reports an estimated cost of $0.32 to $0.39 per task at those settings in the cited index data.

Are the reported costs guaranteed prices?

No. They are per-task estimates in Meyer’s comparison. Token prices are listed separately, and actual costs can depend on the task and usage.

Does the comparison prove Sol is as capable as Opus?

No. The cited index scores place Sol xhigh at 51 and Opus xhigh at 56. Meyer also notes that a one-point difference is within the index’s noise and recommends testing models on the intended workload.

What should teams test before changing models?

Meyer recommends shadow-testing candidates on their own tasks. Teams would need to compare quality, costs and the time required for human review; the source does not provide those results for other organizations.

Source: ThorstenMeyerAI.com